Imagine waking up to a coffee maker that predicts your needs, a self-driving car, and a phone that writes essays for you. This AI-powered world exists today.
In this course, you will explore artificial intelligence concepts that are provided in simple terms for everyone, regardless of technical background. You will discover how machines think and how AI is transforming the world around us.
By course end, you will grasp AI and machine learning basics, understand large language models like Amazon Nova, and recognize their opportunities and ethical questions. Whether you are exploring AI for career growth or simply to understand the world-changing technology, this course provides your essential foundation.
Course objectives
In this course, you will learn to do the following:
Define key artificial intelligence and machine learning (AI/ML) concepts using everyday language.
Identify different types of AI systems and explain how they function.
Recognize how large language models (LLMs) work and their practical applications.
Discover real-world AI applications across various industries.
Understand the ethical considerations that guide responsible AI development.
Introduction to the Foundations of AI and ML
Section objectives
In this section, you will learn to do the following:
Identify basic concepts of artificial intelligence and machine learning.
Identify the relationship between AI, machine learning (ML), and data science.
Have you ever wondered how Amazon Prime Video seems to know exactly what show you might want to watch next? Or how your email sorts of spam without you having to check each message? These everyday technologies use AI, ML, deep learning, and data science to make your life easier.
History and evolution of artificial intelligence
AI has had quite a journey over the decades. The following are some key moments.
The Birth of AI
The term, artificial intelligence, was coined in 1956 at a conference at Dartmouth College. Early researchers were optimistic they could build machines with human-like intelligence within a generation.
Reality Sets In
Researchers discovered that creating true intelligence was much harder than expected. Funding decreased during what became known as the AI Winter.
The Return
AI made a comeback with practical applications in specific domains. IBM's Deep Blue chess computer defeated world champion Garry Kasparov in 1997.
Big Data and Deep Learning
The explosion of internet data and increased computing power led to breakthroughs. Machine learning algorithms improved dramatically, especially deep learning methods.
AI Goes Mainstream
AI has become part of everyday life, from digital assistants like Alexa to recommendation systems on streaming services. Language models can write human-like text, and computer vision systems can recognize objects better than humans in some cases.
Differences Between AI and ML
The following sections distinguish artificial intelligence from machine learning.
Artificial Intelligence
Artificial Intelligence (AI) is teaching computers to perform tasks that typically require human intelligence.
Imagine if you could train your computer to recognize your friends in photos, understand questions when spoken aloud, or drive a car. That is AI in action.
At its core, AI is about creating computer systems that can perform the following:
Solve problems
Learn from experience
Understand natural language
Recognize patterns
Make decisions
Machine learning
Machine Learning (ML) is a specific approach to creating AI. Instead of programming explicit instructions for every situation, you give the computer examples and let it figure out the patterns on its own.
For example, teaching a computer to recognize dogs in pictures. With traditional programming, you would have to write specific rules like, if it has four legs, fur, and a tail, it is a dog. That approach falls apart quickly because there are dogs without tails. The scenario also didn't consider cats, which share features like four legs, fur, and a tail.
With machine learning, the computer analyzes thousands of pictures labelled dog and not dog. The computer then identifies patterns in these images and builds its own rules for recognizing dogs. When shown a new image, it can apply these learned patterns to decide if it's looking at a dog.
Relationship Between AI, ML, Deep Learning, and Data Science
The following definitions explain the relationship between AI, machine learning, deep learning, and data science.
Figure 1 Relationship Between AI, ML, Deep Learning, and Data ScienceSelect image to enlarge
Artificial intelligence
AI is a technology with human-like problem-solving capabilities. Organizations use AI to analyze data to enhance operations.
Machine learning
Machine learning is a type of artificial intelligence that performs data analysis tasks without explicit instructions. Machine learning technology can process large quantities of historical data, identify patterns, and predict new relationships between previously unknown data.
Deep learning
Deep learning is an AI method that teaches computers to process data in a way inspired by the human brain. You can use deep learning methods to automate tasks that typically require human intelligence, such as describing images or transcribing a sound file into text.
Data science
Data science is the study of data to extract meaningful insights for business. This multidisciplinary approach combines principles and practices from mathematics, statistics, and AI. It also incorporates computer engineering to analyze large amounts of data. This helps data scientists to ask and answer questions like what happened, why things happened, and what will happen in the future.
RECALL & APPLY
Practice what you learned
Try answering in your own words, then compare with the suggested answer. These activities are for self-study and are not automatically graded.
01Which part of your customer journey would benefit most from AI support?
Topic: Course scope and learning outcomes
Reveal suggested answer
Accept a specific operational problem and a measurable outcome. Revisit the example as the course progresses.
02Why did larger datasets matter to AI development?
Topic: AI history and development
Reveal suggested answer
More varied examples helped models learn patterns that were difficult to encode as individual rules.
03Is every AI application a machine learning application?
Topic: AI and machine learning
Reveal suggested answer
No. A system can use explicitly programmed rules. Machine learning is one approach within the broader AI field.
Capture a key idea, an example, or a question for your instructor.
Before exploring AI solutions for Amazon Connect Customer, review the fundamental AI terminology. These concepts form foundational understanding of AI systems that enhance customer service platforms. This knowledge helps you understand how artificial intelligence works in contact center environments.
Similar to how customer service representatives need to learn specific job skills, AI systems require certain components to function effectively.
Figure 2 AlgorithmSelect image to enlarge
Algorithm
An algorithm is simply a set of steps to solve a problem, like a recipe for computers. We all use algorithms in everyday life without calling them that.
Figure 3 ModelSelect image to enlarge
Model
A model is what you get after an AI system learns from data. Think of it like a student's brain after studying for a test. The student (or AI) has created an internal understanding that can now be applied to new situations.
Figure 4 TrainingSelect image to enlarge
Training
Training is the process of teaching an AI system by showing it examples. The training is similar to how you might learn to identify birds by looking at different types with a guidebook.
Figure 5 InferenceSelect image to enlarge
Inference
Inference is when an AI system applies what it has learned to new situations. The inference is like taking what you learned in driving lessons and using it to drive on a road you have never been on before.
Figure 6 DatasetSelect image to enlarge
Dataset
A dataset is simply a collection of information used to train or test an AI. Think of the dataset like a textbook full of examples for the AI to learn from.
By understanding these basic AI terms, you can develop the foundation needed to understand AI capabilities. These fundamentals work together to create systems that can enhance customer interactions and streamline contact center operations. With this terminology in mind, you are better equipped to understand how AI solutions are built and deployed.
Role of Data in Machine Learning
What Is the Role of Data in Machine Learning
In machine learning, data is everything. It is the foundation that makes learning possible. Without data, AI systems would have nothing to learn from.
Clean, diverse data teaches AI better than massive amounts of similar examples. AI systems need accurate information, just like a student needs reliable textbooks for learning.
Every AI task requires specific data types. Speech recognition needs audio samples. Image recognition demands diverse pictures. Each job has unique data requirements.
Understanding how machines learn from information
Although the advanced math can be complex, the basic process is something everyone can understand. The following explains the AI learning process:
Start with a task: First, define what the AI must learn. This might be recognizing faces in photos, translating languages, or predicting which movies someone might like.
Gather example data: Next, collect examples related to this task. If you are teaching an AI to recognize apples, gather thousands of apple images and images of things that aren't apples.
Feature recognition: The AI examines the examples to identify features or patterns. For apple recognition, it might notice characteristics like round shapes, red or green colors, and particular textures.
Pattern detection: As the AI looks at more examples, it starts to recognize patterns that differentiate categories. It might learn that apples are typically round, but so are oranges, so roundness alone isn't enough to identify an apple.
Trial and error: The AI makes predictions based on what it's learned so far and checks if it's right. If it misidentifies a red ball as an apple, it adjusts its understanding.
Refinement: With each attempt, the AI adjusts its understanding to improve accuracy. This might involve giving more weight to certain features (color and texture) and less to others (exact size).
Insider tip: Garbage in, garbage out is a fundamental principle in machine learning. Low quality data causes low quality results.
Remember: Machines do not truly understand what they are learning in the way humans do.
When an AI learns to identify pictures of cats, it does not know what a cat actually is, it just recognizes patterns of pixels that humans have labelled as cats.
RECALL & APPLY
Practice what you learned
Try answering in your own words, then compare with the suggested answer. These activities are for self-study and are not automatically graded.
01A contact classifier labels a new chat. Is this training or inference?
Topic: Core AI vocabulary
Reveal suggested answer
Inference. The application is applying an existing trained model to a new example.
02What happens if training examples exclude a common customer accent?
Topic: Data quality and model learning
Reveal suggested answer
The model may perform less well for that group. The dataset and evaluation examples need suitable representation.
Capture a key idea, an example, or a question for your instructor.
Have you ever wondered why a virtual assistant can tell you the weather but struggles to understand sarcasm in your voice? Or why your photo app can recognize your friends' faces but cannot tell you what they are thinking? The answer lies in understanding the different categories of AI systems and their capabilities.
In this section, you will explore the various types of artificial intelligence. From simple systems that follow preset rules to sophisticated learning systems that can improve over time. You will also explore the current state of AI technology and what might be possible in the future. Understanding these categories will help you recognize the strengths and limitations of the AI you encounter every day.
Section objectives
In this section, you will learn to do the following:
Identify narrow AI and general AI systems.
Compare rule-based AI systems with learning-based systems.
Identify different machine learning approaches and their real-world applications.
Introduction
When discussing AI systems, you should understand narrow and general AI. This distinction explains why AI in Amazon Connect Customer has specific capabilities and limitations. Narrow AI excels at specialized tasks like customer interactions. General AI, seen in science fiction, would reason broadly like humans but does not exist yet.
By understanding this fundamental difference, you can set realistic expectations for what AI can accomplish in your contact center operations.
Narrow AI
Narrow AI performs specific tasks extremely well but only those particular jobs. These systems excel at their assigned functions but cannot apply their skills elsewhere.
Figure 7 Narrow AISelect image to enlarge
Narrow AI examples
Spam filters protect your inbox but can't summarize your important messages.
Translation tools convert your sentences between languages but miss cultural nuances.
Voice assistants tell you weather forecasts but get confused by philosophical questions.
Narrow AI is a specialized tool in your digital toolkit. Excellent for specific tasks, yet ineffective for everything else.
General AI
General AI refers to systems with human-like intelligence, such as being able to understand, learn, and apply knowledge across different domains.
Figure 8 General AISelect image to enlarge
General AI examples
Transfers knowledge between different fields.
Reasons about abstract concepts.
Understands context and nuance.
Sets its own goals and pursue them.
Responds with something resembling consciousness or self-awareness.
Here is the important part: true general AI doesn't exist yet.
RECALL & APPLY
Practice what you learned
Try answering in your own words, then compare with the suggested answer. These activities are for self-study and are not automatically graded.
01Why can a strong spam classifier still fail at a different task?
Topic: Narrow AI and general AI
Reveal suggested answer
Its learned patterns and interface address spam classification, not every possible reasoning or language task.
Capture a key idea, an example, or a question for your instructor.
Another important way to categorize AI systems is by how they make decisions. Do they follow preset rules, or do they learn patterns from data?
Rule Based Systems
Rule-based systems follow explicit if-then rules created by human programmers. These systems have the following characteristics:
Follow logical rules created by experts
Have predictable behavior
Cannot improve without humans changing their rules
Work well for problems with clear rules and limited variables
A classic example is a tax preparation program. The program doesn't learn new tax rules on its own. Humans must program new rules when tax laws change.
Learning Systems
Learning systems use data to develop their own rules and improve over time. These systems have the following characteristics:
Identify patterns in data without explicit programming
Can improve with more experience and data
Might discover unexpected connections
Require training instead of rule-writing
Example: Modern spam detection learns from user behavior. When you mark an email as spam, the system learns from that example and gets better at identifying similar emails in the future.
Compare the systems
Neither approach is universally superior, but each has strengths for different situations. The following table compares the strengths of each system.
Rule-based systems are better
Learning systems are better
Precision is crucial (like medical diagnosis or legal applications).
The problem is too complex for humans to define all the rules.
The problem domain has clear, unchanging rules.
The environment changes frequently.
Available data is limited.
Large amounts of data are available.
Mistakes would be costly or dangerous.
The system needs to personalize to individual users.
Both approaches have valuable applications. Rule-based systems are suitable for handling structured workflows with clear rules. Learning systems are more suited for adapting to evolving customer needs. Understanding these differences helps you choose the right AI approach for each contact center challenge.
Learning approaches
Figure 9 Learning ApproachesSelect image to enlarge
Introduction to learning approaches
Machine learning teaches computers to make predictions based on data. You train a model using an algorithm and example data. Then, your application uses this model to generate real-time predictions at scale..
Three main learning methods
Explore the fundamental approaches that power AI systems in contact centers and beyond, each with unique strengths for different customer service challenges.
Supervised learning
In supervised learning, you provide the AI system with labeled data. Labeled data is information that comes with the correct answers already attached, like flashcards where questions have answers on the back.
The system then learns to predict the correct output for new, unseen inputs. It's like learning with a teacher who provides practice problems and the answers. Real-world examples of supervised learning include the following:
Email spam filters
Medical diagnosis from images
Price prediction for homes
Facial recognition systems
Unsupervised learning
Unlike supervised learning, unsupervised learning does not start with labeled data. The system identifies patterns, structures, or relationships within the data on its own. Real-world examples of unsupervised learning include the following:
Customer segmentation for marketing
Recommendation systems, such as recommendations based on customers who bought similar items
Anomaly detection for fraud prevention
Topic discovery in document collections
Reinforcement learning
Unlike supervised learning, reinforcement learning works through trial and error. The AI agent learns by taking actions and receiving feedback through rewards or penalties. Through this process, it gradually improves its strategy to maximize long-term rewards.
Real-world examples of reinforcement learning include the following:
Game playing (like AWS DeepRacer)
Autonomous vehicles learning to navigate traffic
Industrial robotics learning optimal movement patterns
Energy management systems optimizing resource usage
Choosing a learning approach
To create successful AI implementation, it is crucial to understand which learning approach fits your specific contact center needs. Each method addresses different types of business challenges.
Supervised learning
Unsupervised learning
Reinforcement learning
Use supervised learning for the following situations:
You have labeled training data available.
Your problem has clear input-output pairs.
You need specific, predictable responses.
Your task involves classification or regression.
Supervised learning is best suited for customer sentiment analysis and call categorization. It requires historical data with known outcomes.
All three approaches solve different challenges. Choosing the right approach depends on your data and business goals.
Learning approach selection in practice
Supervised learning
Use supervised learning for the following situations:
You have labeled training data available.
Your problem has clear input-output pairs.
You need specific, predictable responses.
Your task involves classification or regression.
Unsupervised learning
Use unsupervised learning for the following situations:
You lack labeled data.
You want to discover hidden patterns.
You need to group similar items together.
You are exploring data without specific predictions in mind.
Unsupervised learning helps identify emerging customer concerns and conversation themes. It finds valuable insights when you don't know what you're looking for.
Reinforcement learning
Use reinforcement learning for the following situations:
Your problem involves sequential decision-making.
You can define clear reward signals.
Your agent needs to learn through trial and error.
You are building systems that interact with environments.
Reinforcement learning optimizes dynamic call routing and conversational flows. It improves through ongoing interactions with customers and systems.
All three approaches solve different challenges. Choosing the right approach depends on your data and business goals.
RECALL & APPLY
Practice what you learned
Try answering in your own words, then compare with the suggested answer. These activities are for self-study and are not automatically graded.
01Which approach suits a fixed eligibility rule?
Topic: Rules and learned patterns
Reveal suggested answer
An explicit rule is usually easier to inspect and enforce when the policy is precise and stable.
02Which method fits grouping similar contacts without predefined categories?
Topic: Machine learning approaches
Reveal suggested answer
Unsupervised learning, because the aim is to discover groups or structures in unlabeled data.
03How would you approach sentiment classification with labeled historical calls?
Topic: Selecting a learning approach
Reveal suggested answer
Start with a supervised learning formulation and evaluate performance on representative held out examples.
Capture a key idea, an example, or a question for your instructor.
AI systems vary dramatically in their complexity and capabilities. Some can only perform simple tasks within narrow domains, whereas others demonstrate impressive versatility across multiple applications. Understanding this spectrum helps set realistic expectations about what different AI technologies can do.
Simple AI Systems
Simple AI systems typically do the following:
Focus on a single, well-defined task.
Follow straightforward logic or learned patterns.
Have limited adaptability to new situations.
Require human intervention when facing unexpected scenarios.
The following are simple AI system examples:
Rule-based bots that follow predefined conversation flows
Basic recommendation systems that suggest products based on past purchases
Document classification systems that sort text into predefined categories
Simple voice recognition systems that respond to specific commands
These systems are like specialized tools. They do one job well but aren't flexible.
Complex AI Systems
Complex AI systems typically do the following:
Handle multiple related tasks or domains.
Combine multiple AI techniques and models.
Adapt to new situations within their domain.
Process and integrate different types of information.
The following are complex AI system examples:
Virtual assistants that understand context across conversations
Autonomous vehicles that navigate unpredictable environments
Advanced medical diagnostic systems that consider numerous factors
Large language models that can generate text across various topics and styles
These systems are more like versatile assistants. They can handle a range of situations within their area of expertise.
AI System Complexity
The following factors can contribute to AI system complexity:
Multimodal processing: Is the system effective at working with different types of data, such as text, images, and audio, simultaneously?
Memory and context: Does the system remember previous interactions and maintain context over time?
Adaptation capability: How well can it adjust to new situations without requiring retraining?
Integration of multiple models: Does the system combine several specialized AI models to achieve its goals?
Learning approach: Does it use multiple learning approaches together, such as supervised, unsupervised, reinforcement?
Consider that even the most complex AI systems today are still fundamentally narrow AI. They are just very sophisticated within their domains. The apparent intelligence of these systems comes from their specialized capabilities rather than general human-like understanding.
RECALL & APPLY
Practice what you learned
Try answering in your own words, then compare with the suggested answer. These activities are for self-study and are not automatically graded.
01Does adding several models remove the need for defined scope?
Topic: Simple and complex AI systems
Reveal suggested answer
No. The integrated system still needs a clear purpose, evaluation criteria, and an understood operating boundary.
Capture a key idea, an example, or a question for your instructor.
So far, you have explored narrow AI, general AI, rule-based systems, and learning systems. But a new category of AI has emerged that changes how we think about what AI can do in the workplace: agentic AI.
Unlike traditional AI systems that wait for instructions and respond to specific queries, agentic AI systems can reason about problems, remember context across interactions, and take independent action to achieve goals. Think of the difference between a calculator (you ask, it answers) and a skilled colleague (they understand your goal, figure out the steps, and get it done).
What Is Agentic AI
Agentic AI represents a fundamental shift in how AI systems operate. These systems go beyond simple input-output responses to demonstrate autonomous goal-directed behavior.
Reasoning
Agentic AI systems can break down complex problems into steps, evaluate options, and determine the best course of action — much like a human problem-solver would approach an unfamiliar situation.
For example, when a customer contacts support with a billing discrepancy, an agentic AI system doesn't just retrieve account information. It reasons through the issue: checking recent transactions, identifying the anomaly, determining the cause, and deciding on the appropriate resolution.
Memory
Agentic AI systems maintain context across interactions. They remember what happened earlier in a conversation, what was discussed in previous sessions, and what actions have already been taken.
This means a customer doesn't have to repeat their issue every time they interact with the system. The AI agent remembers the full history and picks up where things left off.
Action
Perhaps most importantly, agentic AI systems can take independent action. They can use tools, call APIs, update records, send messages, and complete tasks. They can not just suggest next steps, but actually run them.
An agentic AI system might look up a customer's order, check the shipping status, initiate a refund, send a confirmation email, and update the case record as part of resolving a single customer request.
Tool Use
Agentic AI systems interact with external tools and services through standardized protocols. The Model Context Protocol (MCP) is one such standard that enables AI agents to retrieve information from knowledge bases, query databases, and complete actions in external systems.
This tool-use capability is what transforms a language model from a conversational interface into a capable agent that can accomplish real work.
Humorphism
AWS uses the term "humorphism" for this design philosophy, where AI is designed to collaborate like a human teammate rather than operate like a conventional software tool.
In a traditional software model, you tell a tool exactly what to do, step by step. In the humorphism model, you describe what you need and the AI figures out how to accomplish it, similar to how you would delegate a task to a capable team member.
This distinction is central to understanding how Amazon Connect Customer has evolved. Rather than providing a set of tools for contact center agents to operate, Amazon Connect Customer now provides AI agents that work alongside human agents. These agents help with reasoning through problems, remembering context, and taking action independently.
How Agentic AI relates to other categories
Agentic AI sits alongside the categories you have already learned about. The following comparison shows where it fits.
Category
Characteristics
Example
Narrow AI
Excels at one specific task
A spam filter
General AI
Human-like reasoning across all domains
Does not yet exist
Rule-based
Follows explicit if-then rules
A tax preparation program
Learning-based
Improves from data over time
A recommendation engine
Agentic AI
Reasons, remembers, and acts toward goals
An AI agent that resolves customer issues end-to-end
Agentic AI systems are still narrow AI. They operate within specific domains. However, within those domains, they demonstrate a level of autonomy and capability that earlier narrow AI systems did not possess. They combine learning-based approaches (using LLMs for reasoning) with tool use and memory to achieve goal-directed behavior.
Understanding agentic AI is essential because it underpins the entire Amazon Connect portfolio of solutions you will explore throughout this badge readiness path.
RECALL & APPLY
Practice what you learned
Try answering in your own words, then compare with the suggested answer. These activities are for self-study and are not automatically graded.
01What turns a language response into an action?
Topic: Agentic AI capabilities
Reveal suggested answer
A tool or API performs an operation outside the model. The surrounding application controls permissions and execution.
02A refund assistant can explain policy and issue a refund. What must the design distinguish?
Topic: Agentic AI in the wider AI landscape
Reveal suggested answer
It must distinguish answering a question from executing a transaction, including authority, confirmation, and failure handling.
Capture a key idea, an example, or a question for your instructor.
Have you ever asked Amazon Alexa a question, used Amazon Bedrock to generate content, or seen Amazon Lex power a chatbot conversation? These advanced conversational AI tools are increasingly powered by large language models (LLMs), one of the most rapidly evolving AI technologies today. These systems can write essays, translate languages, create poetry, answer complex questions, and even write computer code. These AI tools can do all of this by predicting which text should come next in a sequence.
In this section, you will explore what makes LLMs work, such as the special terminology used to describe them, their capabilities, and important limitations. You will learn how these systems differ from other types of AI. You will also discover how phrasing your requests (prompts) dramatically affects the results.
Section objectives
In this section, you will learn to do the following:
Compare LLMs and traditional AI.
Identify the key terminology associated with LLMs and their functions.
Identify the capabilities and limitations of modern language models.
Identify how context and prompt engineering affect LLM performance.
Key Terminology, Capabilities, and Limitations
Key Terminology
When reading about LLMs, you will encounter several specialized terms. Understanding this vocabulary will help you make sense of how these systems work and how to use them effectively.
Prompt and prompt engineering
A prompt is the input text given to an LLM to elicit a response. It can be a question, instruction, partial text, or any other text that tells the model what kind of output you want. Examples include the following:
"What are three ways to improve sleep quality?"
"Write a poem about autumn leaves in the style of Robert Frost."
"Translate this sentence to Spanish: 'I would like to order dinner.'"
Prompt engineering is the process you use to guide generative AI solutions to generate your desired outputs using techniques such as the following:
Chain-of-thought prompting
Few-shot learning examples
System prompts compared to user prompts
Temperature and other generation parameters
Why it matters: How you phrase your prompt dramatically affects the quality and nature of the response you get from an LLM.
To learn more about prompt engineering, see What is Prompt Engineering?
Tokens
Tokens are the basic units that LLMs process. They are essentially the pieces the model breaks text into. A token might be the following:
A full word (sunshine)
Part of a word (sun + shine)
A character (!)
A space between words
For example, the sentence, "I love artificial intelligence!" might be broken into tokens like ["I", "love", "artificial", "intelligence", "!"].
Why it matters: When using LLMs, you will often see limits expressed in tokens rather than words or characters. For English text, one token is roughly three fourths of a word on average.
Embeddings
Embeddings are numerical representations of words or tokens. They capture meaning by positioning similar concepts near each other in a mathematical space. For example, in this mathematical space:
Dog would be close to puppy and canine.
King minus man plus woman might equal something close to queen.
Paris would be near France just as Tokyo is near Japan.
Why it matters: Embeddings allow LLMs to understand relationships between words and concepts, enabling them to generate coherent and contextually appropriate text.
Context Window
The context window refers to how much text the model can see and consider when generating a response. It is the maximum amount of text (in tokens) that can be processed at one time. Examples include the following:
Early LLMs might have had a context window of 512 tokens (about a page of text).
Modern LLMs might have context windows of 8,000, 32,000, or even over 100,000 tokens.
Why it matters: A larger context window allows the model to reference information from much earlier in a conversation or document. This makes it more consistent and capable of handling longer texts.
Parameters
Parameters are the adjustable values that determine an LLM's behavior, learned during training. The number of parameters often indicates a model's capability and complexity.
Examples include the following:
GPT-3: 175 billion parameters
GPT-4: Estimated 1.76 trillion parameters (though the exact number isn't public)
Why it matters: Understanding parameter counts helps compare models and understand their capabilities and limitations.
Comparing Training and Inference
Training is the process of creating the model by exposing it to vast amounts of text data, allowing it to learn patterns. Inference is when the trained model is actually used to generate text or provide responses to user inputs.
Why it matters: The training process requires enormous computing resources but happens once. Inference (using the model) requires far less computing power. It occurs each time someone interacts with the system.
Fine tuning
Fine-tuning is the process of additional training on specific data to adapt a general LLM for particular purposes or to follow certain guidelines. For example, a general LLM might be fine-tuned to do the following:
Specialize in medical information.
Follow a company's brand voice.
Provide more concise responses.
Avoid certain topics or phrasings.
Why it matters: Fine-tuning is how general-purpose LLMs become specialized tools for specific applications, improving their performance for particular use cases.
Hallucinations
Hallucinations occur when LLMs generate plausible sounding but false or inaccurate information. This is one of the most significant challenges with current LLM technology.
Examples include the following:
Inventing fake citations or research papers
Creating fictional historical events
Generating incorrect technical specifications
Why it matters: Understanding hallucinations is crucial for responsible LLM use, especially in professional or educational contexts where accuracy is essential. Preventing hallucinations requires guardrails like fact-checking, confidence thresholds, human oversight, and regular model evaluations.
Capabilities of LLMs
LLMs are capable of producing large increases in productivity when applied correctly to a use case. Knowing the capabilities of your LLM will help determine suitability for the intended actions.
Generate human like text
LLMs can produce coherent, fluent text in various styles, from academic papers to poetry to business emails. They can adapt their writing style based on instructions.
Summarize information
LLMs can condense long documents or articles into shorter summaries while preserving key points.
Answer factual questions
For many common questions, especially about well-documented topics, LLMs can provide accurate information they have encountered in their training data.
Follow complex instructions
Modern LLMs can understand and execute multi-step instructions or tasks described in natural language.
Creative writing
LLMs can generate creative content like stories, poems, songs, and scripts based on provided prompts or themes.
Code generation and explanation
Many LLMs can write computer code in various programming languages and explain how existing code works.
Language translation
LLMs can translate text between languages they have been trained on, often with good accuracy for common language pairs.
Analyze sentiment and tone
LLMs can analyze text to determine whether it expresses positive, negative, or neutral sentiments.
Limitations of LLMs
The way you ask a question to LLMs can dramatically change the quality of the answer you receive. Although AI assistants seem incredibly knowledgeable, they sometimes hallucinate information that is not true. Understanding limitations becomes a key consideration when deciding if an LLM is applicable for your use case.
Hallucinations
LLMs can confidently generate information that appears plausible but is actually incorrect or made up. They can create false references, invent statistics, or fabricate facts that sound convincing.
Limited knowledge cutoff
LLMs only know information included in their training data up to a certain date. They do not have access to real-time information or events that occurred after their training cutoff date unless they are connected to search tools.
No true understanding
Despite appearances, LLMs do not truly understand text the way humans do. They predict patterns rather than comprehend meaning, which leads to limitations in reasoning and common sense.
Difficulty with complex reasoning
Although they are improving, LLMs still struggle with multi-step logical reasoning, especially for math problems or situations requiring cause-and-effect analysis.
Bias and stereotyping
LLMs can reflect and amplify biases present in their training data, potentially producing outputs that contain stereotypes or unfair representations.
Contextual limitations
LLMs have a limited context window. They can only see a certain amount of text at once, causing them to forget or lose track of information in very long conversations.
No external verification
LLMs typically cannot check external sources to verify information unless specifically connected to other tools or the internet.
Lack of Self awareness
LLMs don't have true self-awareness or consciousness. They do not know what they know or recognize when they are making mistakes.
Hallucination example
This example shows a possible hallucination could occur in an LLM response.
User Query: "Tell me about the 2021 Mars landing of the Harmony spacecraft and its discovery of microbial fossils."
LLM Hallucinated Response: "The Harmony spacecraft landed on Mars in March 2021 as part of a joint NASA-ESA mission. Its rover, Concordia, made headlines when it discovered fossilized microorganisms in the Arcadia Planitia region. These 2.8 billion-year-old microfossils showed cellular structures similar to Earth's cyanobacteria but with unique Martian adaptations. Dr. Elena Matsuoka, the mission's lead scientist, called it 'definitive evidence of ancient life on Mars,' prompting a mission extension and the planned Harmony II follow-up mission set for 2026."
Explanation of the Hallucination: This response is entirely fabricated. There was no "Harmony spacecraft" that landed on Mars in 2021, no "Concordia rover," and no discovery of microbial fossils. The scientist mentioned and all details about the mission are fictional. This demonstrates how an LLM can confidently generate plausible but completely false information when responding to queries about non-existent events. To aid in the identification and reduction of hallucinations, Amazon Bedrock Guardrails may be used to reduce possible hallucinations. To learn more, visit the Amazon Bedrock Guardrails page.
RECALL & APPLY
Practice what you learned
Try answering in your own words, then compare with the suggested answer. These activities are for self-study and are not automatically graded.
01Why can a long conversation lose useful earlier details?
Topic: Large language model terminology
Reveal suggested answer
Earlier information may fall outside the context supplied to the model or receive insufficient attention. Applications must manage context deliberately.
02Does adding a document to a prompt retrain the model?
Topic: Context, parameters, and fine tuning
Reveal suggested answer
No. It supplies context for inference. Fine tuning changes model parameters through a separate training process.
03A fluent answer cites a policy that does not exist. What is the problem?
Topic: LLM capabilities and hallucinations
Reveal suggested answer
The answer is unsupported even though it sounds convincing. Check the actual policy source and correct the response.
04How should a support workflow handle an uncertain policy answer?
Topic: LLM limitations
Reveal suggested answer
Retrieve the approved source or involve a qualified agent. Do not treat confidence of wording as evidence of accuracy.
Capture a key idea, an example, or a question for your instructor.
The LLM landscape is evolving rapidly. Since you learned the fundamentals of large language models, several developments have changed how LLMs are deployed in real-world systems, particularly in contact centers. In this lesson, you learn about four key advances: voice-native models, standardized tool-use protocols, LLM-enhanced NLU, and the emergence of third-party model integration.
Amazon Nova Sonic
Traditional LLMs process text. Amazon Nova Sonic is different. It is a voice-native LLM designed from the ground up for natural, expressive voice interactions. Rather than converting speech to text, processing it, and converting back (which introduces latency and loses vocal nuance), Nova Sonic works directly with audio.
What makes it voice native
Nova Sonic processes speech directly as its input and output modality. This means it understands not just what someone says, but how they say it. Tone, pacing, emphasis, and emotion are all part of the model's understanding.
This eliminates the latency of traditional speech-to-text-to-LLM-to-text-to-speech pipelines and produces more natural conversational experiences.
Role in Amazon Connect Customer
Nova Sonic powers agentic self-service in Amazon Connect Customer. When a customer calls and interacts with an AI agent, Nova Sonic enables that agent to:
Understand spoken requests naturally, including complex or ambiguous phrasing
Respond with expressive, human-like speech in real time
Maintain conversational flow without awkward pauses or robotic tone
This is what makes the agentic self-service experience feel like talking to a capable colleague rather than navigating a menu system.
Compared to traditional approaches
Aspect
Traditional pipeline
Nova Sonic
Processing
Speech → text → LLM → text → speech
Speech → model → speech
Latency
Higher (multiple conversion steps)
Lower (direct processing)
Nuance
Lost in text conversion
Preserved in audio
Conversation feel
Robotic, transactional
Natural, expressive
Model Context Protocol
In the previous lesson on agentic AI, you learned that AI agents can use tools to take action. But how does an AI agent know which tools are available, what they do, and how to use them? The Model Context Protocol (MCP) solves this problem.
MCP is a standardized protocol that enables AI agents to interact with external tools and data sources. Think of it as a universal adapter. Just as USB provides a standard way to connect devices to a computer, MCP provides a standard way to connect AI agents to the tools they need.
How MCP works
MCP defines a structured way for AI agents to:
Discover what tools are available (for example, a CRM lookup tool, a refund processing tool, or a knowledge base search tool)
Understand what each tool does and what inputs it requires
Invoke tools with the correct parameters
Interpret the results returned by those tools
This standardization means the same AI agent can work with many different tools without needing custom integration code for each one.
Why MCP matters for contact centers
In Amazon Connect Customer, MCP enables AI agents to:
Retrieve customer information from knowledge bases
Look up order status in backend systems
Process returns or schedule appointments
Access multiple data sources in a single interaction
Without MCP, each tool connection would require custom development. With MCP, new tools can be added and immediately used by AI agents through the standard protocol.
MCP in the broader environment
MCP is not proprietary to Amazon Connect Customer. It is an open standard gaining adoption across the AI industry. This means skills and patterns learned here apply broadly, and third-party tools built to the MCP standard can be integrated into Amazon Connect Customer AI agents.
Amazon Lex assisted NLU overview
Traditionally, Amazon Lex used slot-based natural language understanding (NLU) to interpret customer intent. This approach required defining specific intents, sample utterances, and slots for each conversation path. Now, Amazon Lex offers LLM-enhanced NLU through a feature called Assisted NLU.
Amazon Lex now offers large language models as an enhanced option for understanding customer intent through Assisted NLU. In Primary mode, the LLM serves as the default means of processing user input. In Fallback mode, the LLM activates only when the traditional NLU confidence score is below threshold. Both modes coexist with the slot-based system rather than replacing it. This represents a significant evolution in how conversational AI systems are built.
Traditional Slot-Based NLU
Assisted NLU (LLM-Powered)
Impact on development
Traditional Slot-Based NLU With traditional NLU, developers must:
Define each possible intent manually (for example, "CheckBalance", "TransferFunds", "ReportLostCard")
Provide sample utterances for each intent (dozens of example phrases)
Define slots for each piece of information needed (account number, amount, date)
Handle edge cases with explicit fallback logic
This approach works well for predictable interactions but struggles with unexpected phrasing, complex requests, or conversations that span multiple intents.
Third party AI integration overview
Amazon Connect Customer no longer relies exclusively on AWS-native AI models. Third-party speech and voice AI providers can now be integrated alongside native services, giving organizations more choice in how they build voice experiences.
Two notable integrations are now available for self-service interactions:
Deepgram for speech-to-text (STT) — known for high accuracy and speed in transcribing spoken language
ElevenLabs for text-to-speech (TTS) — known for highly natural, expressive voice synthesis
This multi-model approach means organizations can select the best AI model for each specific task rather than being limited to a single provider. For example, a contact center might use ElevenLabs for its premium voice quality in customer-facing interactions while using Amazon Polly for internal notifications where cost efficiency is prioritized.
The ability to mix native AWS models with third-party providers reflects the broader industry trend toward composable AI architectures. These are systems built from best-of-breed components rather than monolithic platforms.
Amazon Lex Assisted NLU
Traditional Slot Based NLU
Traditional Slot-Based NLU With traditional NLU, developers must:
Define each possible intent manually (for example, "CheckBalance", "TransferFunds", "ReportLostCard")
Provide sample utterances for each intent (dozens of example phrases)
Define slots for each piece of information needed (account number, amount, date)
Handle edge cases with explicit fallback logic
Assisted NLU LLM Powered
Assisted NLU (LLM-Powered) With Assisted NLU enabled in Amazon Lex:
The LLM understands intent from natural language without needing exhaustive sample utterances
Complex, multi-part requests are interpreted correctly ("I want to check my balance and also dispute the charge from yesterday")
Unexpected phrasing is handled gracefully
Context from earlier in the conversation informs understanding
This dramatically reduces the development effort required to build conversational experiences and improves accuracy for real-world customer language. Amazon Lex remains the conversational AI service — the LLM is the engine powering its understanding.
Impact on development
Aspect
Slot-based NLU
LLM-powered NLU
Setup effort
High (define all intents, utterances, slots)
Lower (model generalizes from fewer examples)
Handling unexpected input
Poor (falls back to error)
Strong (interprets from context)
Multi-intent requests
Requires complex flow design
Handled naturally
Maintenance
Manual updates for new phrases
Adapts with minimal tuning
Third party AI model integration
Two notable integrations are now available for self-service interactions:
Deepgram for speech-to-text (STT) — known for high accuracy and speed in transcribing spoken language
ElevenLabs for text-to-speech (TTS) — known for highly natural, expressive voice synthesis
RECALL & APPLY
Practice what you learned
Try answering in your own words, then compare with the suggested answer. These activities are for self-study and are not automatically graded.
01What should you assess beyond the sound of a generated voice?
Topic: Amazon Nova Sonic and voice interaction
Reveal suggested answer
Assess response relevance, latency, interruption handling, task completion, and the behavior when the system cannot help.
02Does a standard tool interface make every action appropriate?
Topic: Model Context Protocol and tool access
Reveal suggested answer
No. The application still needs authorization rules, suitable input validation, and review of consequential actions.
03What should a test set include beyond the designer’s sample utterances?
Topic: Assisted NLU and model integration
Reveal suggested answer
Paraphrases, ambiguous requests, corrections, incomplete information, and unsupported requests.
04What stays important when an LLM improves intent recognition?
Topic: Traditional and assisted NLU
Reveal suggested answer
Required data collection, confirmation, business validation, and the action that fulfills the request.
Capture a key idea, an example, or a question for your instructor.
An LLM is an AI system trained on massive amounts of text data to recognize and produce human language. Think of an LLM as an extremely sophisticated text prediction system. It's similar to the autocomplete feature on your phone, but at a much larger scale and with far greater capabilities.
Why do they call them large
The large in the name refers to the following:
The vast amount of text data used to train them (trillions of words)
The number of parameters they contain (billions or trillions)
The extensive computing power required to build them
Understanding the purpose of LLMs
The primary purpose of LLMs is to understand and generate human language, allowing them to do the following:
Answer questions and provide information.
Create written content in various formats and styles.
Translate between languages.
Summarize long documents.
Engage in conversational interactions.
Assist with writing and editing tasks.
Generate creative content like stories or poetry.
How LLMs work
At a basic level, LLMs work by predicting the next word or token in a sequence based on all the previous words. Through their training, they have learned patterns and relationships between words and concepts across billions of examples. For example, if you type "The capital of Spain is," an LLM would predict "Madrid" as the next word because it has seen this pattern many times in its training data. Modern LLMs become useful by maintaining prediction capabilities over extended text, enabling coherent generation of paragraphs and longer content.
Figure 11 How LLMs workSelect image to enlarge
How LLMs differ from traditional AI systems
The AI landscape has transformed dramatically with the emergence of LLMs. These powerful systems represent a fundamental shift from traditional AI approaches in several key ways. LLMs process information differently, learning from vast amounts of text rather than following explicit programming. Their ability to understand context and generate human-like responses sets them apart from their predecessors.
General language ability
Traditional AI systems are typically designed for specific, narrow tasks with predetermined functions and limited flexibility. These specialized systems excel at their designated purposes but cannot efficiently adapt to different contexts or tasks.
Examples of narrow-task traditional AI include the following:
Image recognition systems that identify objects in photos but cannot explain the content
Recommendation engines that suggest products based on user data but cannot justify their selections
Fraud detection systems that flag suspicious transactions without explaining their reasoning
In contrast, LLMs possess general language capabilities that enable them to perform across diverse domains without task-specific programming. A single LLM can do the following:
Craft creative content like poems, stories, or marketing copy.
Explain complex scientific concepts in accessible terms.
Generate professional communications, such as emails or reports.
Translate between languages while preserving meaning.
Answer questions across numerous knowledge domains.
This versatility stems from LLMs' foundational understanding of language patterns and contextual relationships rather than from specialized programming for each individual task.
Learning approach
Traditional AI systems often require task-specific training data and features explicitly defined by humans. LLMs use a different approach with pre-training on vast general text data followed by fine-tuning. Traditional AI systems often require the following:
Task-specific training data
Features explicitly defined by humans
Separate models for different tasks
LLMs use a different approach, such as the following:
Pre-training on vast general text data followed by fine-tuning
Self-learning of features and patterns from the data
Adapting one model to many tasks
Input and output flexibility
Traditional AI systems typically have structured inputs and outputs. LLMs are more flexible, accepting natural language instructions and generating varied outputs based on prompts.
Traditional AI systems typically have structured inputs and outputs, such as the following:
Predefined categories (for classification tasks)
Numerical predictions (for regression tasks)
Specific data formats required
LLMs are much more flexible because they can do the following:
Accept natural language instructions.
Generate varied outputs based on how they are prompted.
Adapt to different formats and styles of communication.
Context processing
Traditional AI often processes each input independently in the following ways:
Each image is classified separately.
Each transaction is evaluated on its own.
LLMs can maintain context over extended interactions by doing the following:
Remembering earlier parts of a conversation
Building on previously mentioned information
Maintaining consistency across a generated text
Trade offs Between Generative AI, LLMs, and Traditional AI
Although generative AI and LLMs offer impressive capabilities beyond traditional AI systems, important trade-offs exist.
Costs
LLMs require substantial computing resources for training and inference, making them more expensive to develop and deploy than many traditional AI solutions.
Task suitability
Traditional AI systems remain more efficient and reliable for specific structured tasks, especially those requiring precise, deterministic outcomes or operating in data-constrained environments.
Possibility of hallucinations
LLMs are prone to hallucinations and can generate plausible-sounding but incorrect information. This creates risks traditional rule-based systems typically don't face.
Interpretability
Traditional AI often offers clearer decision paths, whereas LLMs operate as complex black boxes that are difficult to audit.
Streamlined implementation
Existing traditional AI technologies offer proven, lightweight solutions for many business needs without the complexity of implementing cutting-edge LLMs.
Despite their challenges, LLMs offer tremendous value through unmatched natural language processing (NLP) and creative problem-solving capabilities. Their cross-domain adaptability replaces multiple specialized systems, and human-like interactions enhance user experiences. Ongoing advancements promise improved efficiency and reliability, making these trade-offs increasingly worthwhile for many applications.
RECALL & APPLY
Practice what you learned
Try answering in your own words, then compare with the suggested answer. These activities are for self-study and are not automatically graded.
01Why can one language model support both summarization and translation?
Topic: What an LLM does
Reveal suggested answer
Training captures language patterns that can be applied through different prompts, though quality still varies by task and language.
02When might a specialized model be preferable?
Topic: LLMs and traditional AI
Reveal suggested answer
When the task has a narrow output, strong accuracy requirements, or resource constraints that a specialized solution can meet more directly.
03What evidence should support choosing an LLM for a workflow?
Topic: AI solution tradeoffs
Reveal suggested answer
A relevant evaluation comparing output quality, operational cost, latency, and failure behavior against the available alternatives.
Capture a key idea, an example, or a question for your instructor.
Context in LLMs refers to all the text the model can see when generating a response, including the following:
Your current query or instruction
Previous messages in a conversation
Any additional information you have provided
System instructions (often invisible to users)
The model can only respond based on this visible context. It cannot recall previous conversations unless you explicitly refer to them or they are still within the context window.
Understanding prompt engineering
Prompt engineering is the practice of crafting inputs to get the best possible outputs from language models. It is essentially learning how to talk to AI systems effectively. The way you phrase your requests to an LLM, known as prompt engineering, dramatically affects the quality, accuracy, and usefulness of the responses you receive.
The following are basic prompt engineering techniques.
Be specific and clear
Specific, clear prompts eliminate ambiguity, producing desired AI outputs efficiently while reducing the need for repeated refinements. Review the following example:
Vague: "Tell me about planets."
Specific: "Explain the key differences between rocky and gas-giant planets in our solar system in basic terms for a 10-year-old."
Provide examples Few shot learning
Showing examples helps the model understand the pattern you want. Review the following example:
Convert these sentences to past tense:
"I walk to the store" becomes "I walked to the store."
"She runs quickly" becomes "She ran quickly."
"They build a house" becomes "They built a house."
Specify format and length
Tell the model how you want information structured. Review the following example:
Create a three-column table comparing apples, oranges, and bananas, based on the following:
1) Nutritional benefits
2) Growing conditions
3) Common varieties Keep each cell to 15 words or fewer.
Use role prompting
Asking the model to adopt a perspective or role can shape its response. Review the following example:
As an experienced math teacher helping a struggling eighth grader, explain how to solve for x in the equation 3x + 7 = 22.
Advanced prompt engineering techniques
Advanced techniques move beyond basic instructions to use deeper understanding of how language models process information and generate responses. These techniques can help you achieve more nuanced, accurate, and contextually appropriate outputs while reducing common issues like hallucinations or inconsistencies.
Chain-of-thought prompting
System instructions
Controlling creativity and precision
Handling hallucinations
Common prompt engineering mistakes
Even experienced users can fall into common pitfalls when crafting prompts for AI systems. Understanding these frequent mistakes is crucial for improving your prompt engineering skills and getting more reliable results. By learning to recognize and avoid these common errors, you can save time, reduce frustration, and achieve more consistent outcomes in your interactions with AI.
The following are the most prevalent mistakes and how to address them:
Being too vague: Vague prompts lead to generic responses. Add specificity about audience, purpose, tone, and format.
Overloading with instructions: Too many contradictory or complex requirements can confuse the model. Focus on the most important instructions.
Not iterating: Prompt engineering often requires refinement. Adjust your prompt and try again when responses are unsatisfactory.
Forgetting about context limitations: Remember that models can only see a limited amount of text. Very long conversations might lose important context from earlier messages.
Advanced prompt engineering examples
Stepwise problem solving
Encourage the model to show its reasoning process in the following way:
Question: A shirt costs $25. If it's discounted by 20 percent and then there's an additional 10 percent off, what is the final price?
Think through this step by step.
System instructions
Many LLM applications allow setting system-level instructions that shape all responses. The following is an example:
You are a helpful assistant that specializes in explaining scientific concepts using everyday analogies. Keep explanations under 100 words.
Controlling creativity and precision
Some systems allow adjusting temperature settings, such as the following:
Higher temperature: More creative and varied, but potentially less accurate responses
Lower temperature: More predictable and conservative, and often more factual responses
Handling hallucinations
When accuracy is critical, you can add instructions like the following:
If you're unsure of any information, please explicitly state that you don't know rather than guessing.
RECALL & APPLY
Practice what you learned
Try answering in your own words, then compare with the suggested answer. These activities are for self-study and are not automatically graded.
01Why does “help this customer” produce inconsistent answers?
Topic: Context and clear prompts
Reveal suggested answer
The instruction omits the customer’s issue, approved information, the assistant’s role, and the required outcome.
02How would you improve “summarize this call”?
Topic: Examples and output requirements
Reveal suggested answer
Specify the audience and fields, such as customer issue, actions taken, outcome, and next step. Provide the transcript and any required length limit.
Have you ever wondered how Amazon Music seems to know exactly what podcast you might enjoy listening to next? Or how recipes are shown on your Alexa that you might want to cook? These everyday conveniences are powered by AI and ML technologies that have quietly transformed how businesses operate and how we interact with digital services.
Section objectives
In this section, you will learn to do the following:
Identify how AI and ML are transforming business decision-making processes.
Describe natural language processing applications in everyday use.
Describe how AI enables creative content generation across different media.
Identify automation and virtual assistant applications across industries.
Business Decision Making, Virtual Assistants, and Recommendations
AI Powered Business Intelligence
Businesses today face a flood of data, such as sales figures, customer behavior, market trends, and inventory levels. Making sense of all this information would be difficult for humans alone. That is where AI-powered business intelligence comes in.
AI transforms business decisions by analyzing vast datasets to reveal hidden patterns and trends. Companies use these insights to optimize operations and predict market changes with unprecedented accuracy. AI systems help businesses turn raw data into actionable insights through several key capabilities.
Data analysis at scale
AI systems help businesses turn raw data into actionable insights by analyzing massive datasets far beyond what human analysts could process manually.
These systems can identify patterns across millions of transactions or customer interactions, detecting trends that would be impossible to spot through manual analysis.
Predictive analytics
AI does not just analyze past data; it can predict future trends. The following are examples of predictive analytics:
Sales forecasting: Based on current patterns, we expect a 12 percent increase in winter coat sales in the Northeast region.
Customer behavior: Customers who purchase this product typically return within 45 days to buy these complementary items.
Risk assessment: This loan applicant has an 85 percent probability of repayment based on our model.
Automated reporting and visualization
AI systems can automatically generate reports and visual representations that make complex data understandable.
These tools can create custom dashboards that update in real-time, highlighting the most relevant information for different stakeholders and adapting to changing business conditions.
Real world examples
The following industry examples illustrate how AI systems help business:
Retail: Retailers use AI to optimize inventory levels, preventing both shortages and excess stock. This is done by analyzing historical sales data, seasonal trends, local events, weather forecasts, and social media buzz.
Healthcare: Hospitals use AI systems to improve operations and patient care by predicting patient admission rates to staff appropriately, optimizing surgery schedules, and analyzing treatment outcomes.
Financial services: Banks and investment firms rely on AI for numerous decision-support functions including algorithmic trading, credit scoring models, and anti-fraud systems.
Virtual Assistants and Conversational AI
Customer service bots are AI systems that handle inquiries through advanced language processing algorithms. These digital assistants analyze context and access databases to provide relevant responses to common questions. When faced with complex issues, they escalate to human agents while continuously learning from interactions.
Voice assistant request processing cycle
Audio input
The voice assistant activates and performs the following actions to capture the user's speech through a microphone:
The system records the raw audio waveform of your voice.
Background noise is filtered out using noise cancellation algorithms.
Voice Activity Detection (VAD) identifies when the user starts and stops speaking, so the system knows when a complete utterance has been captured.
Example: User says, "What's the weather like today?"
Speech to text
The audio recording is converted into written text with the following actions:
The audio is broken into small segments (usually 10-20 milliseconds each).
These segments are analyzed to identify phonemes (speech sounds).
Phoneme sequences are compared against language models to recognize words.
The system produces a text transcript of what was said.
Example: The audio is input as text, "What's the weather like today?"
Natural Language Processing
The system uses the following actions to analyze the text to understand its meaning and context:
The text is broken down into tokens (words and punctuation).
Part-of-speech tagging identifies nouns, verbs, adjectives, and so forth.
Dependency parsing determines relationships between words.
Named entity recognition identifies specific objects, places, or concepts.
The system builds a semantic representation of the user's request.
Example: The system recognizes that weather and today are the key entities.
Intent classification
The system uses the following actions to determine what the user is trying to accomplish:
Machine learning models classify the request into predefined categories.
The system identifies the user's goal or intent.
Confidence scores are calculated for different possible intents.
The highest-scoring intent is selected.
Example: The intent is classified as weather inquiry with location as current and time as today.
Response generation
The system creates an appropriate answer to the user's request with the following actions.
Based on the identified intent, the system retrieves necessary information.
For a weather request, it connects to a weather service API.
Data is gathered, such as temperature, conditions, and forecast.
A natural-sounding response is constructed following response templates.
Example: System generates, "The current temperature is 72 degrees with partly cloudy skies."
Text to speech
The text response is converted into spoken audio with the following actions.
Text is broken down into phonetic representations.
A voice synthesis model generates natural-sounding speech.
Prosody (rhythm, stress, intonation) is applied to sound more human.
The audio response is played through speakers.
Example: "The current temperature is 72 degrees with partly cloudy skies" is spoken aloud.
Speech to speech
The spoken input is converted directly into spoken output with the following actions:
Audio input is captured and converted into phonetic representations.
A voice conversion model processes the speech patterns and applies target voice characteristics.
Prosody (rhythm, stress, intonation) from the original speech is preserved or modified as needed.
The transformed audio is generated and played through speakers.
Example: A person says "Hello, how are you today?" in English, and the response is converted to sound like a different speaker. The response may be in a different accent while maintaining the original meaning and emotional tone.
Recommendation systems and personalization
If you have ever been pleasantly surprised by a Recommended for You suggestion that perfectly matched your tastes, you've experienced an AI recommendation system in action.
These systems work by using the following:
Collaborative filtering: This approach finds patterns based on user similarity, such as "Users who liked the same movies also enjoyed this one."
Content-based filtering: This approach analyzes the characteristics of items, such as "Because you enjoyed action movies with strong female leads, you might like this film."
Hybrid approaches: Most modern recommendation systems combine multiple methods for better results.
Some real-world applications of these systems include the following:
Streaming media: Streaming provider customers rely on content recommendations for their next binge set.
Ecommerce: The Frequently bought together suggestions on Amazon.com significantly impact purchasing decisions.
Social media: Platforms use AI to personalize content feeds based on your interactions.
News and content: News sites and apps personalize what stories appear first based on your interests.
Speech to speech
Speech-to-speech AI represents the next evolutionary leap, eliminating the text intermediary altogether. Instead of converting text to speech, these advanced systems can take spoken input in one language and convert it directly to spoken output in another language. The technology enables natural conversations across language barriers without requiring text intermediaries or causing significant delays.
For example, an executive can now speak to an international team in real time, preserving voice inflections while delivering the message in the team's language.
RECALL & APPLY
Practice what you learned
Try answering in your own words, then compare with the suggested answer. These activities are for self-study and are not automatically graded.
01How does a forecast differ from a guaranteed outcome?
Topic: AI in business decisions
Reveal suggested answer
A forecast is an estimate based on data and assumptions. Teams should inspect uncertainty and update their plans as evidence changes.
02At which step would a weather assistant obtain current weather data?
Topic: The voice assistant processing cycle
Reveal suggested answer
The application retrieves data after identifying the request. Speech recognition alone does not supply the weather information.
03Why might a recommendation system combine several methods?
Topic: Speech interaction and recommendations
Reveal suggested answer
Each method captures different evidence. Combining them can address missing data or improve relevance for different users and items.
Capture a key idea, an example, or a question for your instructor.
Earlier in this section, you explored how AI and ML are transforming business decision-making, powering virtual assistants, and enabling creative applications. Now you will learn how agentic AI is being applied across entire industries through the Amazon Connect portfolio.
In the first half of 2026, AWS announced the expansion of Amazon Connect Customer from a single contact center product into four agentic AI solutions, each purpose-built for a specific industry. These solutions demonstrate how the same underlying AI capabilities (reasoning, memory, action, and tool use) take different forms depending on the domain.
Four industry solutions
The following four solutions show how agentic AI adapts to specific industry needs. Each solution uses specialized AI agents trained for its domain.
Amazon Connect Health
Amazon Connect Health deploys five specialized AI agents for healthcare organizations:
Patient verification — confirms identity securely across channels
Appointment management — schedules, reschedules, and sends reminders
Patient insights — surfaces relevant medical history during interactions
Medical coding — assigns appropriate codes from clinical notes
Key characteristics:
HIPAA-eligible by design
Integrates with Electronic Health Record (EHR) systems
AI agents operate within strict compliance boundaries
This demonstrates how agentic AI can function in highly regulated environments where accuracy and privacy are non-negotiable.
Amazon Connect Decisions
Amazon Connect Decisions applies agentic AI to supply chain operations, combining:
30 years of Amazon's operational science
25+ specialized AI tools for logistics and planning
SCOT (Supply Chain Optimization Technologies) foundation models
AI agents in this solution can reason about complex supply chain scenarios such as inventory positioning, demand forecasting, and route optimization. They can recommend or take action based on real-time data.
This demonstrates how agentic AI handles problems with many variables, constraints, and interdependencies that would overwhelm traditional rule-based systems.
Amazon Connect Talent
Amazon Connect Talent applies agentic AI to the hiring process:
Every candidate receives the same fair, standardized experience
Scoring is based on competencies, not subjective impressions
This demonstrates how agentic AI can bring consistency and scale to processes that traditionally depended on individual human judgment. At the same time, it maintains the human-like quality of voice interaction through Nova Sonic.
Amazon Connect Customer
Amazon Connect Customer is the evolution of the original Amazon Connect contact center product, now rebranded and enhanced for agentic self-service.
At its core, Amazon Connect Customer enables AI agents powered by Nova Sonic to handle complex customer requests end-to-end — reasoning through multi-step problems, accessing tools, and resolving issues autonomously. To make building these experiences accessible, the platform includes a no-code conversational AI canvas (via NLX) that lets designers visually compose self-service flows without writing code. This combination of powerful AI reasoning and visual design tooling means organizations can deploy sophisticated conversational experiences rapidly.
This is the solution most directly relevant to the contact center use cases you will explore in detail throughout the remaining courses in this badge path.
Common patterns across solutions
Although each solution targets a different industry, they share common agentic AI characteristics.
Pattern
Description
Example
Specialized AI agents
Purpose-built agents for specific tasks
Health's five distinct agents
Domain knowledge
Models trained on industry-specific data
Decisions' SCOT foundation models
Compliance by design
AI operates within regulatory boundaries
Health's HIPAA eligibility
Human-like interaction
Natural voice via Nova Sonic
Talent's structured interviews
Autonomous action
Agents complete tasks end-to-end
Customer's self-service resolution
The pattern you should recognize is this: agentic AI is not one solution applied generically. It is a design philosophy built on reasoning, memory, and action. It takes different shapes depending on the domain's unique requirements, constraints, and data.
RECALL & APPLY
Practice what you learned
Try answering in your own words, then compare with the suggested answer. These activities are for self-study and are not automatically graded.
01Why is a generic agent insufficient for every industry?
Topic: Agentic AI across industries
Reveal suggested answer
Domain workflows, data access, acceptable errors, and evaluation requirements differ. The agent must be configured for its intended use.
02What do these examples have in common?
Topic: Amazon Connect portfolio examples
Reveal suggested answer
They apply agentic capabilities to a defined domain using relevant information, tools, and operating constraints. Discuss these as the course’s portfolio examples.
03What would you change when moving an agent from support to supply chain planning?
Topic: Shared patterns across agentic solutions
Reveal suggested answer
Change the domain information, tools, permissions, objectives, and test cases. Reusing the general architecture does not validate the new use case.
Capture a key idea, an example, or a question for your instructor.
AI's growing ability to generate creative content represents one of the most fascinating developments in modern technology. These systems now write stories, compose music, and create visual art that rivals human-made work. The boundaries between human and machine creativity continue to blur as AI design tools become more sophisticated and accessible.
Text generation
AI systems can now produce various types of written content, including the following:
Creative writing, such as stories, poetry, scripts, and lyrics
Practical writing, such as marketing copy, blog posts, and educational materials
Prompt: "Create me an example of text generation to showcase creative writing with a maximum of 100 words"
Output: In Eldergrove, clockmaker Orion crafted timepieces infused with emotions. A mysterious woman sought a clock to rewind a cherished moment, paying with a memory. He created a moonlit marvel. She vanished, leaving jasmine's scent. Though heartbroken, Orion found joy in his work. Tales of his magical clocks spread, immortalizing love's fleeting beauty. Time, in Eldergrove, was rewritten, not just measured.
Image generation
AI can create images based on text descriptions using technologies. These systems can produce photorealistic images of scenes that don't exist, artistic renderings in specific styles, product visualizations, and character designs.
Prompt: "Create me an artist impression of an Eldergrove from the provided story."
Video generation
AI video generation tools like Amazon Nova Reel can now create short video clips or even longer sequences based on text descriptions or image inputs. These systems can produce animated scenes, transform still images into motion, generate realistic human movements, and even create entire short films. Although still developing, this technology is rapidly advancing, offering new possibilities for filmmakers, marketers, and content creators.
Prompt: "Create me a video of the story from the attached image."
Conversational AI using speech to speech
AI now engages in natural voice conversations with real-time processing, maintaining context across multi-turn dialogues, and adapting speaking style to match user preferences. Voice-based systems handle interruptions, emotional nuances, and conversational flow while preserving the immediacy of spoken communication. The technology creates natural human-machine conversation through instantaneous voice processing, dynamic emotional expression, and fluid dialogue that mirrors human speech patterns.
This audio is a conversation between a fictitious user and Amazon Nova Sonic. Nova Sonic is a conversational AI model that processes speech-to-speech. The first voice represents the user asking a question and the second voice is Nova Sonic providing a response.
Transcript Output
User: Hi Nova Sonic, please tell me about yourself.
Nova Sonic: Hey there, I'm an AI system here to assist you with any questions or tasks you have. I think of me as your friendly tech buddy.
RECALL & APPLY
Practice what you learned
Try answering in your own words, then compare with the suggested answer. These activities are for self-study and are not automatically graded.
01What makes a useful creative generation prompt?
Topic: Creative AI applications
Reveal suggested answer
A clear purpose, audience, medium, style direction, and constraints. Evaluate the output against those requirements.
Capture a key idea, an example, or a question for your instructor.
AI is transforming physical systems and automation, enabling robots and machines to perform increasingly complex tasks with greater flexibility and less human intervention.
Review the following to learn more about the various ways robotics and automation are being applied.
Industrial robotics and manufacturing
Modern manufacturing uses AI-enhanced robots for the following:
Adaptive assembly: AI-powered robots can recognize part variations and adjust their movements.
Quality control: AI vision systems inspect products with greater accuracy than human inspectors.
Predictive maintenance: AI analyzes sensor data to predict equipment failures before they happen.
Example: Modern automotive manufacturing plants use AI-powered robots that can work collaboratively alongside human workers, adapting to different car models without reprogramming.
Warehouse and logistics automation
Ecommerce and logistics companies use AI-powered systems for the following:
Autonomous mobile robots (AMRs): Warehouse robots that navigate dynamically around obstacles and people
Intelligent sorting systems: AI-powered conveyor systems that identify and route packages
Last-mile delivery: Autonomous vehicles, drones, and sidewalk robots for final delivery
Example: Amazon employs more than half a million mobile robots in its fulfilment centers to help store, sort, and retrieve products.
Consumer and service robotics
The following are examples of how service robots aid humans in a variety of ways:
Home robots: Robot vacuums that map homes and optimize cleaning paths
Healthcare robotics: Surgical assistance robots with enhanced precision
Customer service robots: Hotel and retail robots that assist customers
Example: The da Vinci Surgical System provides surgeons with enhanced precision and control. It translates the surgeon's hand movements into smaller, more precise actions performed by miniature instruments inside the patient's body.
The future of robotics
Emerging developments in robotics include the following:
Advanced dexterity: Robots that can manipulate delicate or irregular objects
Social robotics: Robots designed for natural human-robot interaction
Field robotics: Autonomous systems for agriculture, construction, and exploration
Multi-robot coordination: Swarms of robots working together on complex tasks
Example: Modern autonomous robots demonstrate advanced mobility and adaptability, navigating complex terrain and performing dynamic tasks that were impossible for earlier robotics systems.
AI is revolutionizing robotics and automation across industries, from manufacturing floors to warehouses, homes, and hospitals. These applications represent much more than technological novelties. They are transforming business operations, enhancing human capabilities, and creating new possibilities for customer experiences.
RECALL & APPLY
Practice what you learned
Try answering in your own words, then compare with the suggested answer. These activities are for self-study and are not automatically graded.
01How does a physical robot’s failure differ from an incorrect text answer?
Topic: Robotics and automation
Reveal suggested answer
Physical actions can affect people and equipment directly. The design needs appropriate sensing, operating limits, and failure handling.
Capture a key idea, an example, or a question for your instructor.
Have you ever wondered if the algorithms that recommend products to you online might treat some customers differently? Or whether an AI making hiring suggestions might favor certain types of candidates? As AI becomes more powerful and widespread, important ethical questions arise about how these systems work and their impact on society.
In this section, you will explore the ethical considerations that come with AI development and deployment. You will examine issues of bias and fairness, and look at how AI systems affect our privacy. You will become familiar with concerns about LLMs and consider questions about aligning AI with human values. Understanding these ethical dimensions is crucial for everyone living in a world increasingly shaped by this technology.
Section objectives
In this section, you will learn how to:
Identify key ethical challenges related to bias and fairness in AI systems.
Identify fundamental data privacy concerns in AI development and use.
Identify specific ethical issues associated with LLMs.
Describe concepts of AI alignment, transparency, and social impact.
Bias and fairness
Bias and fairness
When a bank's AI approves loans unevenly across similar groups, is it perpetuating bias or detecting real patterns? When your voice assistant records conversations for improvement, who can access this data and where is it stored?
These ethical dilemmas emerge as AI becomes increasingly woven into our everyday experiences.
Bias and fairness concepts
AI systems can reflect, amplify, or even introduce biases that affect different groups of people unfairly. Biases include the following:
Training data bias: AI systems learn from historical data, which often contains existing societal biases. For example, a hiring algorithm trained on past hiring decisions might perpetuate gender or racial disparities.
Selection bias: The way data is selected or collected can skew results. For example, customer feedback collected only through smartphone apps excludes people without smartphones.
Measurement bias: The method you choose to measure success can create bias. For example, a content recommendation system optimized for engagement might promote controversial or inflammatory content.
Algorithm design bias: The choices made when designing AI systems can introduce bias, such as including certain features while excluding others, which can impact different groups differently. For example, an algorithm trained mostly on male patient data might miss certain diagnoses in women due to their different symptom patterns.
RECALL & APPLY
Practice what you learned
Try answering in your own words, then compare with the suggested answer. These activities are for self-study and are not automatically graded.
01A survey reaches only mobile app users. What bias could follow?
Topic: Ethics, bias, and fairness
Reveal suggested answer
Selection bias. Customers who do not use the app are missing, so the findings may not represent the full customer population.
Capture a key idea, an example, or a question for your instructor.
Elevate your customer engagement strategy with the power of AI in Amazon Connect. This comprehensive course explores innovative AI capabilities that transform how businesses interact with customers across their contact center operations. Discover how to create unified customer profiles with automated data mapping, resolve identity concerns, and design targeted outbound campaigns using natural language segment creation. You will learn to use AI for more personalized customer interactions, improved first-contact resolution, and more efficient campaign management.
In this course, you will learn to do the following:
Implement AI capabilities in Amazon Connect to enhance customer engagement across contact center operations.
Configure AI-powered Amazon Connect Customer Profiles and Identity Resolution to create comprehensive unified customer views.
Design targeted outbound campaigns using AI-driven segmentation and natural language tools.
Measure the business impact of AI-powered customer engagement solutions on operational efficiency and customer experience.
Introduction to AI Powered Customer Engagement
Section objectives
In this section, you will learn to do the following:
Compare traditional contact centers with AI-enhanced engagement models to identify key opportunities for transformation.
Implement AI-powered features to recognize customers across channels and deliver personalized experiences.
Evaluate how the partnership between AI and human agents improves key business metrics and customer satisfaction.
Section introduction
Have you ever called customer service and repeated your information multiple times to different agents? Or have you received generic marketing messages that had nothing to do with your needs?
In this section, you will explore how Amazon Connect AI capabilities are reshaping customer engagement and why these capabilities matter for your business. You will discover how AI helps businesses recognize customers across different channels. This helps businesses understand customer history and needs, so they can engage with customers in more relevant ways. Whether you are managing a support team or designing customer journeys, understanding these AI foundations is valuable. They will help you deliver experiences that feel personalized and efficient instead of robotic and frustrating.
Customer Data in the Contact Center
Traditional contact centers operate with a major constraint: fragmented customer information. Agents often need to juggle multiple systems to gather basic information about who they are talking to and why.
This can create a number of problems with the contact center experience, including the following:
Disjointed customer experience –When customer data lives in separate systems, agents spend valuable time switching between applications instead of helping customers. By the time the agent has gathered the necessary information, the customer is already frustrated.
Repetitive customer identification –One of the most frustrating experiences for customers is having to repeatedly identify themselves as they are transferred between departments or agents. This happens when contact centers lack a unified customer profile that travels with the customer throughout their journey.
Generic service instead of personalized engagement –Without a complete picture of the customer, agents default to generic service scripts instead of tailored interactions. They miss opportunities to reference previous purchases, anticipate needs, or acknowledge loyalty.
Real-world example: You recently moved and updated your address with a company's billing department. When you call customer services about a different issue, the agent asks you toconfirm your address. However, the agent still sees your old address because the billing system and customer services system are not sharing information.
Contact center evolution
Contact centers have evolved through several stages over the years as explained in the following:
1970s 1990s
Basic Call Centers The Phone Only Era
Phone-based support with minimal customer data.
Basic call centers revolutionized customer service by moving interactions from stores to centralized telephone support. This transformation standardized service delivery across entire organizations. Companies achieved significant operational savings while maintaining consistent customer communication for the first time.
Customer viewpoint: Why am I on hold for 20 minutes?
1990s 2010s
Multichannel Contact Centers Breaking Down the Walls
Added email, chat, and social media, but often as separate operations.
Multichannel centers fueled ecommerce growth by supporting customers through emerging digital channels. New service expectations developed as consumers demanded help beyond traditional business hours. Companies built specialized teams for each channel, which created new career paths in digital customer engagement.
Customer viewpoint: Can I just email you instead of calling?
2010 2020
Omnichannel Contact Centers Making Connections
Connected channels but still with fragmented customer data.
Omnichannel approaches elevated customer experience as a primary business differentiator beyond price and product. Organizations began measuring success through customer satisfaction metrics instead of operational efficiency alone. Seamless experiences across touchpoints became essential as consumer loyalty increasingly depended on effortless service interactions.
Customer viewpoint: Why do I need to repeat myself every time I contact you?
2020+
AI Powered Contact Hubs The Intelligent Revolution
Intelligent AI-powered systems that unify customer data and provide AI-assisted insights.
AI-powered hubs transform service operations from reactive cost centers into proactive revenue generators. Predictive capabilities empower companies to address customer needs before problems fully develop. Human agents now focus on complex problem-solving while technology handles routine interactions. Service insights drive product development and marketing strategies across the entire organization.
Customer viewpoint: The system already knew what I needed before I explained it.
Unified customer data
The foundation of effective customer engagement is having a complete, unified view of each customer by bringing all customer information together. When properly implemented, this comprehensive customer profile becomes the central hub from which meaningful interactions flow.
A unified customer profile includes data such as the following:
Basic identity information – Names, contact details, account numbers, and authentication history
Purchase and billing history – Complete transaction records, subscription status, and payment preferences
Previous interactions across all channels – Conversation history from calls, chats, emails, and social media engagements
Service cases and their resolutions – Past issues, how they were resolved, and follow-up outcomes
Preferences and behavioral patterns – Communication preferences, product usage habits, and engagement tendencies
Without this unified foundation, customer data remains trapped in isolated systems, which creates fragmented experiences. When agents lack access to complete information, customers face repetitive questions and inconsistent service.
Even the most sophisticated AI tools will struggle to deliver meaningful improvements in customer engagement when working with disjointed data. Personalization begins with this consolidated view that empowers both human agents and AI systems to understand the full customer context.
Avoid these critical engagement pitfalls
Before revolutionizing your customer experience strategy, be aware of the following crucial pitfalls that even seasoned organizations could encounter:
Never launch new engagement channels without establishing proper data infrastructure first.
Do not rush into automation solutions while your customer data remains fragmented across systems.
Remember that without a comprehensive view of the entire customer journey, your personalization efforts will fall short.
By addressing these fundamental challenges before implementing advanced technologies, you can build a solid foundation that maximizes the return on your customer experience investments.
Personalizing Customer Interactions
AI is transforming customer interactions from generic scripts to personalized conversations. Amazon Connect AI capabilities help businesses recognize customers instantly, understand their needs more deeply, and deliver more relevant experiences.
Beyond basic recognition
Identity Resolution in Amazon Connect uses both rule-based and machine learning (ML) approaches to identify and consolidate matching customer profiles. This helps create a unified view of the customer across channels and systems as follows:
Rule-based matching – Uses exact matching on key identifiers such as phone numbers, email addresses, and account numbers
ML matching – Goes beyond basic rules to detect patterns and relationships that indicate the same customer across different records
With Amazon Connect Identity Resolution, businesses can maintain conversation continuity across channels. This helps honor preferences consistently regardless of how customers identify themselves. Companies can deliver truly personalized experiences without requiring customers to repeatedly explain who they are.
Identity Resolution in action
Customer Profile Alejandro "Ale" Rosalez
Alejandro has used both his personnel email and work email when contacting your business. He has also contacted your business using both the name Alejandro and the name Ale.
Alejandro "Ale" Rosalez
Traditional systems would create separate records for Alejandro and name Ale and each email used. This fragments the customer's history across multiple profiles.
Identity Resolution in Amazon Connect can recognize these are likely the same person based on multiple data points. The system analyzes behavior patterns, device information, and location data.
The numbered markers in the following figure show the identity resolution process.
Figure 12 Personalizing Customer InteractionsSelect image to enlarge
1 - Ingest data
Collect customer data from all sources, with real-time and batch data transfer through software as a service (SaaS), database, and data warehouse connectors.
2 - Map to a profile
Build and update customer profiles in real time as data is ingested.
3 - Resolve identities
Use ML-powered identity resolution.
4 - Merge and enrich profiles
Merge duplicate records using configurable exact and probabilistic matching.
Use out-of-the-box insights and segments from customer interactions.
AI enhanced context for better first contact resolution
AI capabilities of Amazon Connect systematically analyze multiple data points to provide agents with comprehensive context at the moment of customer contact. This proactive intelligence increases the likelihood of resolving issues during the first interaction.
Figure 13 Personalizing Customer InteractionsSelect image to enlarge
As illustrated in the diagram, Amazon Connect places personalization at the center of the customer experience by intelligently connecting the following:
Contact information – Identifying the customer across channels
Contact history – Maintaining continuity between interactions
Case information – Bringing relevant support details forward
Purchase history – Understanding what products and services the customer uses
Customer insights – Using predictive analytics to anticipate needs
This unified approach transforms typical fragmented support experiences into seamless interactions where customers feel recognized and understood from the first moment of contact.
Calculated attributes
Calculated attributes represent a key AI capability that elevates customer service beyond basic identification to truly intelligent engagement. These dynamically generated insights transform raw behavioral data into predictive customer understanding that drives meaningful personalization.
Calculated attributes create actionable intelligence as follows:
Identifying a customer's preferred communication channel – Analyzing patterns to determine whether a customer responds best to chat, email, SMS, or phone contact
Calculating average order frequency – Determining typical purchase cycles to anticipate needs and identify potential upsell opportunities
Recognizing past behavior such as frequency of contact – Establishing baselines for normal customer behavior that help identify potential satisfaction issues
Detecting rising frustration based on recent interactions – Analyzing tone, word choice, and contact patterns to predict customer sentiment
For example, a customer typically contacts support through chat during business hours. If the customer unexpectedly calls after hours, the system might flag this as an urgent issue requiring special attention. This pattern deviation intelligence makes it possible for support teams to prioritize appropriately and respond with the appropriate level of urgency.
Figure 14 Personalizing Customer InteractionsSelect image to enlarge
The system adapts with each interaction, building a deeper understanding of customer needs to guide automated systems and human agents.
Human AI Partnership
The most effective implementations of AI for customer engagement augment human agents instead of replacing them. AI can augment agents as follows:
AI handles data gathering and analysis.
AI suggests possible solutions and provides context.
Human agents apply judgment, empathy, and creativity.
Together, they deliver better outcomes than either could alone.
For example, when a customer contacts support about a product issue, Amazon Connect can immediately recognize them. It can pull up their purchase history and identify the likely product they are calling about. Amazon Connect can recommend solutions based on similar cases before the agent even greets the customer. This leaves the agent free to focus on the human elements of the interaction that AI cannot replicate.
Benefits of AI powered customer engagement
Organizations implementing AI-powered customer engagement solutions see improvements across multiple dimensions of their business. The following tangible benefits demonstrate why forward-thinking companies are adopting these technologies to transform their customer experience operations. Explore each benefit below to understand how AI creates value for both customers and businesses.
Improved first contact resolution
When customers get their issues resolved on the first contact, it delivers the following benefits:
Lower support costs (fewer repeat contacts)
Higher customer satisfaction
Reduced customer effort
"More than 80% of companies cite improved CSAT as the top driver for proactive outreach." AI for Business Success, Jan 2024. Metrigy.
Reduced handle time without sacrificing quality
AI helps agents work more efficiently without cutting corners by doing the following:
Pre-populating customer information from unified profiles
Capturing and categorizing interaction details
Suggesting appropriate next steps
Enhanced personalization at scale
AI makes personalization scalable as follows:
Recognizing customers across channels instantly
Providing relevant context at the right moment
Suggesting personalized next-best actions
Improved operational efficiency
Beyond customer experience benefits, AI delivers significant operational improvements such as the following:
More efficient agent onboarding
Optimized workforce management
Reduced technical debt
Measurable business outcomes
The benefits translate to concrete business results:
Cost Savings
Reduced average handle time
Higher first-contact resolution
More efficient agent training
Revenue Protection and Growth
Improved customer retention
Increased cross-sell/upsell
Higher customer lifetime value
Customer Experience Improvements
Higher customer satisfaction scores
Reduced customer effort
Improved Net Promoter Score (NPS)
As customer expectations continue to rise, the gap between companies that embrace AI-powered engagement strategies and those that continue to use fragmented, reactive approaches widens. Leading organizations will view AI not merely as automation, but as a tool for understanding customers at a deeper level.
RECALL & APPLY
Practice what you learned
Try answering in your own words, then compare with the suggested answer. These activities are for self-study and are not automatically graded.
01What should be improved before adding more automation to disconnected customer systems?
Topic: AI powered customer engagement
Reveal suggested answer
Establish the data foundation and integration needed to understand the customer and the current issue.
02Why is adding chat different from connecting the customer journey?
Topic: Contact center evolution
Reveal suggested answer
Adding a channel provides another interface. Continuity requires sharing relevant identity, history, and case context across those interfaces.
03Two records share a nickname but have different account details. Should they merge automatically?
Topic: Identity and personalization
Reveal suggested answer
A nickname alone is insufficient. Review the configured matching and merging criteria and the consequences of a false match.
04How can a calculated attribute be more useful than a long activity list?
Topic: Customer context and calculated attributes
Reveal suggested answer
It condenses relevant behavior, such as contact frequency, into a value the workflow can use, while the underlying history remains useful for investigation.
05Why should AHT be reviewed together with FCR?
Topic: The human and AI partnership
Reveal suggested answer
A shorter contact is not an improvement if it creates unresolved issues and repeat contacts. Measure both speed and resolution quality.
Capture a key idea, an example, or a question for your instructor.
In this section, you will learn to do the following:
Configure Amazon Connect Customer Profiles with generative AI data mapping to unify customer data from multiple sources.
Implement Identity Resolution using both rule-based and ML approaches to create consolidated customer profiles.
Apply best practices for profile merging to improve agent efficiency and customer experience.
Section introduction
Imagine answering a customer call and immediately having access to your customer's complete history, preferences, and account details. You have their data organized in one place without switching between multiple systems.
With Amazon Connect Customer Profiles, you can combine contact history from Amazon Connect with information from external applications to provide this improved experience.
In this section, you will learn how generative AI streamlines data mapping and how identity resolution automatically finds and merges duplicate customer records.
Components of Amazon Connect Customer Profiles
Amazon Connect Customer Profiles combine customer contact history with important information such as their account number, personal details, contact information, and interaction preferences. After it is enabled, Amazon Connect Customer Profiles automatically creates a unique profile for every customer who contacts your organization.
When a customer contacts your organization, Amazon Connect can identify the customer. Typically, it uses a phone number to identify the customer for a voice call or email address for chat. Amazon Connect Customer Profiles then retrieves the customer's existing profile or creates a new one. The profile may contain the following:
Personal information (name, date of birth, email addresses)
Account details (account numbers, status)
Contact history (previous calls, chats, and interactions)
Case information (open issues, resolved problems)
Data from external systems (customer relationship management records, order history)
Benefits for agents and customers
Benefits for Agents
Having all customer information in one place benefits human support agents in the following ways:
Reduced handle time – Agents no longer need to switch between systems.
Personalized service – Agents can greet customers by name and acknowledge their contact history.
Figure 15 Components of Amazon Connect Customer ProfilesSelect image to enlarge
Benefits for Customers
The benefits for customers are as follows:
Reduced repetition – Customer details are stored, which reduces the need to repeat the same information.
Faster resolution – Customers experience faster issue resolution because agents have customer data when they connect.
Personalized service – Agents can greet customers by name or route them to an agent faster due to account status.
Consistent experience – Customers experience is more consistent across different communication channels.
Figure 16 Components of Amazon Connect Customer ProfilesSelect image to enlargeFigure 17 Components of Amazon Connect Customer ProfilesSelect image to enlarge
1 - Customer profile
This section includes information such as account number, additional information, birth date, email, multiple addresses, name, and party type.
2 - Cases
This section includes status, reference ID, title, source, updated date, and more information related to cases ingested from third-party applications. This is in addition to cases created and managed using Amazon Connect Cases.
3 - More information
This section includes customer-defined attributes and might contain information like cell phone number and shipping address.
This information is sorted alphabetically to help an agent quickly locate the information they need.
4 - Contact history
This section includes dates, times, and duration when this customer contacted your contact center in the past.
5 - Product purchase history
All the assets purchased by a customer can be populated here. The data is ingested from an external application that you have integrated with Amazon Connect Customer Profiles.
RECALL & APPLY
Practice what you learned
Try answering in your own words, then compare with the suggested answer. These activities are for self-study and are not automatically graded.
01Which data would help an agent handle a repeat delivery complaint?
Topic: Amazon Connect Customer Profiles
Reveal suggested answer
The customer’s identity, order and shipping history, previous contacts, open case, and any promised follow up.
02What should the instructor point out in the profile screenshot?
Topic: The customer profile workspace
Reveal suggested answer
Locate the profile, cases, additional attributes, contact history, and purchase history. Explain how each area supports a different part of the conversation.
Capture a key idea, an example, or a question for your instructor.
Modern contact centers handle massive volumes of customer data from multiple sources. Understanding how to effectively integrate this data is crucial for delivering exceptional customer experiences.
Comparing traditional and AI powered data integrations
The following sections compare traditional data integration with generative AI data mapping in Amazon Connect.
Traditional Data Integration Methods
Traditional data integration relies on manual processes and predefined rules.
Organizations typically use extract, transform, load (ETL) pipelines where data engineers manually map fields between systems.
For example, mapping a customer's phone_number field from a customer relationship management (CRM) system to the contact_number field in Amazon Connect requires explicit programming.
Key characteristics of traditional methods include the following:
Manual field mapping – Engineers identify and code relationships between data fields.
Static rule-based transformations – Predefined logic handles data format conversions.
Batch processing – Data moves in scheduled intervals rather than in real time.
High maintenance overhead –Changes require developer intervention and testing.
The next section explains generative AI data mapping.
Generative AI-Powered Data Mapping
Generative AI-powered data mapping in Amazon Connect Customer Profiles is an intelligent system that analyzes your customer data. The data mapping feature uses data from various sources and automatically determines how to organize and combine it into unified profiles. This capability reduces the time needed to create unified customer profiles.
For example, an insurance company can use the generative AI-powered data mapping in Amazon Connect to automatically consolidate customer information. This makes it possible for the company to map customer data from their claims database, CRM system, and payment portal into unified profiles. Agents can view a comprehensive customer history on a single screen and reduce call handling time.
Key characteristics of generative AI-powered features include the following:
Intelligent pattern recognition – AI identifies similar data types across systems automatically.
Semantic understanding – Natural language processing recognizes that customer_phone and contact_number represent the same concept.
Real-time adaptation – Continuous learning improves mapping accuracy over time.
Self-healing integrations – Automatic adjustments occur when source systems change.
The following sections describe the mapping process and configuration.
Although traditional data integration methods provide proven reliability, generative AI-powered automated mapping represents the future of contact center data management. It offers unprecedented speed, accuracy, and scalability for modern customer service operations.
Generative AI Powered Data Mapping Process
The data mapping process consists of the four phases described below.
How does data mapping with generative AI work
Generative AI transforms how you map customer data in Amazon Connect Customer Profiles. The system analyzes your incoming data from various sources and intelligently determines relationship patterns between fields.
Generative AI-powered data mapping in Amazon Connect Customer Profiles automatically determines how to organize and combine it into unified profiles.
This significantly reduces the time needed for unified customer profiles, which empowers you to deliver more personalized customer experiences efficiently.
Let's see how it works in action.
Phase 1 Data Collection and Analysis
Figure 18 Phase 1 Data Collection and AnalysisSelect image to enlarge
The system fetches source attributes and, if available, sample data from your data source. For an Amazon Simple Storage Service (Amazon S3) data source, the first CSV file found in the selected Amazon S3 bucket and prefix will be used as the sample data.
For other data sources such as Salesforce or Zendesk, it fetches attributes through Amazon AppFlow.
Phase 2 Smart Field Mapping
Figure 19 Phase 2 Smart Field MappingSelect image to enlarge
A large language model (LLM) processes each custom attribute from your data source and maps them to standard customer profile attributes.
For example, if your system has fields labeled CustomerEmailAddress, Email_Contact, and EmailID, the AI recognizes these are all referring to email addresses and maps them appropriately.
Phase 3 Key Attribute Selection
Figure 20 Phase 3 Key Attribute SelectionSelect image to enlarge
After mapping fields, the LLM selects suitable attributes that can serve as keys (standard identifiers) for customer profiles. These include the following:
Unique identifier – You must have a unique identifier for your data to avoid ingestion errors. This identifier, also known as the unique key, distinguishes data and enables indexing for search and updates. There can be only one unique identifier.
Customer identifier – You must have at least one customer identifier to avoid ingestion errors. Also known as the profile key, Amazon Connect Customer Profiles uses it to determine if data can be associated to existing profiles or create new ones by searching other profiles for this identifier. You can have multiple customer identifiers.
Product identifier – You must have at least one product identifier to avoid ingestion errors. Also known as the asset key, Amazon Connect Customer Profiles uses it to distinguish from other customer product purchase data and determine profile associations by searching other profiles for this identifier. You can have multiple product identifiers.
Case identifier – You must have at least one case identifier to avoid ingestion errors. Also known as the case key, Amazon Connect Customer Profiles uses it to distinguish from other customer case data and determine profile associations by searching other profiles for this identifier. You can have multiple case identifiers.
Order identifier – You must have at least one order identifier to avoid ingestion errors. Also known as the order key, Customer Profiles uses it to distinguish from other customer order data and determine profile associations by searching other profiles for this identifier. You can have multiple order identifiers.
Additional search attributes (optional) – You can choose attributes in your data source object that you want to index to be searchable. By default, all your identifiers are indexed.
After mapping fields, the LLM selects suitable attributes that can serve as keys (standard identifiers) for customer profiles. These include the following:
Unique identifier – You must have a unique identifier for your data to avoid ingestion errors. This identifier, also known as the unique key, distinguishes data and enables indexing for search and updates. There can be only one unique identifier.
Customer identifier – You must have at least one customer identifier to avoid ingestion errors. Also known as the profile key, Amazon Connect Customer Profiles uses it to determine if data can be associated to existing profiles or create new ones by searching other profiles for this identifier. You can have multiple customer identifiers.
Product identifier – You must have at least one product identifier to avoid ingestion errors. Also known as the asset key, Amazon Connect Customer Profiles uses it to distinguish from other customer product purchase data and determine profile associations by searching other profiles for this identifier. You can have multiple product identifiers.
Case identifier – You must have at least one case identifier to avoid ingestion errors. Also known as the case key, Amazon Connect Customer Profiles uses it to distinguish from other customer case data and determine profile associations by searching other profiles for this identifier. You can have multiple case identifiers.
Order identifier – You must have at least one order identifier to avoid ingestion errors. Also known as the order key, Customer Profiles uses it to distinguish from other customer order data and determine profile associations by searching other profiles for this identifier. You can have multiple order identifiers.
Additional search attributes (optional) – You can choose attributes in your data source object that you want to index to be searchable. By default, all your identifiers are indexed.
Phase 4 Timestamp Processing
Figure 21 Phase 4 Timestamp ProcessingSelect image to enlarge
Finally, the system parses timestamps to maintain the correct chronological order of records. This ensures that customer interactions and updates appear in the right sequence.
The entire process typically takes minutes. This improves speed and accuracy compared to manual data mapping, which could take hours or days.
Setting up Generative AI Powered Data Mapping in Amazon Connect Customer Profiles
Contact center administrators can review and complete the setup of customer profiles for data mapping to occur. This will provide agents with relevant customer information and dynamically personalized interactive voice response (IVR) and self-service assistants to improve customer satisfaction and agent productivity.
Welcome to this demonstration on setting up generative AI-powered data mapping in Amazon Connect Customer Profiles.
To start, sign in to the AWS Management Console.
In the search box at the top of the page, enter Amazon Connect. In the displayed list of services, choose Amazon Connect.
On the Amazon Connect page, the instances in your account are displayed. Choose the Instance alias of the instance that you want to set up automated data mapping in Amazon Connect Customer Profiles on.
On the Instance page, under Applications, choose Customer Profiles.
On the Data source integrations tab, choose Add data source integration to begin the process.
Select a Data source from any of the 70+ available no-code data connectors such as Adobe Analytics, Salesforce, or Amazon Simple Storage Service (S3). Each connector is designed to pull specific types of customer data. Configure the connector you have selected, then choose Next.
When you reach the Map data step, select the Auto-generate mapping option. This is where the generative AI capabilities are used to automatically map the attributes of your data to Amazon Connect Customer Profiles attributes. Then, choose Next to continue.
You will now be presented with a summary of all the automatically mapped Amazon Connect Customer Profiles attributes. This is your opportunity to review and make any necessary adjustments. When you are ready, choose Next to continue.
When you are satisfied with the configuration, choose Add data source integration to complete the connector configuration and begin ingesting your customer data into Amazon Connect Customer Profiles.
Your new connector will now be listed in your Amazon Connect Customer Profiles Data source integrations. When the connector is complete, it will show as Active.
RECALL & APPLY
Practice what you learned
Try answering in your own words, then compare with the suggested answer. These activities are for self-study and are not automatically graded.
01Why must a proposed field mapping be reviewed?
Topic: Data integration and generative mapping
Reveal suggested answer
Similar names can represent different business meanings. Incorrect mappings can corrupt identity, dates, or customer context.
02Why do identifiers and timestamps deserve special attention?
Topic: The data mapping process
Reveal suggested answer
Identifiers determine association and updates. Timestamps affect record ordering and which data appears current.
03What is the critical review point before completing the integration?
Topic: Data mapping demonstration
Reveal suggested answer
Check each mapping, required identifier, and timestamp interpretation. Verify that sample records create or update the intended profiles.
Capture a key idea, an example, or a question for your instructor.
There can be multiple profiles when customer records are captured across multiple channels and applications for the same customer. Without a common unique identifier linking these profiles, the same customer ends up with separate, disconnected profiles across different systems.
Identity Resolution in Amazon Connect Customer Profiles tackles this challenge by finding each profile and consolidating them. Identity Resolution uses both rules-based matching and ML matching to ensure that each customer has just one comprehensive profile.
Rule-based matching relies on predefined rules to determine if two profiles represent the same customer. This method compares specific attributes between profiles using a set of matching rules.
These rules examine attributes such as the following:
Account number
Address components (city, postal code, country)
Email addresses
Phone numbers
Name components (first name, last name, middle name)
Date of birth
You can configure up to 15 different matching rules to catch various scenarios where profiles might represent the same customer. The strength of rule-based matching is its precision and transparency. You know exactly why two profiles were matched.
You can choose how profiles are compared across attribute types and which attribute to use for matching from each type.
For example, if you want to match multiple email types, choose many-to-many to match across profiles and attribute types that your business uses.
You can choose from multiple attribute types.
Email type
Choose from the following:
EmailAddress
BusinessEmailAddress
PersonalEmailAddress
Phone number type
Choose from the following:
PhoneNumberNumber
HomePhoneNumber
MobilePhoneNumber
Address type
Choose from the following:
Address
BusinessAddress
MaillingAddress
ShippingAddress
matching (matches across sub-types) as follows:
ONE_TO_ONE – The system can only match if the sub-types are exact matches.
For example, when the EmailAddress fields of Profile A and B match, the two profiles are matched on the EmailAddress type.
MANY_TO_MANY – The system can match attributes across the sub-types of an attribute type.
For example, if EmailAddress for Profile A matches BusinessEmailAddress for Profile B, the profiles are matched based on EmailAddress type.
ML matching
For more sophisticated matching capabilities, Amazon Connect offers ML-based Identity Resolution. This approach uses AI to identify similarities that might not be caught by basic rules.
ML-based Identity Resolution reviews the following personal identifiable information (PII) attributes in each profile:
Names (first, middle, last)
Email addresses (personal, business)
Phone numbers (home, mobile, business)
Addresses (business, mailing, shipping, billing)
Date of birth
Unlike rule-based matching, ML-based matching can detect similarities even when information is not exactly the same.
For example, it might recognize that Ana Carolina Silva and Ana Silva with similar addresses are likely the same person.
Setting up Identity Resolution in Amazon Connect Customer Profiles
Transcript Setting up Identity Resolution in Amazon Connect Customer Profiles
Welcome to this demonstration on setting up Identity Resolution for your Amazon Connect Customer Profiles.
To start, sign in to the AWS Management Console.
In the search box at the top of the page, enter Amazon Connect. In the displayed list of services, choose Amazon Connect.
On the Amazon Connect page, the instances in your account are displayed. Choose the Instance alias that you want to set up Identity Resolution for Amazon Connect Customer Profiles.
On the Instance page, under Applications, choose Customer Profiles.
On the Amazon Connect Customer Profiles page, in the Identity Resolution section, choose Enable identity resolution to begin the process.
A message appears asking you to confirm that you understand that enabling Identity Resolution will not result in any profile merging, and that you will need to enable merging separately within Identity Resolution. After reading the message, choose Enable Identity Resolution to proceed.
You've now successfully enabled Identity Resolution with rule-based matching and machine learning-based matching for your Amazon Connect Customer Profiles domain.
After enabling Identity Resolution for a new domain with rule-based matching, the matching will start immediately if you have an integration running. For existing domains, the matching process will start within one hour. For machine learning-based matching, the Identity Resolution job will run for the first time within 24 hours of enabling the feature.
RECALL & APPLY
Practice what you learned
Try answering in your own words, then compare with the suggested answer. These activities are for self-study and are not automatically graded.
01Does enabling Identity Resolution immediately merge all matches?
Topic: Identity Resolution matching
Reveal suggested answer
No. The course distinguishes identifying matches from enabling merging. Review that separation during the demonstration.
02Why might a personal email need to match a business email field?
Topic: Attribute matching and setup
Reveal suggested answer
The same address may be stored under different subtypes in different source systems. Matching configuration controls whether those comparisons are allowed.
Capture a key idea, an example, or a question for your instructor.
After Identity Resolution identifies similar profiles (either through rules or ML), the next step is to consolidate them through the following auto-merging process:
Apply consolidation criteria that you define to determine which profiles should be combined.
Create a consolidated profile that combines information from all matching profiles.
Preserve information from the original profiles, and retain all values if there are conflicts.
Update references to the original profiles to point to the new consolidated profile.
Auto merging process
The Identity Resolution process in Amazon Connect Customer Profiles, automatically combines fragmented customer data from multiple sources into a unified customer profile.
In the following example, John Doe creates an account with his personal email. Later, he contacts support using his work email. Without Identity Resolution, the system would create two separate profiles. With Identity Resolution, if other attributes, such as the name and phone number match, the system recognizes these are the same person and merges the profiles.
The numbered markers in the following figure show the auto merging process.
Figure 22 Auto merging processSelect image to enlarge
1 - Fragmented first-party (1P) data collection
Customer data from various touchpoints:
Advertising: Mobile ID
Email: Name and email address
Site: Name, ID and email
Retail: Full name, ID, email, and address
App: Mobile ID and phone number
Contact center: Name and phone number
2 - Automatic matching and merging
The system processes fragments through a three-stage workflow:
1P deterministic match: Identifies exact matches like identical email addresses
1P probabilistic match: Identifies exact matches like identical email addresses
Auto-merging records: Combines matched information into a single profile using custom workflows
3 - Unified profile creation
The result is a comprehensive customer profile containing the following information:
Full name: John Doe
ID: CRM123
Email address: J.Doe@email.com
Phone number: 212-867-5309
Address: 123 4th St NY, NY
Mobile ID: M1234
This automated process helps businesses recognize the same customers across different channels, which can result in better customer service experiences and more effective marketing.
Setting up Auto Merging in Identity Resolution for Amazon Connect Customer Profiles
When similar profiles are detected by an Identity Resolution job, the process can automatically merge them into a unified profile based on auto-merging rules that you specify.
Transcript Setting up Auto Merging in Identity Resolution for Amazon Connect Customer Profiles
Welcome to this demonstration on setting up auto-merging in Identity Resolution for Amazon Connect Customer Profiles.
To start, sign in to the AWS Management Console.
In the search box at the top of the page, enter Amazon Connect. In the displayed list of services, choose Amazon Connect.
On the Amazon Connect page, the instances in your account are displayed. Choose the Instance alias of the instance that you want to set up Identity Resolution in Amazon Connect Customer Profiles on.
Now, on your Amazon Connect instance Overview page in the navigation pane, under Applications, choose Customer Profiles.
Next, in the Identity Resolution section, choose View Identity Resolution.
On the Identity Resolution page, in the Identity Resolution settings section, you will see the Rule-based resolution and Machine learning resolution. Choose the resolution that you want to setup auto-merging on. For this demonstration, choose the Rule-based resolution.
Now, on the Rule-based resolution settings page, in the Merge rule-based matches section, choose Edit.
The Missing timestamp pop-up appears and indicates whether you have custom object type mappings. Amazon Connect uses a timestamp attribute and timestamp format to determine when a profile was last updated. If the Missing timestamp pop-up appears, it means there is a timestamp missing from your custom objects. You can add it using the PutProfileObjectType API. If your object type does not have a proper timestamp attribute, you can acknowledge that a default timestamp will be applied for records ingested into Amazon Connect Customer Profiles. For this demonstration, select the acknowledgment. Then, choose Next.
On the Edit merge rule-based matches page, in the Merge matches section, you will define when to merge profiles based on the selected rule. Select the Merge matches found by rule-based matching checkbox.
Next, choose a rule level for merging from the list available. For this demonstration, choose Rule 1.
Then, choose Save.
In the Merge data pop-up, enter confirm in the textbox.
Then, choose Merge data.
Now, on the Rule-based resolution settings page, notice that the rule-based resolution settings have been updated.
Next, decide if you want to review matched profile IDs by having them written to an Amazon Simple Storage Service (Amazon S3) bucket. For this demonstration, assign an S3 bucket to the rule-based resolution. In the Match results location section, choose Edit.
Now, on the Edit match results location page, in the S3 location - optionalsection, select the Write profile ID matches to Amazon S3 checkbox.
To choose your bucket destination, either enter the S3 URI or choose Browse S3. For this demonstration, choose Browse S3.
In the Choose an archive in S3 pop-up, under Buckets, enter the name of the S3 bucket.
Now, choose the S3 bucket.
Then, select Choose.
Now, with the S3 bucket URI entered into the S3 URI textbox, choose Save.
On the Rules-based resolution settings page, an alert indicates you have successfully assigned the S3 bucket location.
Next, choose View Identity Resolution.
On the Identity Resolution page, notice that the rule-based resolution has an Active status for both Find matches and Merge matches. Also, S3 location displays the chosen S3 bucket.
Now, you will complete the machine learning resolution. In the Identity Resolution settings section, choose Machine learning resolution.
On the Machine learning resolution settings page, in the Merge matches section, choose Edit.
Now, on the Edit merge machine learning matches page, in the Merge matches section, define when to merge profiles based on attribute rules you create. Select the Merge matches found by machine learning matching checkbox.
Next, under Rule 1, choose the Attributes menu, and then choose one or more attributes from the list. For this demonstration, choose AccountNumber and EmailAddress.
To add more rules, choose Add merge rule. For the second rule, choose AccountNumber only, and for the third rule, choose EmailAddress only.
Then, choose Save.
In the Merge data pop-up, enter confirm in the textbox.
Then, choose Merge data.
Now, on the Machine learning resolution settings page, notice that the machine learning resolution settings have been updated.
Next, decide if you want to review matched profile IDs by having them written to an S3 bucket. For this demonstration, you will assign an S3 bucket to the machine learning resolution. In the Match results location section, choose Edit.
Now, on the Edit match results location page, in the S3 location - optional section, select the Write profile ID matches to Amazon S3 checkbox.
To choose your bucket destination, either enter the S3 URI or choose Browse S3. For this demonstration, choose Browse S3.
In the Choose an archive in S3 pop-up, under Buckets, enter the name of the S3 bucket.
Next, choose the S3 bucket.
Then, select Choose.
Now, with the S3 bucket URI entered into the S3 URI textbox, choose Save.
On the Machine learning resolutionsettings page, an alert indicates you have successfully assigned the S3 bucket location.
Next, choose View Identity Resolution.
On the Identity Resolution page, notice that the machine learning resolution has an Active status for both Find matches and Merge matches. Also, S3 location displays the chosen S3 bucket.
That's it. You successfully set up auto-merging in Identity Resolution and assigned S3 buckets to store matched profile IDs for Amazon Connect Customer Profiles.
RECALL & APPLY
Practice what you learned
Try answering in your own words, then compare with the suggested answer. These activities are for self-study and are not automatically graded.
01What risk does an overly broad merge rule create?
Topic: Auto merging customer profiles
Reveal suggested answer
It can consolidate different people into one profile, producing incorrect customer context and potential data exposure.
02What should you review before enabling a merge rule?
Topic: Auto merging demonstration
Reveal suggested answer
Review the attributes, rule level, timestamps, expected match examples, and likely false matches. Use an authorized training environment for the demonstration.
Capture a key idea, an example, or a question for your instructor.
In this section, you will learn to do the following:
Configure Amazon Connect outbound campaigns with AI-powered call classification to optimize agent productivity.
Create targeted customer segments using both natural language prompts and data-driven inspiration cards.
Implement segmentation strategies that improve campaign performance and customer engagement metrics.
Section introduction
Imagine you are tasked with reaching out to thousands of customers who haven't made a purchase in the past month. How would you identify these customers? How would you make sure that your outreach is personalized and effective? This is where outbound campaign intelligence comes into play.
Setting up Outbound Campaigns
Outbound campaigns are proactive communications initiated by businesses to their customers, and they are essential for everything from appointment reminders to marketing promotions.
Amazon Connect outbound campaigns can help with various proactive communications. Review the following examples of outbound campaigns:
Appointment reminders
Marketing promotions
Delivery notifications
Billing reminders
Service follow-ups
Customer satisfaction surveys
Preparing to set up Amazon Connect outbound campaigns
Before beginning your first campaign, you need to complete the following prerequisites:
In the following demonstration, you will learn about enabling outbound campaigns in the Amazon Connect. Amazon Connect outbound campaigns allow you to automate outbound campaign voice communications to your customers.
Transcript Configuring Outbound Campaigns in Amazon Connect
Welcome to this demonstration on how to set up Amazon Connect outbound campaigns, formerly known as high-volume outbound communications.
To start, sign in to the AWS Management Console.
In the search box at the top of the page, enter Amazon Connect. In the displayed list of services, choose Amazon Connect.
On the Amazon Connect page, the instances in your account are displayed. Choose the Instance alias of the instance that you want to enable outbound campaigns on.
Next, in the navigation pane, under Channels and communications, choose Outbound campaigns.
On the Outbound campaigns page, choose Enable. If you don't have this option, verify whether outbound campaigns is available in your AWS Region.
Next, under the Encryption section, you will need to either enter your own AWS KMS key or choose to create a new one. For this demonstration, you will create a new KMS key.
Choose Create an AWS KMS key. A new browser tab will open for the AWS Key Management Service (AWS KMS) console.
On the Create key page, in the Configure key step, review that the Key type is set to Symmetric and Key usage is set as Encrypt and decrypt. Then, choose Next.
On the Add labels step, enter a descriptive alias and description for your key. In this demonstration, for Alias, enter KMS-AmazonConnectOutboundCampaigns. For Description, enter KMS for Amazon Connect Outbound Campaigns. When you are done, choose Next.
On the Define key administrative permissions - optionalstep, choose Next.
On the Define key usage permissions - optionalstep, choose Next.
On the Edit key policy - optionalstep, choose Next.
Finally, on the Review step, choose Finish to complete the key creation process.
Now, return to the Amazon Connect console tab. Choose the AWS KMS key field, and review your list of available keys. Your newly created key should appear in the list. In this demonstration, you will choose the KMS-AmazonConnectOutboundCampaigns key.
With everything configured, choose Enable outbound campaigns. The process will take a few minutes to complete. After outbound campaigns is enabled, you can begin creating outbound campaigns for voice calls in Amazon Connect.
If the enabling process fails, you might need to verify that you have the required AWS Identity and Access Management (IAM) permissions for the key, including kms:DescribeKey, kms:CreateGrant, and kms:RetireGrant.
To enable your total instance limits for outbound campaigns, you need to make sure that your campaign permissions are up to date. Choose Upgrade permission.
This completes the demonstration on setting up Amazon Connect outbound campaigns.
RECALL & APPLY
Practice what you learned
Try answering in your own words, then compare with the suggested answer. These activities are for self-study and are not automatically graded.
01Why is the flow’s call progress branch important?
Topic: Outbound campaign setup
Reveal suggested answer
It determines how the campaign handles a person, an answering machine, or an uncertain outcome.
02Where would you investigate an enablement failure involving the key?
Topic: Outbound campaign demonstration
Reveal suggested answer
Review the key configuration and the required permissions identified in the course procedure. Confirm the intended instance and region as well.
Capture a key idea, an example, or a question for your instructor.
Customer segments are specific groups of people who share common characteristics or behaviors. Customer segments are dynamically evaluated based on attributes that you define and can change over time when the value of the attributes change. In Amazon Connect, segments serve as the foundation for targeted outbound campaigns.
Instead of sending the same promotional message to all customers, you might create different segments, such as the following:
Customers who abandoned their shopping carts recently
High-value customers who haven't purchased in 30 days
New customers who made their first purchase in the past week
Real world example
The following diagram illustrates how a customer can be segmented based on interaction type (inbound compared to outbound contacts) and priority levels. At the center is Amazon Connect, which serves as the hub for all customer interactions.
On the left side of the diagram, inbound contacts are categorized into the following three priority segments:
Cart abandoners (urgent priority) with an estimated 3,200 contacts who abandoned their carts within 24–72 hours
High-value customers (high priority) with 2,450 estimated contacts who haven't made purchases in more than 30 days and have historical spend exceeding $500 each month
New customers (medium priority) with approximately 1,850 contacts who made their first purchase within 7 days
The right side of the diagram displays the following outbound contact strategies:
Recovery and conversion efforts (urgent priority) targeting 200 contacts with limited-time offers and purchase hesitation solutions
Reengagement strategies (high priority) focusing on 350 contacts through personalized discounts and loyalty program enrollment
Welcome and onboarding (medium priority) reaching 1,500 contacts with product tutorials and support information
Figure 23 Understanding Customer SegmentsSelect image to enlarge
Segmentation benefits
The more targeted your segments, the more personalized and effective your campaigns can be. The following are some examples:
Higher engagement rates: When customers receive relevant messages, they're more likely to engage.
Improved customer experience: Personalized communications demonstrate to customers that you understand their needs.
Better resource allocation: By focusing on the right customers at the right time, you optimize your outreach efforts.
AI in Campaign Management
AI has revolutionized outbound campaign management, enabling businesses to work smarter, not harder. From detecting whether a human or machine answered a call to receiving recommendations based on trends in the customer data, AI-powered tools make campaigns more efficient.
In Amazon Connect, AI powers the following key aspects of outbound campaigns.
Call Classification
Call classification uses machine learning to determine whether a call was answered by a person or an automated system.
Amazon Connect call classification technology analyzes several factors:
Background noise analysis: It can detect background noise patterns typically associated with pre-recorded messages.
Speech patterns: The system recognizes long strings of words common in voicemail greetings.
Human response patterns: Call classification identifies typical human responses, such as "Hello," followed by pauses.
Amazon Connect call classification connects agents only with live customers.
Segment AI Assistant
Generative AI-powered segmentation helps non-technical business users to build audiences using natural language queries.
Natural language segment creation represents a significant breakthrough in making data more accessible to non-technical users. Instead of navigating complex filter interfaces or writing database queries, you can describe the customers you want to target in everyday language.
Amazon Connect segment AI assistant interprets your natural language description and translates it into a structured segment definition that identifies the exact customers matching your criteria.
Inspiration Cards
This generative AI-powered feature presents segment ideas tailored to specific customer data and trends, streamlining segment creation.
Inspiration cards are AI-powered recommendations that analyze your customer data to suggest potentially valuable segments for your campaigns. Unlike traditional segmentation that requires you to define criteria from scratch, inspiration cards proactively identify patterns and opportunities in your data.
Understanding call classification
Call classification uses machine learning to determine whether a call was answered by a person or an automated system.
Call progress analysis
When your outbound campaign makes a call, the Check call progress flow block branches based on the call classification outcome. The following are a few examples:
If a human answers, it branches to connect to an agent.
If an answering machine responds, it branches to leave a pre-recorded message.
If the ML model cannot determine the answer type, it branches to play a message before connecting to an agent.
This intelligence matters because studies show that many calls to consumers go to voicemail. Without call classification, agents would waste countless hours listening to voicemail greetings and leaving messages.
Figure 24 AI in Campaign ManagementSelect image to enlarge
Agent productivity benefits
Beyond call classification, AI improves agent productivity in several key ways:
Minimizing idle time: Predictive dialing helps agents spend more time speaking with customers.
Prioritizing live connections: Because answering machines, wrong numbers, and no-answers are filtered out, agents stay focused on live connections.
Providing context: When a call connects to an agent, AI can instantly display relevant customer information.
Figure 25 AI in Campaign ManagementSelect image to enlarge
Customer experience impact
AI does not just benefit agents and businesses, it also improves the customer experience in the following ways:
Reduced spam perception: When calls connect to live agents immediately, customers are more likely to engage.
Personalization: AI helps agents have more relevant conversations by providing customer context.
Figure 26 AI in Campaign ManagementSelect image to enlarge
RECALL & APPLY
Practice what you learned
Try answering in your own words, then compare with the suggested answer. These activities are for self-study and are not automatically graded.
01How would “high value customers” become a testable segment?
Topic: Customer segments
Reveal suggested answer
Define the spend metric, threshold, and time period, then inspect the resulting members and exclusions.
02What should happen when the classifier cannot determine the answer type?
Topic: AI in campaign management
Reveal suggested answer
Use the flow’s defined uncertain outcome handling. Demonstrate that the design must account for this result explicitly.
03Why is a high dialing rate insufficient as a success measure?
Topic: Campaign productivity and customer experience
Reveal suggested answer
It does not show whether customers received useful, timely service or whether the campaign achieved its intended outcome.
Capture a key idea, an example, or a question for your instructor.
Natural language segment creation represents a significant breakthrough in making data more accessible to non-technical users. Instead of navigating complex filter interfaces or writing database queries, you can describe the customers you want to target in conversational language.
Amazon Connect segment AI assistant interprets your natural language description and translates it into a structured segment definition that identifies the exact customers matching your criteria.
How natural language segment creation works
Behind the scenes, the segment AI assistant uses advanced natural language processing algorithms to do the following:
Analyze your description to identify key criteria and conditions.
Map these criteria to available customer attributes in your domain.
Construct a logical segment definition with the appropriate filters and relationships.
Apply this definition to your customer data to generate the segment.
For example, instead of manually configuring multiple filters and conditions, you could enter: Customers who have spent over $2,000 in the last 60 days.
Figure 27 Segment AI AssistantSelect image to enlarge
Best practices for your segment AI assistant prompts
Amazon Connect segment AI assistant relies heavily on the quality of prompts to generate effective responses for customer service agents. The difference between vague and better formed prompts can significantly impact the assistant's performance and usefulness. The quality of your segment depends significantly on how you phrase your prompt. Use the following best practices when composing prompts for the segment AI assistant:
Be specific: Include precise criteria instead of vague descriptions.
Reference existing attributes: When possible, use the names of attributes that exist in your data.
Include clear timeframes: Specify time periods clearly, such as in the last quarter.
Start simple: Begin with straightforward prompts and gradually add complexity.
Use business terminology: Frame your prompt in terms of business objectives.
Examples of vague and better formed prompts
Figure 28 Segment AI AssistantSelect image to enlargeFigure 29 Segment AI AssistantSelect image to enlarge
Refining AI generated segments
After the AI generates your segment, you can do the following:
Review the segment definition: Check that all conditions reflect your intent.
Adjust thresholds: Fine-tune values like purchase amounts or timeframes.
Add or remove conditions: Enhance the segment with additional criteria or simplify.
Preview results: See how many customers match your segment and review a sample.
Figure 30 Segment AI AssistantSelect image to enlarge
Inspiration Cards
Inspiration cards are Amazon Connect AI-powered recommendations that analyze your customer data to suggest potentially valuable segments for your campaigns. Unlike traditional segmentation that requires you to define criteria from scratch, inspiration cards proactively identify patterns and opportunities in your data.
These cards appear on the Customer segmentspage in Amazon Connect and present ready-to-use segment ideas based on actual trends in your customer profiles data.
Figure 31 Segment AI AssistantSelect image to enlarge
Inspiration cards generated recommendations
Inspiration cards use advanced analytics and AI to generate their suggestions. They do the following:
Analyze historical data: Examine customer behavior patterns over time.
Identify trends: Spot significant changes or patterns in customer activities.
Apply business logic: Categorize these patterns into meaningful business contexts.
Generate actionable segments: Create predefined segment criteria based on these insights.
The system organizes these suggestions into the following business-focused themes:
Promotion: Segments ideal for marketing promotions and special offers
Retention: Segments highlighting customers who might need attention to prevent churn
Support: Segments identifying customers who might benefit from proactive service
RECALL & APPLY
Practice what you learned
Try answering in your own words, then compare with the suggested answer. These activities are for self-study and are not automatically graded.
01What must you inspect after entering a segment prompt?
Topic: Natural language segment creation
Reveal suggested answer
Inspect the attributes, operators, thresholds, timeframes, AND or OR relationships, and a sample of matched customers.
02Improve the prompt “find customers who spend a lot.”
Topic: Segment prompts and refinement
Reveal suggested answer
Use a defined metric and time period, for example customers whose recorded spend exceeds 2000 in the last 60 days, using attributes available in the domain.
03Should a suggested retention audience immediately enter a campaign?
Topic: Inspiration cards
Reveal suggested answer
First review its definition, sample membership, suitability, and campaign requirements. A suggestion is a starting point for design.
Capture a key idea, an example, or a question for your instructor.
In this section, you will learn to do the following:
Identify common challenges that AI technology addresses in contact center agent workflows.
Compare key performance metrics before and after AI agent assistance implementation.
Apply Amazon Q in Connect features to resolve specific agent productivity challenges.
Evaluate the quality of AI-generated summaries against documentation best practices.
Section introduction
Have you ever called customer service and been put on hold while the agent searches for information? Or maybe you have been an agent yourself, struggling to find answers while a customer waits impatiently on the line. AI in contact centers aims to address these frustrating experiences.Amazon Connect delivers first-party AI across all channels, with unlimited AI that remains tied only to your underlying channel usage rather than AI consumption. Organizations of any size can use AI across all touchpoints, which empowers agents to reach their full potential with an intelligent digital partner. Think of it as giving each agent a super-smart assistant that works alongside them to handle routine tasks. Agents can focus on what humans do best: solving complex problems and building relationships with customers.
Current Agent Challenges and AI Solutions
Contact center agents face numerous challenges that affect their productivity, job satisfaction, and ability to deliver excellent customer service. Agents can use Amazon Connect AI solutions to help in the following ways:
Find solutions quickly
Resolve issues consistently
Manage customer emotions
Document after contact
Finding solutions quickly
Agents often need to search through multiple knowledge bases while customers wait and assemble a solution based on the information they find. They can then present this solution to the customer. Agents need to navigate disparate systems to complete the required actions.
Figure 32 Current Agent Challenges and AI SolutionsSelect image to enlarge
Amazon Connect AI solution:
Amazon Q in Connect, a personalized real-time agent assistant, analyzes customer conversations in real time. This provides agents with relevant information and actions to address issues. Amazon Q in Connect uses your knowledge bases to provide responses. It personalizes these responses based on customer-specific information, including order history, case details, loyalty status, and more.
For example, if a customer asks about a return policy for damaged items, the agent receives the appropriate policy details without needing to search. The agent is also provided with the steps to take to resolve the issue.
Consistent issue resolution
Some customer issues require consistent handling, regardless of the experience level of the agent. These issues have a well-defined process and steps that the agent needs to complete based on the conversation.
Amazon Connect AI solution:
Amazon Q in Connect guided agent workflows recommend step-by-step guidance that seamlessly walks agents through processes that require consistency.
Figure 33 Current Agent Challenges and AI SolutionsSelect image to enlarge
For example, when getting assistance to update information on a customer's account, a step-by-step guide can be launched to complete the process directly from the agent workspace. This intuitive support ensures consistent, high-quality service delivery for every customer scenario regardless of agent tenure.
Customer emotion management
It is difficult for agents to consistently identify customer emotions and conversation trends, especially in text-based channels such as chat.
Amazon Connect AI solution:
Contact Lens provides real-time analytics and sentiment analysis. AI can detect customer emotions and alert supervisors when interactions might need special attention.
Contact Lens detects negative sentiment in real time during customer interactions. For example, it recognizes frustration in customer statements such as, "This is the third time I've contacted you, and nobody seems to care." Supervisors receive alerts about deteriorating conversations, so they can choose to join or guide agents. This proactive approach helps prevent customer loss by addressing emotional signals in conversations.
Figure 34 Current Agent Challenges and AI SolutionsSelect image to enlarge
After contact documentation
Agents typically spend 30-45 seconds documenting each call, which adds up to hours of non-customer-facing time each week.
Amazon Connect AI solution:
Contact Lens automatically generates detailed summaries of customer conversations and captures key issues, actions taken, and next steps. This reduces the time agents spend on after contact documentation. These summaries can be copied for use in case notes or automatically added to systems of record without requiring agents to take action.
For example, Contact Lens created an automatic summary when a customer called about returning a watch. The summary captured key details like the return reason and exchange steps. The agent could immediately assist the next customer instead of spending time on documentation.
Figure 35 Current Agent Challenges and AI SolutionsSelect image to enlarge
Key Metrics Improved Through AI Agent Assistance
When contact centers implement AI agent assistance tools, they typically see improvements across multiple performance metrics. Understanding these metrics helps you set realistic goals and measure success after implementation.
AI typically reduces AHT through faster information access and reduced after contact work.
AI typically improves FCR by giving agents better information and process guidance.
AI-assisted agents typically achieve higher CSAT scores.
AI-enhanced customer service can lead to improved NPS scores.
Contact centers using AI assistance often see improved speed to competency and reductions in onboarding time, training time, and agent turnover.
Key metric examples
Learn about each metric and how AI enhances the experiences of customers in the following real-world examples.
AHT
Imagine an agent helping a customer with a complex return. Instead of putting customers on hold to search multiple systems, the AI assistant works instantly. It provides return policies, purchase history, and resolution steps all in one place. This reduces the total time it takes to complete the request.
FCR
A customer calls about internet connectivity issues. The AI system guides the agent through diagnostic steps and suggests solutions based on the customer's equipment. It can also help schedule technician visits when needed. The agent resolves the customer's issue on the first attempt. This is achieved without transferring the call to a more tenured resource or requiring a follow-up.
CSAT
A customer inquires about retirement planning options. The AI system provides the agent with personalized recommendations based on the customer's age, income, and risk tolerance. When the customer gets frustrated about the forecast gap, the agent can be empathetic and recommend the best options for the customer's unique circumstances. The customer is impressed by the tailored advice and quick service, which leads to a high satisfaction rating.
NPS
A frequent flyer calls an airline about a missed connection. AI instantly recognizes the customer's elite status, provides the agent with rebooking options, and suggests offering a lounge pass as compensation. The customer is delighted by the swift, personalized service and becomes more likely to recommend the airline.
Agent performance and retention
Consider a new agent who assists with a complex billing inquiry. Instead of feeling overwhelmed, they confidently navigate the interaction with AI-powered guidance to resolve the issue successfully. This positive experience contributes to their job satisfaction and likelihood of remaining with the company.
Introduction to Agent Assistance with Amazon Q in Connect
In this section, you will learn to do the following:
Identify specific ways Amazon Q in Connect reduces agent handling time and improves first contact resolution (FCR).
Identify how to configure Amazon Q in Connect knowledge bases to address common customer inquiries.
Recognize how to resolve common knowledge retrieval issues in Amazon Q.
Evaluate the effectiveness of Amazon Q responses for different types of customer scenarios.
Agent Assistance with Amazon Q in Connect
Amazon Q in Connect is a generative AI customer service assistant that helps contact center agents resolve customer issues quickly and accurately. Amazon Q in Connect listens to customer conversations in real time and identifies helpful information. It then presents this information to agents automatically, which eliminates the need for manual searching. Unlike traditional knowledge management systems that require manual keyword searches, Amazon Q in Connect works proactively without agent initiation. It uses conversation context to find the right information and generates natural language responses.
Amazon Q in Connect uses the three stages shown in the following figure.
Figure 36 Agent Assistance with Amazon Q in ConnectSelect image to enlarge
Conversation understanding
The system uses natural language understanding (NLU) to analyze what the customer is saying during calls or chats.
Information retrieval
Based on that understanding, Amazon Q in Connect searches through your connected knowledge sources to find relevant information.
Response generation
Using information from knowledge bases and customer data, the system provides agents with suggested responses and recommended actions. It also offers links to detailed documentation and step-by-step guides for efficient task completion.
Enable Amazon Q in Connect
A Domain, also known as an Assistant, is created when enabling Amazon Q in Connect. You can then associate knowledge bases to your domain, which will be used as the source of information during conversations.
Amazon Q in Connect comes preconfigured with AI agents for different tasks and are created within a domain. These agents can assist human agents in real time or support manual queries that agents might make. Amazon Q in Connect also provides AI agents for other use cases, such as self-service. You can use these default AI agents or create your own for more control over the behavior, associated knowledge bases, guardrails, chunking strategies, and more.
Key points to consider include the following:
Multiple domains can exist, but they operate independently.
One domain can link to multiple Amazon Connect instances.
Each Amazon Connect instance can only link to one domain at a time.
A default knowledge base can be used across the domain, or you can use separate knowledge bases by configuring AI agents.
Amazon Connect instances can be reassigned to different domains as needed.
Integration with Guided Workflows to Assist Agents
Guided workflows are step-by-step instructions that walk agents through specific processes. When integrated with Amazon Q in Connect, these workflows become even more powerful, automatically appearing based on the conversation.
Amazon Q in Connect Guided Workflows
When Amazon Q in Connect detects a specific customer intent, it can trigger the appropriate guided workflow to help the agent resolve that issue. Consider the following:
The customer explains their issue.
Amazon Q in Connect detects the intent.
The system recommends a relevant guided workflow.
The agent follows the workflow steps to resolve the issue.
For example, if a customer says, "I need to dispute a charge on my credit card," Amazon Q in Connect can do the following:
Recognize the intent as a dispute transaction.
Recommend a guided workflow that walks the agent through the dispute process.
Figure 37 Integration with Guided Workflows to Assist AgentsSelect image to enlarge
Flow Configuration
To enable Amazon Q in Connect in your contact flows, add an Amazon Q in Connect block to your contact flow. This block associates the Amazon Q domain with the current contact.
For voice calls only, do the following:
Add a Set recording and analytics behavior block.
Configure it for Contact Lens conversational analytics in real time.
Place this block anywhere in the flow.
Note: Contact Lens conversational analytics is required for Amazon Q to work with voice calls but is not required for chat interactions or self-service use cases.
Figure 38 Integration with Guided Workflows to Assist AgentsSelect image to enlarge
Providing Agent Access
Agents need access to Amazon Q in Connect to take advantage of its benefits during calls. The Admin security profile already includes all Amazon Q permissions by default. For more information about security profiles, see Security profile permissions for Amazon Q in Connect.
For agents to access and use Amazon Q in Connect, assign the following permissions in their security profile:
Amazon Q - Access – Agents can search for and view content. They can also receive automatic recommendations during calls if Contact Lens conversational analytics is enabled.
Custom views - Access – This is required if using the step-by-step guides integration.
Agent access to Amazon Q in Connect
Amazon Q in Connect offers two access methods for agents: directly through the Agent Workspace or through embedded integration in custom applications.
Amazon Connect Agent Workspace
If you are using the agent application provided with Amazon Connect, after you enable Amazon Q in Connect, share the following URL with your agents so they can access it:
Figure 40 Amazon Connect Agent WorkspaceSelect image to enlarge
Introduction to Enhancing Customer Service with AI Powered Tools
In this section, you will learn to do the following:
Identify key indicators in real-time transcription that require immediate agent action.
Differentiate between types of customer sentiment signals and select appropriate response strategies.
Apply Contact Lens AI summarization features to standardize post-call documentation.
Evaluate when to use different AI-powered tools based on specific customer interaction scenarios.
Real time Transcription Seeing What You Hear
Real-time transcription converts speech to text as the conversation happens. When a customer calls in, Amazon Connect Contact Lens analyzes the audio stream immediately and converts spoken words into text almost instantly. Real-time transcription with Amazon Connect Contact Lens offers significant advantages for customer service operations.
Let us explore how this capability enhances contact center effectiveness:
Capturing complex information
When a customer provides detailed information like reference numbers or addresses, you do not need to ask them to repeat information.
Agents can review the complete transcript during After Contact Work to verify details or use in back office operations. Reducing errors in documentation and follow-up actions. Supervisors can monitor calls more effectively without listening to entire conversations.
Providing accurate records of complex customer details, eliminating reliance on agent memory or incomplete notes.
Avoiding misunderstandings
If an agent is unable to recall what a customer has said, they can refer back to the transcript to check. Reducing the need to make an outbound contact asking them to repeat.
Helping clarify confusing statements or terminology that might otherwise lead to errors.
Handling accents and difficult audio
Sometimes phone connections are poor or accents make understanding difficult. Transcription helps bridge these communication gaps.
Improving comprehension for agents reviewing difficult conversations.
Focusing on the conversation
Instead of taking notes, you can focus on active listening while knowing that the transcript will capture the details. The transcript can then be summarized automatically and available at the end of the contact.
Creating more engaged customer interactions without worry about missing details.
Learning from past segments
While on a chat contact, if you need to reference something the customer mentioned earlier, you can scroll up in the transcript. For a voice contacts, agents can review the complete transcript during After Contact Work to reference information from earlier in the call. You can analyze conversation flow to identify patterns or missed information from completed transcripts.
Allowing more informed follow-up actions and improvements to customer journeys
Demystifying the conversation
Do your agents or customers commonly use acronyms, jargon, or other industry specific terms? It's common to have annual percentage rate said as APR, or explanation of benefits said as EOB. Custom vocabulary enables transcription of full words or phrases. This feature makes it easier for agents, supervisors, and quality analysts to better understand the complete context of conversations.
Specialized terminology appears accurately in documentation for improved training and analysis.
Supervisor assistance
Supervisors can access completed transcripts to provide targeted feedback and coaching to agents. This ensures guidance focuses on actual conversation details rather than recalled information.
Providing more effective training, when based on accurate records of customer interactions.
Best practices
The following are some best practices when using real-time transcription:
Do not rely on it exclusively – While transcription is remarkably accurate, it's not perfect. Use it as a tool to complement your listening skills, not replace them.
Verify critical information – For important details like credit card numbers or medical information, always verify directly with the customer.
Use it for follow-up – Reference the transcript to ensure you've addressed all the customer's concerns before ending the call.
Check for errors – Occasionally review the transcript for any obvious errors, especially with unusual names or technical terms.
Sentiment Analysis Understanding the Emotional Journey
Amazon Connect Contact Lens analyzes the sentiment of both the customer and the agent in a conversation as positive, negative, or neutral. It then considers the following two factors for each participant to assign a score that ranges from -5 to +5 for each period of the call:
Frequency – The number of times the sentiment is positive, negative, or neutral.
Sentiment streaks – The consecutive turns with same sentiment.
The overall sentiment score is the average of the scores assigned during each portion of the call.
Figure 41 Sentiment Analysis Understanding the Emotional JourneySelect image to enlarge
Agent Nikki handles a call from an initially frustrated customer. Amazon Connect Contact Lens detects negative sentiment at the start of the call. However, by midway through the call, the sentiment shifts to neutral and ends positive. This showcases Nikki's effectiveness at turning around negative situations.
Review the sentiment trend in the following figure and the explanations for its numbered points.
Figure 42 Sentiment Analysis Understanding the Emotional JourneySelect image to enlarge
Call start
Customer calls with an issue about their account.
Midpoint
Customer is frustrated at not being able to resolve her issue.
Resolution
The agent was able to resolve the customer's issue.
Completion
The customer is pleased with the outcome.
With sentiment analysis, you can see shifts happening in real time and adjust your approach accordingly. This can mean the difference between losing a customer or regaining a customer's trust. If this information was only available after the call completed, it could be too late to regain the customer's trust.
Post Contact Summarization with Amazon Connect Contact Lens
After contact work (ACW) significantly impacts agent productivity. Trying to remember all the details of a call while the next customer is waiting is stressful. Streamlining these post-interaction tasks can save your contact center substantial operational hours throughout the year, which leads to improved efficiency and reduced costs.
Common ACW Tasks
ACW typically involves the following:
Summarize what the customer needed.
Document what actions the agent took.
Record any follow-up tasks.
This documentation is crucial for several reasons. It helps agents who might handle the customer in the future, provides valuable data to identify trends, and creates an audit trail for compliance purposes.
The next section describes common after contact work challenges.
Common ACW Challenges
Agents may face the following common ACW challenges:
Forget important details from the beginning of the call.
Spend too much time writing comprehensive notes.
Create inconsistent documentation that varies from agent to agent.
Feel rushed and make documentation errors.
Now that you have reviewed common ACW challenges, move on to the remaining content.
Contact Lens summarization process
Amazon Connect Contact Lens uses generative-AI to automatically generate summaries of customer interactions. Think of it as having an assistant who listens to your calls and creates notes for you and then gives you an elevator pitch to what happened.
Explore how this works behind the scenes.
Recording and transcription
The system captures the audio from both sides (agent and customer) of the conversation and converts it to text.
Natural language processing
AI analyzes the transcript to identify key elements like main issues, actions taken, outcomes, and follow-up items.
Summary generation
Using these identified elements, AI creates a concise, structured summary.
For example, instead of reading through a 10-minute transcript, you might see a summary like: "Customer called about missing reimbursement for canceled flight. Agent verified eligibility and processed $250 refund. Customer will receive email confirmation within 48 hours."
Using Contact Lens summarization
The following figure and numbered explanations describe how Contact Lens summarization supports daily operations.
Figure 43 Post Contact Summarization with Amazon Connect Contact LensSelect image to enlarge
During the call
Focus entirely on helping your customer. You don't need to take extensive notes or worry about remembering every detail.
After the call ends
While you are in ACW mode, the system starts generating a summary.
Review
When the summary is ready (usually within seconds), it will appear ready to review.
Once you have reviewed the summary, you can close the contact and move on to your next customer. This is much faster than writing notes from the beginning. The summaries follow a consistent format that highlights issues, outcomes, and action items, which makes it easier for anyone reviewing the contact later.
Bringing It All Together The AI Enhanced Agent Experience
Call begins
As soon as the customer starts speaking, real-time transcription begins capturing their words, and sentiment analysis starts tracking their emotional state.
Issue identification
Amazon Q in Connect uses the transcript to provide real-time assistance, including solutions and next best actions. Contact Lens automatically identifies the issue for automated disposition and analytics.
Personalization
You are alerted to the negative customer sentiment and adjust your tone and approach to be more empathetic.
Resolution
As you work toward a solution, you can see sentiment improving in real time, which confirms you're on the right track
Documentation
After the call, the transcript and sentiment analysis contribute to an AI-generated post-contact summary. This streamlines your ACW.
Real-time sentiment analysis and AI suggestions transformed this standard billing dispute into an opportunity for exceptional service. When an agent receives automated guidance and early warning signals of customer dissatisfaction, they can prevent escalations and deliver faster resolutions.
This technology-driven approach reduces operational costs while boosting both customer satisfaction and agent confidence.
RECALL & APPLY
Practice what you learned
Try answering in your own words, then compare with the suggested answer. These activities are for self-study and are not automatically graded.
01Which agent activity should AI assistance improve first in your environment?
Topic: AI assistance for contact center agents
Reveal suggested answer
Accept a defined workflow problem with an observable baseline, such as time spent searching or incomplete case notes.
02How do these capabilities support an inexperienced agent?
Topic: Agent challenges and AI support
Reveal suggested answer
They reduce information gathering and procedural uncertainty while the agent remains responsible for appropriate customer handling.
03Which measures would you compare before and after deployment?
Topic: Measuring agent assistance
Reveal suggested answer
Use a consistent baseline and comparable contact types. Review efficiency, resolution, customer experience, and agent outcomes together.
04What could cause an assistant to give an irrelevant recommendation?
Topic: Amazon Q in Connect assistance
Reveal suggested answer
The detected issue, retrieved source, knowledge quality, or configuration may be wrong. Investigate those stages separately.
05Why does the domain association matter during troubleshooting?
Topic: Domains, knowledge bases, and AI agents
Reveal suggested answer
The contact must use the intended assistant configuration and knowledge sources. The wrong association can produce missing or irrelevant assistance.
06What should a voice demonstration verify before testing recommendations?
Topic: Guided workflows and flow configuration
Reveal suggested answer
Verify the relevant recording and analytics behavior, Amazon Q block, domain association, and the agent’s access.
07The feature is enabled but an agent cannot see it. What do you check?
Topic: Agent access and workspace
Reveal suggested answer
Check the agent’s security profile, workspace or embedded application configuration, and the assistant association.
08Should an agent accept every transcribed account number without checking?
Topic: Real time transcription
Reveal suggested answer
No. Speech recognition can make errors. Confirm critical identifiers through the organization’s established process.
09Why should a supervisor inspect the conversation behind a sentiment score?
Topic: Sentiment across a conversation
Reveal suggested answer
The score is an estimate. The surrounding language and issue context are needed to decide whether and how to intervene.
10What should an agent verify before using a generated summary?
Topic: Post contact summaries
Reveal suggested answer
Verify the customer issue, actions, outcome, commitments, and any follow up. Correct omissions or unsupported statements.
11Describe an effective handoff from AI assistance to agent judgment.
Topic: The assisted agent journey
Reveal suggested answer
The assistant supplies relevant facts and guidance. The agent checks applicability, responds with empathy, and decides how to handle exceptions.
Capture a key idea, an example, or a question for your instructor.
Introduction to AI Powered Supervisor Capabilities
Supervisors Role and AI Technologies
AI's role for supervisors
AI in Amazon Connect Customer combines advanced technologies that analyze customer interactions in real-time and post-contact. Unlike traditional tools that simply record calls or track metrics, AI-powered solutions can actually understand conversations and identify important patterns. For example, imagine a customer calling about a billing issue. Traditional tools might just record the call and track how long it lasted. Amazon Connect has built-in AI that can detect payment disputes and rising customer frustration, alerting supervisors before issues escalate.
Amazon Connect Customer uses the following key AI capabilities to support supervisors.
Core AI technologies in Amazon Connect Customer
Review the following core AI technologies in Amazon Connect Customer.
Generative AI
Technology that can create new content based on patterns learned from training data.
Generative AI can produce human-like text summaries, categorizations, and other content derived from customer interactions.
For example, a customer has a 15-minute call with an insurance company discussing a complex claim involving water damage, contractor estimates, and coverage questions. After the call, generative AI analyzes the conversation and creates a concise summary with key issues and actions. The customer care agent reviews the summary, and makes necessary edits. The AI system files it while drafting follow-ups that may be needed later.
Natural language processing
Natural language processing (NLP) allows computers to understand and meaningfully respond to human language through artificial intelligence.
NLP bridges the gap between human communication and computer understanding by analyzing text or speech.
For example, a customer calls a telecom company's support line and says, "My internet keeps dropping every evening." The contact center's NLP system identifies service disruption as the intent, identifies evening as the time, routes to tech support, and retrieves account data for the support agent.
Natural language understanding
Natural language understanding (NLU) is a specialized subset of NLP. NLU technology is focused specifically on comprehending the meaning and intent behind language.
NLP handles the overall processing of language, whereas NLU drills deeper into extracting specific intents, entities, and meaning.
For example, a customer calls a telecom company's support line and says, "My internet isn't working right when I need it most." The NLU system interprets vague complaints as service issues, infers evening from context, detects unstated frustration, and prioritizes routing accordingly.
Automatic speech recognition
Automatic speech recognition (ASR) converts spoken language into text. ASR uses acoustic modelling to transcribe speech accurately. It handles various accents and filters background noise.
ASR enables real-time call transcription, making voice conversations searchable and analyzable. Supervisors can quickly scan conversation content without listening to recordings. This technology supports sentiment analysis on voice channels.
For example, a customer calls an airline's support line and says, "I need to change my flight to Dallas on Thursday." The contact center's ASR system accurately transcribes the spoken words into text in real-time, despite the customer's accent and background noise from an airport. This transcription instantly reaches the agent and booking system, which searches for Dallas Thursday flights before the agent responds.
Sentiment analysis
This technology analyzes transcribed text using NLP to detect positive, negative, or neutral sentiment in customer and agent conversations, helping supervisors identify interactions that might need attention.
For voice calls, the system first transcribes speech to text, then NLP classifies each speaker turn as positive, negative, or neutral based on the words used. For chat, text is analyzed directly. A score between -5 (most negative) to +5 (most positive) is generated for each portion, with the overall score calculated as the average across all portions.
For example, a customer calls a bank's support line saying, "I've been trying to access my account all day and nobody has helped me." The sentiment analysis system analyzes the transcribed text using NLP and classifies the turn as negative based on the language patterns. The system flags the interaction as having negative sentiment, alerting supervisors.
Speech analytics
Speech analytics analyzes voice data for insights beyond basic transcription. Speech analytics is broader than sentiment analysis, evaluating aspects like tone, pace, non-talk time, interruptions, and conversation dynamics.
Sentiment analysis and speech analytics are different but related technologies. Sentiment analysis uses NLP on transcribed text to classify each speaker turn as positive, negative, or neutral based on language patterns.
For example, a customer calls a healthcare provider about a billing issue. During the call, speech analytics detects extended silences, faster customer speaking, and cross talk. The system flags this call for review, revealing communication issues and suggesting agent training needs.
Text analytics
Text analytics derives meaningful patterns and insights from written text. Text analytics identifies topics, themes, and anomalies across text-based interactions.
For example, the text analytics system scans thousands of weekend customer chats and detects a 400 percent rise in checkout error messages. The system alerts IT to investigate payment processing issues before social media complaints begin. The system also notifies managers to prepare response strategies for this emerging problem.
Knowledge retrieval and question answering
Knowledge retrieval systems locate information from knowledge bases. Question answering tools extract specific answers from content. Both technologies respond intelligently to natural language queries.
For example, a customer calls a travel agency asking, "What's your cancellation policy for Mediterranean cruises during hurricane season?" The knowledge retrieval system quickly analyzes the question and displays policy information on the agent's screen. The agent provides immediate, accurate details without putting the customer on hold.
PII detection and redaction
This technology identifies and can remove sensitive personally identifiable information (PII) from text and recordings.
These algorithms recognize patterns associated with sensitive information like credit card numbers or addresses.
For example, a banking customer sends a chat message: "I can't access my account. My username is jstiles1980, my password is Spring2023!, and my account number is 1111-2222-3333.
ML for pattern recognition
This technology uses ML to identify patterns in customer interactions and contact center data.
These algorithms learn from historical data to recognize trends, themes, and clusters across conversations.
For example, a credit card company's ML system analyzes millions of service interactions over the most recent six-month period. The system discovers customers interactions that discuss international travel without mentioning transaction fees often result with customers calling back with disputes. The contact center creates a new protocol for agents to explain these fees during travel conversations. This change reduces follow-up calls and boosts customer satisfaction scores.
AI tools in Amazon Connect Customer do not replace effective supervisors, they make them more efficient. These technologies handle routine tasks and spot important patterns automatically. Supervisors can now focus on what matters most for your business.
The most effective contact centers combine human skills with AI capabilities. This partnership helps teams respond faster and understand customers better. As tech evolves, one constant remains: when skilled people use intelligent tools, everyone wins.
Shifting from Reactive to Proactive Monitoring
Quality Monitoring
Traditional quality management and monitoring are largely reactive. Supervisors typically review a small percentage of calls after they occur or respond when an agent asks for help. AI transforms this model by enabling a proactive approach with real-time monitoring, alerts, and data driven insights.
For example, without AI, you might discover a compliance issue days after it occurred. With Amazon Connect Customer AI capabilities such as Conversational Analytics, you can receive an alert the moment an agent forgets to read a required disclosure. This facilitates immediate issue resolution. Another is the ability to gain insights across the entire contact center while not requiring the manual evaluation of contacts to see trends.
With traditional supervision, interventions happen only after issues are identified through review, potentially too late to change the outcome.
AI-powered supervision allows real-time monitoring. Supervisors can detect issues during calls before they escalate. This prevents discovering problems through post-call reviews when intervention is too late. The key difference is the sequence. AI detects and alerts sooner than manual review therefore allowing earlier action.
Figure 44 Quality MonitoringSelect image to enlarge
Amazon Connect Customer AI tools eliminate traditional limitations by doing the following:
Simultaneously analyzing every customer interaction
Automatically identifying patterns across thousands of conversations
Surfacing the specific interactions that need human attention
Benefits of AI to Supervisors
Amazon Connect Customer AI capabilities streamline supervisor tasks by automating contact monitoring and analysis. This transforms traditional quality management processes that relied on manual call reviews and report creation.
These efficiencies translate directly to cost savings because of the following:
Reduced time spent on routine monitoring (from hours to minutes)
More focused effort on coaching and agent development
Faster identification and resolution of emerging customer trends and issues
Enhanced quality management
AI transforms quality management by automating routine evaluations, freeing supervisors to invest their expertise where it matters most.
Complete interaction coverage: Instead of reviewing a random 2–3 percent sample of calls, AI can analyze 100 percent of interactions, making sure that no critical issues are missed.
Consistent measurement: AI can provide consistent analysis across all interactions, eliminating human bias in evaluations.
Automated scoring: Many aspects of quality evaluation can be automated, such as compliance statement detection or greeting verification.
Real time intervention possibilities
Perhaps the most transformative benefit is the shift from after-the-fact review to real-time intervention.
Live alerts: Supervisors can receive notifications when AI detects issues like customer frustration, compliance risks, or agents struggling.
In-call assistance: When alerted to an issue, supervisors can provide real-time guidance through chat or call barging.
Proactive intervention: AI can detect which calls might become problematic before issues escalate by detecting changes in customer sentiment.
Voice of the customer: Amazon Connect Customer AI can transform routine customer calls into valuable voice-of-customer data. It automatically identifies common frustration points across thousands of conversations. Amazon Connect Customer can categorize these points by key features such as product or service type and provide sentiment assessment based on those categories.
Marketing insights: AI analytics reveal unmet customer needs before they appear in formal research. Amazon Connect Customer can highlight customer confusion around specific product terms or benefits. These insights help refine campaign targeting based on real customer pain points.
Strategic advantage: The AI analytics in Amazon Connect Customer mines valuable business intelligence from every customer conversation. This tool uncovers competitor mentions that agents might miss. It reveals genuine customer sentiment about pricing strategies. The system exposes emerging market opportunities hidden within daily interactions.
Without AI, these insights would remain buried in conversation data.
For example, a supervisor might spend 15 hours weekly reviewing a limited sample of calls. With automation analyzing 100 percent of interactions, the supervisor can now focus on strategic coaching to enhance customer experience metrics. This helps them make data-driven decisions based on comprehensive insights rather than limited snapshots.
Introduction to Conversational Analytics
Amazon Connect Customer Conversational Analytics
Conversational Analytics overview
Amazon Connect Customer Conversational Analytics is an AI-powered analytics solution built into Amazon Connect Customer. It uses machine learning to analyze conversations between customers and agents across voice and chat channels, providing insights that would be impossible to gather manually at scale. At its core, Conversational Analytics acts like an extremely attentive listener that can process thousands of conversations simultaneously. Conversational Analytics transcribes speech to text and analyzes the language to understand meaning and identify important patterns.
Figure 45 Conversational Analytics overviewSelect image to enlarge
AI generated contact summary
Contact Lens uses generative AI to create concise summaries of customer interactions.
Conversational analytics
Customer sentiment trend: Shows how customer sentiment changes as the contact progresses
Customer sentiment: Shows the distribution of customer sentiment for the entire call
Talk time: Shows the distribution of talk time and non-talk time during the entire call—talk time is further split into agent and customer talk time
Audio analysis
Interactions can be recorded and are split into Customer, Agent, or System/Bot. Customer and Agent sentiment are marked in the timeline of the recording.
Transcript Key highlights and categories
Transcript automatically identifies and labels key parts of customer conversations and displays highlights of the conversations. Managers can view those highlights on the Contact details page.
Contact Lens rules let you automatically categorize contacts, receive alerts, or generate tasks based on keywords.
How Conversational Analytics works with Amazon Connect Customer
Conversational analytics works in two primary modes:
Post-contact analysis: After a conversation ends, Conversational Analytics processes the full interaction to provide comprehensive insights, transcripts, and analytics.
Real-time analysis: During live conversations, Conversational Analytics can analyze the interaction as it happens, alerting supervisors to potential issues that might require immediate intervention.
Conversational Analytics features
The following capabilities work across multiple channels (voice calls, chat conversations, and more), giving you a comprehensive view of all customer interactions:
Conversational Analytics with AI-powered summaries, transcription, and sentiment analysis
Performance evaluation with automated assessments and calibration tools
Real-time monitoring capabilities with supervisor alerts and screen recording
Search, recording, and supervisor intervention tools
Enabling Contact Lens in an Amazon Connect Customer instance
The following demonstration steps help you understand how to enable Contact Lens in a Connect Customer instance.
Welcome to this demonstration on enabling Contact Lens in an Amazon Connect instance.
To start, sign in to the AWS Management Console.
In the search box at the top of the page, enter Amazon Connect. In the displayed list of Services, choose Amazon Connect.
On the Amazon Connect page, the instances in your account are displayed. Choose the Instance alias of the instance that you want to enable Contact Lens on.
Next, in the navigation pane on the left of the screen. Under Applications, choose Analytics tools.
On the Analytics tools page, notice the state of Contact Lens. In this example, Contact Lens is not enabled. To access the settings for Contact Lens, choose Edit.
To enable Contact Lens, select the checkbox. Then choose Save.
On the Analytics tools page, a success alert indicates your instance features have been updated. Notice that the status for Contact Lens has changed to Enabled.
This concludes the demonstration on enabling Contact Lens in an Amazon Connect Customer instance.
Thank you for your participation.
Conversational Analytic Security Features
Conversational Analytics incorporates comprehensive security protocols designed specifically for sensitive contact center environments. Conversational Analytics safeguards customer conversations through the following protective layers that work seamlessly together:
Sensitive data redaction: This feature identifies and removes sensitive data like credit card numbers, names, and other PII information.
Data encryption: All data is encrypted both in transit and at rest.
Retention controls: Organizations can set data storage policies.
Insider tip: Ensure that you configure access permissions through security profiles to determine which users can access Contact Lens insights. To learn more about security profiles, see Security Profiles for Connect Customer and Contact Control Panel (CCP) Access.
These security features help ensure that organizations in regulated industries can confidently deploy Amazon Connect Customer while maintaining their compliance obligations and protecting sensitive customer information.
Sentiment Analysis, Issue Detection, and Summarization
Understanding sentiment analysis
Sentiment analysis is like having an emotional thermometer for every conversation. It automatically detects the sentiment of interactions, helping you respond appropriately before issues escalate. Conversational Analytics analyzes transcribed text using NLP to determine whether sentiment is positive, neutral, or negative based on the language patterns in each speaker turn. The system tracks sentiment throughout the conversation, showing how it changes over time.
Figure 46 Sentiment Analysis, Issue Detection, and SummarizationSelect image to enlarge
Sentiment analysis example
Sentiment analysis estimates the sentiment of what the customer and agent are saying throughout the interaction. This metric is represented as both a quantitative value (with a range from -5 to +5) and a qualitative value (positive, neutral, mixed, or negative). Scores are based on language patterns in the transcribed text, weighted by both the frequency of positive/negative turns and sentiment streaks (consecutive turns with the same sentiment).
Real time sentiment alerts
Conversational Analytics sends immediate notifications when a conversation's sentiment turns negative, empowering managers to join in-progress contacts and help resolve issues before they escalate.
For example, when a customer chat shows rapidly deteriorating sentiment, a supervisor can receive an alert and review the transcript. The supervisor can then either coach the agent privately or join the conversation directly.
Figure 47 Sentiment Analysis, Issue Detection, and SummarizationSelect image to enlarge
Using sentiment data for coaching
Sentiment analysis provides valuable data for agent coaching and training. By setting up agent alerts, you can use real-time sentiment data for coaching in the following ways:
Identify which agents consistently maintain positive customer sentiment.
Recognize conversation patterns that tend to improve sentiment.
Spot common triggers that lead to negative sentiment.
Create training scenarios based on real interactions that show sentiment shifts.
Issue detection
Issue detection works like an early warning system, automatically flagging potential problems before they become serious.
Conversational Analytics analyzes conversations in real-time, looking for patterns that indicate potential problems including the following:
Technical difficulties described by customers
Requests for supervisor escalation
Mentions of cancellations or refunds
Specific product problems or recurring complaints
Expressions of significant frustration
Real time issue alerts
When Conversational Analytics detects a potential issue, it can do the following:
Send alerts to supervisors.
Categorize the issue type automatically.
Provide relevant context for whoever steps in to help through Tasks.
Trend analysis for systemic issues
Beyond individual contacts, issue detection also consolidates data to help identify systemic problems and trends. Issue groupings include the following:
Product defects affecting multiple customers
Website or app functionality problems
Confusing policies or processes
Training gaps among agents
Seasonal or campaign-related issues
Summarization
Conversational Analytics uses generative AI to create concise summaries of customer interactions. These summaries highlight the following:
Main issues discussed
Actions taken by the agent
Outcomes of the interaction
Follow-up items or commitments
Key customer information or requests
Figure 48 Sentiment Analysis, Issue Detection, and SummarizationSelect image to enlarge
Benefits of automated summarization
Time efficiency: Understand contact outcomes without reviewing entire transcripts or recordings
Structured insights: Highlight customer issues, agent actions, and interaction outcomes
Objective representation: Focus on facts instead of subjective interpretations
Performance tracking: Identify common customer issues and agent coaching opportunities
Compliance Monitoring
Compliance monitoring overview
Conversational Analytics uses advanced speech and text analysis that you can use to identify contacts for compliance with various requirements. Compliance monitoring provides peace of mind in a complex regulatory landscape.
By automatically checking every interaction against compliance requirements, Conversational Analytics helps organizations reduce risk, protect customer information, and demonstrate their commitment to regulatory standards. This proactive approach not only helps avoid penalties but also builds customer trust through consistent, compliant interactions.
Common compliance examples are as follows:
Required disclosures and statements
Prohibited language or topics
Sensitive data collection processes
Authentication procedures
Required information collection
For example, in a financial services contact center, agents might be required to verify identity using specific questions and inform customers about call recording. conversational analytics can automatically check whether actions were completed in each interaction.
Sensitive data protection
Conversational Analytics also helps protect sensitive customer information by automatic redaction of sensitive personal information from call recordings and transcripts.
Beyond checking for required actions, compliance monitoring also helps protect sensitive customer information with the following features:
Automatic redaction: Conversational Analytics can automatically redact sensitive personal information from transcripts and recordings.
PII detection: The system identifies when personally identifiable information is shared.
Compliance reporting: Aggregate data shows compliance rates across agents and teams.
This protection works for both recorded calls and chat transcripts, helping to ensure sensitive information is handled appropriately across all channels.
For example, Conversational Analytics can automatically redact credit card numbers spoken by customers during calls from recordings and transcripts. This feature helps you maintain payment card industry (PCI) compliance.
Compliance consideration
Although the Conversational Analytics redaction feature uses sophisticated technology to protect sensitive information, human verification remains essential to validate complete compliance. The following are examples of situations they might require verification:
Review redacted output
Because of the predictive nature of machine learning, Conversational Analytics might not identify and remove all instances of sensitive data in a generated transcript. You should review any redacted output to ensure the transcript meets your compliance requirements. For more information, see Use sensitive data redaction to protect customer privacy using Contact Lens.
Scrubbing PII or PCI
If PII or PCI data is captured in call recordings, the sensitive data must be scrubbed from the recording and obfuscated from any logs or transcriptions. For more information, see Best practices for PCI compliance in Amazon Connect.
Conversational Analytics Expanded Channels and AI Interactions
Conversational Analytics for Email
The Conversational Analytics capabilities you learned about now extend to additional areas. For the email channel, categorization, PII redaction, and summarization are available. For AI agent (bot) interactions, sentiment analysis, summarization, and compliance monitoring apply.
Conversational Analytics automatically processes email interactions alongside voice and chat. This unifies analytics across all customer communication channels.
What it does
For every email interaction, the system automatically:
Categorizes the email by topic and intent
Redacts PII (personally identifiable information) from stored records
Generates contact summaries so supervisors can quickly understand the interaction without reading entire email threads
This means your analytics workflows for categorization, summarization, and PII redaction now apply to email alongside voice and chat.
Why it matters for supervisors
Email interactions are often longer and harder to review than a quick chat or call. A single email thread might span multiple days and dozens of exchanges. AI-generated summaries make these reviewable at a glance, while automatic categorization helps identify trends across the email channel.
PII redaction ensures that stored analytics data remains compliant with privacy requirements, even when customers include sensitive information in their emails.
Conversational Analytics for AI Agents
Conversational Analytics also covers self-service interactions including conversations handled by AI agents across voice, chat, and messaging channels.
This is a fundamental shift. Previously, analytics focused exclusively on human agent performance. Now supervisors can apply the same analytical lens to AI agent interactions:
Sentiment analysis for bot conversations — are customers getting frustrated with the AI?
Redaction — PII is protected in bot interactions just as it is in human ones
Issue identification — where are AI agents struggling or generating poor outcomes?
This gives supervisors a unified view of the entire customer experience, regardless of whether a human or AI agent handled the interaction. Patterns become visible: if customers consistently express negative sentiment during a specific type of AI agent interaction, that signals a tuning opportunity.
AI powered case summaries
Supervisors can also generate concise summaries that span multiple interactions for a single case or customer issue.
A customer might contact your organization three times about the same billing dispute:
Figure 49 Conversational Analytics for EmailSelect image to enlarge
Previously, a supervisor reviewing this case would need to read transcripts from all three interactions individually.
With AI-powered case summaries, supervisors generate a single concise summary covering all related interactions with one click. The summary captures the key facts, actions taken, and current status across the full history of the case.
This streamlines review workflows and helps supervisors quickly identify cases that need attention without spending time reading through multiple full transcripts.
Introduction to Real Time Analytics in Contact Centers
Understanding Real Time Analytics
Real time analytics overview
Real-time analytics refers to the collection, processing, and analysis of data as it is being generated. In contact centers, this means having access to up-to-the-minute information about customer interactions, agent performance, queue status, and overall operational metrics.
Unlike historical reporting that looks at what happened yesterday or last week, real-time analytics focuses on what is happening now. It is like the difference between watching a live sports game compared to reading about it in tomorrow's newspaper.
Key components of real time analytics
Review the following key components of real-time analytics:
Real-time dashboards
Dashboards transform complex data into understandable charts, graphs, and status indicators that update continuously.
A good dashboard is like the cockpit of an airplane. The dashboard shows you all the critical information you need to manage your contact center safely. Managers can quickly see if service levels are on target, if queues are growing too large, or if too many agents are unavailable.
Figure 50 Key components of real time analyticsSelect image to enlarge
Data collection systems
Real-time analytics begins with systems that capture information as it happens. These systems track every customer interaction across channels like phone, chat, email, and task. They monitor how long customers wait, what agents are doing, and the status of every interaction.
Amazon Connect Customer automatically collects all these data points without requiring any additional configuration. From the moment your instance is set up, the built-in telemetry of Amazon Connect Customer continuously gathers metrics on contacts, queues, and agent activities. These metrics are immediately available for real-time monitoring and analysis.
Think of it like sensors placed throughout your contact center, constantly feeding information to Amazon Connect Customer. These sensors might track when calls come in, how quickly they are answered, and what is happening during the conversation.
Alert systems
Real-time analytics become truly powerful when combined with alert systems that notify the right people when metrics fall outside acceptable ranges.
For example, if customer wait times suddenly spike beyond 5 minutes, the system could automatically send text messages to team leads. The system can also display visual alerts on the dashboard. You do not need to constantly monitor it, but you will know immediately when there is a problem.
Integration capabilities
Amazon Connect Customer real-time analytics seamlessly integrate with both AWS services and external systems to create a comprehensive supervisor system.
Real-time analytics directly connect with the Amazon Connect Customer Forecasting, Capacity Planning, and Scheduling capabilities.
Supervisors can use live metrics like occupancy rates and service levels to make immediate staffing adjustments through the built-in scheduling tool. This helps to ensure optimal coverage during unexpected contact volume fluctuations.
Amazon Connect Customer real-time metrics can be accessed through APIs and streamed through Amazon Kinesis. This allows integration with third-party workforce management tools, CRM platforms, and business intelligence systems.
For example, when real-time analytics detect an unusual spike in contacts about a specific issue, this data can trigger automated workflows in external systems.
Importance of real time analytics
With real-time analytics, supervisors can quickly respond to issues and make positive changes, such as the following:
Immediate problem resolution
When issues arise, every minute counts. Real-time analytics helps supervisors to identify problems instantly, such as system errors, call quality issues, or service disruptions. Then, supervisors can implement immediate corrective actions.
Dynamic resource allocation
Contact centers often experience unexpected surges in volume. With Real-time analytics, supervisors anticipate needs and shift resources where they are needed most, preventing potential bottlenecks before they impact customer experience.
Improved customer experience
When supervisors have visibility into what is happening right now, they can take steps to improve the customer experience immediately.
Agent support and coaching
Common mistakes to avoid with real time metrics
Even with powerful tools, organizations often make mistakes when implementing real-time analytics. Examples of these mistakes include the following:
Information overload
Trying to monitor too many metrics at once can lead to analysis paralysis. Focus on the vital few metrics that drive immediate action, not every possible data point.
Failing to act on insights
Real-time data is only valuable if it drives real-time action. Some organizations invest in sophisticated analytics tools but do not empower their supervisors to make quick adjustments based on what they see.
Neglecting training
Managers and supervisors need proper training to interpret real-time data and understand which actions are most appropriate for different situations.
Real Time Analytics Tools and Features
Real time dashboards
Have you ever watched a busy restaurant kitchen during the dinner rush? The head chef constantly monitors multiple cooking stations, checks food quality, manages timing, and coordinates the staff, all simultaneously. Contact center managers face a similar challenge, and like that chef, they need specialized tools to help them succeed.
The dashboard brings real-time analytics to life, serving as your contact center's command center for monitoring current activity.
Figure 51 Real time dashboardsSelect image to enlarge
Channel filter
To view metrics with a specific channel or channels, add a filter to limit the data shown in the dashboard.
Combine data
View combined graphs of data to give a more complete picture of what is happening in your contact center.
Customize view
Modify metrics displayed in each widget for a more customized view of your contact center.
Real-time dashboards include the following:
Channel specific metrics
Effective dashboards show metrics for channels, such as phone, chat, or email. These channels might highlight to indicate when one channel is performing significantly worse than others.
Queue status displays
These displays show how many customers are waiting, how long they have been waiting, and which service they are waiting for. This information helps managers make immediate staffing adjustments.
Customizable views
With advanced dashboards, users create personalized views based on their role and responsibilities.
Queue and agent performance dashboard
The Queue and agent performance dashboard display the real-time state for each agent. This is shown as Activity on the dashboard. Timers track duration in each state against performance thresholds. The dashboard has the ability to set color-coded indicators to show green, yellow or red custom threshold.
Service level report
Service level appears as a percentage of contacts answered within target time frames. For example, answering 80 percent of calls within 30 seconds would show as 80 percent for SL30. Time frames, queues and the performance indicator colors for Service Level (SL) metrics can all be configured on the report. In the report below you can see that only 1 percent of contacts have been answered with 15 seconds (SL15) for the AnyCompany Finance queue.
Real time alerts
Although dashboards provide visibility, alert systems help ensure that you do not miss critical changes even when you are not actively monitoring the dashboard.
Threshold based alerts
Trend based alerts
Anomaly detection
Alert delivery methods
Real time monitoring and intervention
Real-time monitoring and intervention is the ability to monitor individual interactions as they happen and intervene when necessary.
Live contact monitoring
Monitoring options
Speech and sentiment analysis
Supervisors can listen to live calls or read live chats to assess quality, identify training needs, or step in to assist with difficult situations.
Integrating Real Time Analytics with Other Systems
The effectiveness of real-time analytics multiplies when integrated with the following systems:
Workforce management systems
When real-time analytics shows high call volumes, workforce management integration can automatically identify available staff to call in or agents who could work longer shifts.
CRM integration
Real-time analytics can pull customer information from CRM systems to provide context for current interactions. For example, a simple dashboard might only show that call volumes are high. With a CRM integration, you might find that 80 percent of those calls come from high-value customers.
Quality management integration
When real-time analytics identifies potential quality issues in customer interactions, it can automatically flag these interactions for later review by the quality team.
Automated response systems
Real-time analytics can trigger automated system responses without requiring human intervention. For example, when wait times exceed thresholds, systems can automatically offer callbacks and update website messaging to set realistic response time expectations.
Practical applications of real time analytics
Managing unexpected volume fluctuations
Improving first contact resolution
Supporting agents in the moment
Making data-driven staffing adjustments
Real-time analytics help managers respond quickly by indicating the following:
Identifying volume surges as they begin
Showing which channels and contact types are most affected
Predicting the impact on service levels if no action is taken
Tracking the effectiveness of mitigation efforts
Practical example: When a retail company launches a new product, the contact center sees a 35 percent increase in call volume within the first hour. The manager takes the following immediate action:
Checks agent adherence to identify any scheduled agents who have not logged in.
Reassigns cross-trained agents from email to phone support.
Updates the IVR message to acknowledge higher-than-normal call volume.
Adds a banner to the website with answers to common questions about the new product.
Best practices for implementing real time analytics
Now you will review best practices for real-time analytics with Amazon Connect.
Start with clear objectives
Focus on actionable metrics
Design intuitive visual displays
Create tiered thresholds
Develop clear response protocols
Integrate with workforce management
Continuously refine your approach
Tactics that undermine success
Use the following strategies to overcome various factors that can undermine success in your contact center:
Information overload: Limit main dashboards to 5–7 key metrics with drill-down capabilities for details.
Alert fatigue: Carefully balance early warning with the amount of notifications and regularly adjust thresholds.
Focusing on technology instead of people: Allocate at least as much time to training and change management as to technical implementation.
Neglecting feedback loops: Create regular processes to review real-time data patterns and use them to drive systemic changes.
The Queue and agent performance dashboard display the real-time state for each agent. This is shown as Activity on the dashboard. Timers track duration in each state against performance thresholds. The dashboard has the ability to set color-coded indicators to show green, yellow or red custom threshold.
Figure 52 Queue and agent performance dashboardSelect image to enlarge
Service level report
Service level appears as a percentage of contacts answered within target time frames. For example, answering 80 percent of calls within 30 seconds would show as 80 percent for SL30. Time frames, queues and the performance indicator colors for Service Level (SL) metrics can all be configured on the report. In the report below you can see that only 1 percent of contacts have been answered with 15 seconds (SL15) for the AnyCompany Finance queue.
Figure 53 Queue and agent performance dashboardSelect image to enlarge
Real time alerts
Although dashboards provide visibility, alert systems help ensure that you do not miss critical changes even when you are not actively monitoring the dashboard.
Threshold based alerts
These alerts trigger when a metric crosses a predefined threshold, such as oldest contact age exceeding 10 minutes.
Trend based alerts
These notify you when metrics are moving in a concerning direction, like call volume increasing by 30 percent compared to the previous hour.
Anomaly detection
You could use anomaly detection in Amazon CloudWatch logs to detect unusual patterns from your contact center metrics like queue size and waiting times that might not trigger standard threshold alerts. To learn more about CloudWatch, see Amazon Monitoring and Observability.
Alert delivery methods
Alerts can be delivered through visual indicators on dashboards, tasks, emails, or events to Amazon EventBridge. To learn more about EventBridge, see Amazon EventBridge.
Real time monitoring and intervention
Real-time monitoring and intervention is the ability to monitor individual interactions as they happen and intervene when necessary.
Live contact monitoring
Monitoring options
Speech and sentiment analysis
Supervisors can listen to live calls or read live chats to assess quality, identify training needs, or step in to assist with difficult situations.
Integrating Real Time Analytics with Other Systems
The effectiveness of real-time analytics multiplies when integrated with the following systems:
Workforce management systems
When real-time analytics shows high call volumes, workforce management integration can automatically identify available staff to call in or agents who could work longer shifts.
CRM integration
Real-time analytics can pull customer information from CRM systems to provide context for current interactions. For example, a simple dashboard might only show that call volumes are high. With a CRM integration, you might find that 80 percent of those calls come from high-value customers.
Quality management integration
When real-time analytics identifies potential quality issues in customer interactions, it can automatically flag these interactions for later review by the quality team.
Automated response systems
Real-time analytics can trigger automated system responses without requiring human intervention. For example, when wait times exceed thresholds, systems can automatically offer callbacks and update website messaging to set realistic response time expectations.
Practical applications of real time analytics
Managing unexpected volume fluctuations
Improving first contact resolution
Supporting agents in the moment
Making data-driven staffing adjustments
Real-time analytics help managers respond quickly by indicating the following:
Identifying volume surges as they begin
Showing which channels and contact types are most affected
Predicting the impact on service levels if no action is taken
Tracking the effectiveness of mitigation efforts
Practical example: When a retail company launches a new product, the contact center sees a 35 percent increase in call volume within the first hour. The manager takes the following immediate action:
Checks agent adherence to identify any scheduled agents who have not logged in.
Reassigns cross-trained agents from email to phone support.
Updates the IVR message to acknowledge higher-than-normal call volume.
Adds a banner to the website with answers to common questions about the new product.
Best practices for implementing real time analytics
Now you will review best practices for real-time analytics with Amazon Connect.
Start with clear objectives
Focus on actionable metrics
Design intuitive visual displays
Create tiered thresholds
Develop clear response protocols
Integrate with workforce management
Continuously refine your approach
Tactics that undermine success
Use the following strategies to overcome various factors that can undermine success in your contact center:
Information overload: Limit main dashboards to 5–7 key metrics with drill-down capabilities for details.
Alert fatigue: Carefully balance early warning with the amount of notifications and regularly adjust thresholds.
Focusing on technology instead of people: Allocate at least as much time to training and change management as to technical implementation.
Neglecting feedback loops: Create regular processes to review real-time data patterns and use them to drive systemic changes.
Real-time monitoring and intervention is the ability to monitor individual interactions as they happen and intervene when necessary.
Live contact monitoring
Supervisors can listen to live calls or read live chats to assess quality, identify training needs, or step in to assist with difficult situations.
Monitoring options
Amazon Connect Customer offers the following monitoring options:
Silent monitoring: Supervisors can listen to live calls or read live chats without the agent or customer knowing.
Barging: Supervisors can join a live voice or chat conversation. For voice, the supervisor speaks to both agent and customer. For chat, the supervisor sends messages visible to both parties.
Speech and sentiment analysis
Amazon Connect Customer can analyze conversations in real-time, detecting customer sentiment, identifying key phrases, or flagging compliance issues.
Integrating Real Time Analytics with Other Systems
The effectiveness of real-time analytics multiplies when integrated with the following systems:
Workforce management systems
When real-time analytics shows high call volumes, workforce management integration can automatically identify available staff to call in or agents who could work longer shifts.
CRM integration
Real-time analytics can pull customer information from CRM systems to provide context for current interactions. For example, a simple dashboard might only show that call volumes are high. With a CRM integration, you might find that 80 percent of those calls come from high-value customers.
Quality management integration
When real-time analytics identifies potential quality issues in customer interactions, it can automatically flag these interactions for later review by the quality team.
Automated response systems
Real-time analytics can trigger automated system responses without requiring human intervention. For example, when wait times exceed thresholds, systems can automatically offer callbacks and update website messaging to set realistic response time expectations.
Practical applications of real time analytics
Managing unexpected volume fluctuations
Real-time analytics help managers respond quickly by indicating the following:
Identifying volume surges as they begin
Showing which channels and contact types are most affected
Predicting the impact on service levels if no action is taken
Tracking the effectiveness of mitigation efforts
Practical example: When a retail company launches a new product, the contact center sees a 35 percent increase in call volume within the first hour. The manager takes the following immediate action:
Checks agent adherence to identify any scheduled agents who have not logged in.
Reassigns cross-trained agents from email to phone support.
Updates the IVR message to acknowledge higher-than-normal call volume.
Adds a banner to the website with answers to common questions about the new product.
Improving first contact resolution
Real-time analytics helps improve first contact resolution (FCR) in the following ways:
Identifying when customers are repeatedly calling about the same issue
Highlighting knowledge gaps that prevent agents from resolving issues
Detecting when transfers between departments are increasing
Supporting agents in the moment
One of the most powerful applications is providing immediate support to agents when they need it most.
Practical example: A new agent has been on a call for 12 minutes, significantly longer than the average 5 minutes. The supervisor acts by doing the following:
Using silent monitoring to listen to the call
Recognizing the customer is asking about a complex product feature the agent has not been trained on
Using barge-in mode to guide the customer and agent through the process
After the call, briefly meets with the agent to reinforce the learning
Making data-driven staffing adjustments
Data can be useful for staffing adjustments in the following ways:
Allowing immediate response to unexpected staffing issues, such as call-outs or traffic surges
Allowing intra-day adjustments instead of waiting for the next scheduling cycle
Providing visibility into emerging patterns before they become established trends
Supporting dynamic reallocation of staff across departments as needs shift
Historical data analysis remains valuable for long-term planning and identifying seasonal patterns. The responsiveness that real-time solutions offer for day-to-day workforce optimization cannot be matched by historical analysis.
Best practices for implementing real time analytics
Now you will review best practices for real-time analytics with Amazon Connect.
Start with clear objectives
Define exactly what you want to achieve before implementing real-time analytics.
Focus on actionable metrics
Prioritize metrics that drive immediate action when they change, connect to customer experience, and can be influenced by real-time decisions.
Design intuitive visual displays
Use color coding consistently, group related metrics, use appropriate visualization types, and limit information on each screen.
Create tiered thresholds
Not all issues require the same response or urgency. Create a tiered system, such as the following:
Green: Metrics are not exceeding thresholds.
Yellow: Metrics crossed thresholds but are not critical.
Red: A significant service impact requires immediate action.
Develop clear response protocols
For each type of threshold, develop standard response protocols so managers know exactly what actions to take.
Integrate with workforce management
Make real-time analytics even more powerful by integrating with workforce management systems like Amazon Connect Customer forecasting, capacity planning, and scheduling.
Continuously refine your approach
Regularly review and adjust dashboard layouts, alert thresholds, and response protocols.
Tactics that undermine success
Use the following strategies to overcome various factors that can undermine success in your contact center:
Information overload: Limit main dashboards to 5–7 key metrics with drill-down capabilities for details.
Alert fatigue: Carefully balance early warning with the amount of notifications and regularly adjust thresholds.
Focusing on technology instead of people: Allocate at least as much time to training and change management as to technical implementation.
Neglecting feedback loops: Create regular processes to review real-time data patterns and use them to drive systemic changes.
Note - AI-powered manager assistance is currently available as a preview feature. To request access, contact your AWS account team or an AWS representative.
The real-time analytics tools include AI-powered manager assistance (preview), which lets supervisors ask natural language questions about their metrics instead of navigating dashboards manually.
This feature allows supervisors to ask questions like, "Why is average handle time increasing?" and receive instant diagnostic answers drawn from over 150 metrics.
How it works
Instead of clicking through dashboard filters and visually scanning for anomalies, supervisors type or speak questions in natural language. The AI assistant:
Interprets the question
Queries the relevant metrics across queues, agents, and time periods
Identifies the likely cause
Presents a concise answer with supporting data
For example: "Which queue has the longest wait time right now?" or "Are there any agents who haven't taken a call in the last 30 minutes?"
What kinds of questions it answers
The assistant can address:
Diagnostic — "Why is service level dropping in the billing queue?"
Status — "How many agents are currently available?"
Comparison — "How does today's call volume compare to last Tuesday?"
Alerting — "Is anything unusual happening right now?"
It draws from 150+ metrics to form answers, so supervisors do not need to know which specific metric to look at; they describe what they want to understand.
Impact on supervisor workflow
This shifts supervisors from reactive dashboard monitoring to proactive, question-driven management. Rather than waiting for an alert or noticing a number change, they can investigate hunches, confirm suspicions, and diagnose issues conversationally.
Additional enhancements
Several additional enhancements make real-time analytics more powerful and customizable.
Custom metrics no code
Supervisors can now create custom metrics using mathematical operations without technical skills. For example, combining existing metrics to calculate a "first-contact resolution rate" specific to your organization's definition. No developer involvement needed.
Custom business dimensions
Metrics can now be filtered by business divisions, product lines, or customer segments. Instead of seeing aggregate data across the entire contact center, a supervisor can focus on "enterprise customers in the returns queue" or "small business accounts handled by Team A."
Enhanced real time alerts
Alerts now include the specific context that triggered them, such as which agents, queues, flows, or routing profiles are involved. Previously, an alert might say "service level dropped below 80%." Now it says "service level dropped below 80% in the Billing queue, affecting agents Smith, Park, and Chen."
Flow designer analytics
Supervisors can view aggregate traffic through each step in contact flows, identifying where customers drop off, which branches handle the most volume, and where errors occur. This helps optimize flow design based on real usage patterns rather than assumptions.
Introduction AI Agent Monitoring and Enhanced Evaluations
AI Agent Monitoring and Analytics
AI agent dashboards
AI agent monitoring dashboards provide supervisors with key performance indicators specific to autonomous AI interactions.
Key dashboard metrics
The dashboards surface the following metrics for AI agent interactions:
Hand-off rate — percentage of interactions transferred to a human agent
Conversation turns — average number of exchanges per interaction
Goal success rate — percentage of interactions resolved successfully
Faithfulness score — whether responses are grounded in source material
These metrics update in real time, giving supervisors an always-current view of AI agent performance.
Version comparison
When you update an AI agent's configuration, such as changing its knowledge bases, adjusting guardrails, or modifying its tools, the dashboard lets you compare performance between versions side by side.
This answers the critical question: "Did our change make things better or worse?" You can see whether the new version improved goal success rate, reduced hand-offs, or inadvertently increased conversation turns.
Difference from human agent monitoring
Human agent dashboards focus on availability, adherence, and individual performance coaching. AI agent dashboards focus on:
System-level quality (is the agent working correctly?)
Configuration effectiveness (are the knowledge bases and tools sufficient?)
Boundary detection (where does the agent fail and need human takeover?)
The supervisory action differs too. For human agents, you coach. For AI agents, you tune
When to intervene
Dashboards show you what is happening. The next step is knowing what to do about it. The following patterns indicate an AI agent needs attention.
Signal
What it means
Action to take
Rising hand-off rate
Agent cannot resolve a growing number of requests
Check for new customer query types not covered by knowledge bases
Dropping faithfulness score
Agent is generating unsupported responses
Review and expand knowledge base content; tighten guardrails
Increasing conversation turns
Agent is taking longer to resolve issues
Analyze transcripts for confusion loops; improve tool selection
Low goal success rate
Agent is failing to resolve requests
Investigate specific failure categories; consider adding tools
Negative sentiment in bot analytics
Customers are frustrated with the AI
Review interaction patterns; adjust tone and escalation thresholds
Understanding the faithfulness score
The faithfulness score measures how closely an AI agent's responses align with its configured knowledge sources. Scores range from 0 to 1, where higher values indicate stronger grounding in approved content. When faithfulness scores drop below acceptable thresholds, supervisors should investigate the agent's traces to identify whether the agent is generating unsourced responses or hallucinating information.
AI agent traces
For deeper investigation beyond what dashboards provide, supervisors and developers can access AI agent traces through APIs.
Traces provide a step-by-step record of an AI agent's reasoning and actions during an interaction. They show:
What the agent "thought" at each step (its reasoning)
Which tools it considered and selected
What data it retrieved from knowledge bases
Why it made the decisions it made
This level of detail is essential for diagnosing specific failures. When the dashboard shows a problem, traces help you understand the root cause so you can fix it precisely rather than guessing.
Enhanced Evaluations
Multilingual evaluation support
Automated evaluations previously supported English only. They now operate in five additional languages, with cross-language evaluation support.
Portuguese
French
Italian
German
Spanish
Cross-language evaluation means a supervisor working in English can evaluate interactions that occurred in any of the supported languages. The system handles the translation and analysis, so language barriers do not prevent quality oversight.
New evaluation question types
Evaluation forms now support additional question types beyond the original format, enabling more nuanced assessment.
Multiple choice questions
Evaluators can now include multiple choice questions in evaluation forms. This enables:
Categorizing the type of issue handled
Selecting from predefined quality levels
Classifying interaction outcomes into specific buckets
Multiple choice provides structured data that is easy to aggregate and trend over time.
Date questions
Date-type questions allow evaluators to record specific dates within evaluations. This is useful for:
Tracking when issues were first reported
Recording promised follow-up dates
Noting deadlines communicated to customers
Date fields enable time-based analysis of evaluation data that was previously captured inconsistently in free-text fields.
Automated follow up evaluations
A powerful new workflow allows evaluation results to trigger additional evaluations automatically.
When an initial evaluation reveals a concern — for example, a low score on compliance or accuracy — the system can automatically trigger a follow-up evaluation on subsequent interactions by the same agent. This creates a continuous improvement loop:
Initial evaluation identifies a problem
Follow-up evaluations are automatically scheduled
Subsequent interactions are assessed for improvement
The cycle continues until performance meets the threshold
This removes the manual step of remembering to re-check agents who received poor scores. The system handles the follow-through.
Evaluating AI agent interactions
Evaluations are no longer limited to human agent interactions. Automated evaluations can now assess AI agent self-service interactions, with aggregated insights across all bot conversations.
How it works
The system automatically evaluates AI agent interactions against quality criteria you define:
Did the agent resolve the customer's request?
Did it stay within compliance guidelines?
Was the tone appropriate?
Did it escalate appropriately when needed?
Results are aggregated into dashboards showing AI agent quality trends over time.
Difference from human agent evaluations
Human agent evaluations often focus on soft skills, adherence to scripts, and coaching opportunities. AI agent evaluations focus on:
Configuration effectiveness (is the agent set up correctly?)
Knowledge base adequacy (does it have the information it needs?)
Guardrail performance (is it staying within boundaries?)
Escalation accuracy (is it handing off at the right moments?)
The corrective actions differ: for humans, you coach. For AI agents, you adjust configuration, expand knowledge bases, or refine guardrails.
RECALL & APPLY
Practice what you learned
Try answering in your own words, then compare with the suggested answer. These activities are for self-study and are not automatically graded.
01What changes when review expands beyond a small contact sample?
Topic: AI support for supervisors
Reveal suggested answer
Supervisors can examine wider patterns, but still need validated criteria, appropriate interpretation, and targeted human review.
02How should a supervisor respond to a recurring complaint pattern?
Topic: Proactive quality monitoring
Reveal suggested answer
Review representative contacts, verify the pattern, and determine whether the cause is a product issue, policy problem, or coaching need.
03Which view would help investigate why a customer became frustrated?
Topic: Conversational analytics views
Reveal suggested answer
Use the sentiment timeline to locate the change, then inspect the corresponding transcript and recording where available.
04Is enabling the instance capability the only configuration step?
Topic: Analytics modes and enablement
Reveal suggested answer
No. Review the applicable flow and analytics settings and the permissions needed for users to access the results.
05How does an issue category differ from a sentiment value?
Topic: Sentiment, issue detection, and summaries
Reveal suggested answer
A category describes what the contact concerns. Sentiment estimates emotional tone. They answer different questions and can be used together.
06Does automated redaction prove that every sensitive value was removed?
Topic: Compliance monitoring and redaction
Reveal suggested answer
No. The source explicitly notes that predictive systems can miss information. Review output against the organization’s requirements.
07Why is a case summary useful for a repeated billing complaint?
Topic: Analytics across channels and AI interactions
Reveal suggested answer
It combines the important facts and actions from related contacts so the reviewer can see what remains unresolved.
08What makes an operational metric actionable?
Topic: Real time operational analytics
Reveal suggested answer
A change in the metric has an understood meaning, an accountable owner, and a practical response the team can take.
09What must be known before interpreting a service level percentage?
Topic: Dashboard views and service levels
Reveal suggested answer
The target time, included contacts, time period, and queue or channel scope. Percentages with different definitions are not directly comparable.
10How can an alert become a useful response instead of another notification?
Topic: Alerts and live intervention
Reveal suggested answer
Include the relevant context, route it to the right owner, and define what action is expected at that level of urgency.
11A product launch produces a sudden call surge. What should the supervisor inspect first?
Topic: Responding to contact volume changes
Reveal suggested answer
Confirm the affected queue, current staffing and adherence, and the reasons customers are contacting support before selecting a response.
12Why might a long agent activity duration need context?
Topic: Queue and agent performance examples
Reveal suggested answer
A complex contact may legitimately take longer. Review the interaction and workload before treating duration alone as a performance problem.
13An agent is handling an unusually long call. What is an appropriate first step?
Topic: Monitoring and operational integrations
Reveal suggested answer
Review the contact context through the permitted monitoring workflow. Determine whether the agent needs assistance before intervening.
14How would you reduce alert fatigue?
Topic: Effective real time analytics practices
Reveal suggested answer
Review false or low value alerts, adjust thresholds, clarify priorities, and keep notifications tied to actions that teams can take.
15What should an instructor verify before demonstrating a preview capability?
Topic: AI assisted management and custom metrics
Reveal suggested answer
Confirm that the training environment has access and that its behavior and interface match the planned demonstration. Preserve the source’s preview qualification.
16Why can a lower hand off rate be a poor result?
Topic: AI agent monitoring
Reveal suggested answer
The agent may be retaining contacts it cannot safely or correctly resolve. Review success, quality, and appropriate escalation together.
17Faithfulness declines after a knowledge update. Where should the investigation begin?
Topic: AI agent investigation and traces
Reveal suggested answer
Compare affected interactions, retrieved content, and configuration versions. Look for missing, outdated, or irrelevant knowledge and unsupported answers.
18Why should an evaluation form use observable criteria?
Topic: Evaluation capabilities
Reveal suggested answer
Observable criteria support consistent scoring and make disagreements easier to investigate using contact evidence.
19How does corrective action differ for a human agent and an AI agent?
Topic: Evaluating AI agent interactions
Reveal suggested answer
Human follow up often involves coaching. AI follow up involves knowledge, instructions, tools, guardrails, and configuration testing.
Capture a key idea, an example, or a question for your instructor.
Imagine running a contact center where you have the right number of agents available at the right times. In this contact center, agent performance is evaluated automatically, and both customers and agents are happier. Welcome to the world of AI-powered workforce optimization in Amazon Connect Customer. Contact center workforce optimization involves strategically scheduling the right number of skilled agents to match staffing needs. It is about balancing excellent customer service with operational efficiency. In a typical contact center, workforce optimization involves several key components.
Forecasting
A forecast attempts to predict future contact volume and average handle time by predicting how many customer contacts you will receive and when they will come in.
Amazon Connect Customer uses historical metrics to generate the forecast.
Capacity planning
Capacity planning determines how many agents you will need to handle the forecasted contact volume while meeting your service level goals.
A capacity plan in Amazon Connect Customer helps you estimate the long-term full-time equivalent (FTE) requirements for your contact center up to 64 weeks in the future. It specifies how many FTE agents are required to meet your service level target for a certain period of time.
Scheduling
Contact center schedulers or managers need to create agent schedules for day-to-day workloads that are flexible and meet business and compliance requirements.
Amazon Connect Customer helps you create efficient schedules that are optimized for per-channel service level or average speed of answer (ASA) targets.
You can generate and manage agent schedules based on the following:
A published forecast
Shift profiles (templates for weekly shifts)
Staffing groups (agents that can handle specific types of contacts from a specific forecast group)
Human resources and business rules
Time off management
Contact center leaders coordinate team members' time off requests and manage schedule activities due to operational changes.
With Amazon Connect Customer time off management capabilities, you can manage agent time off requests that comply with preconfigured regional labor and business rules.
Administrators or managers with the appropriate security profile permissions can configure the time off settings. Amazon Connect Customer automatically approves or rejects requests depending on how you've configured both the time off rules and the daily maximum allowed time off hours.
Supervisors or managers with permissions can view agent time off requests and override automatically approved or rejected time off.
Overtime and voluntary time off management
Effectively balancing agent supply with customer demand is essential for meeting service goals while controlling costs. When you properly match the number of agents with incoming contacts, you create an environment where service levels and response times meet targets without unnecessary expenses. This forms the foundation of efficient workforce management.
Overtime (OT) and voluntary time off (VTO) provide the flexibility needed to adapt to changing conditions. When contact volumes spike or agent availability drops, you can use overtime to extend coverage without hiring new staff. In contrast, during slower periods, willing agents can use VTO to take unpaid time off, which reduces costs while maintaining appropriate staffing levels. Together, these capabilities create a responsive workforce model that can quickly adjust to business needs while keeping both customers and financial objectives in focus.
Schedule adherence
Contact center supervisors or managers track schedule adherence to understand when agents are following the schedule that you have created.
This helps ensure you achieve your service level targets, while improving agent productivity and customer satisfaction.
Intraday performance monitoring
Managers and team leaders can observe daily patterns and workload levels to adjust staffing schedules as needed. They can make predictions for rest-of-day contact volumes, average queue answer time, average handle time, and effective staffing.
The Amazon Connect Customer Intraday Forecast Performance Dashboard provides forecasts for the following:
Contact volume and average handle time for queues that have a minimum of 5,000 unique contacts per week per queue-channel for the last 4 weeks. This threshold is evaluated on a rolling basis, with intraday forecasts refreshed every 15 minutes throughout the day covering future intervals of the current day.
Average queue answer time follows the same threshold of 5,000 unique contacts per week per queue-channel over the trailing 4-week period.
Quality management
AI-powered quality management ensures consistent service delivery through intelligent agent evaluation and automated insights. Amazon Connect Customer agent evaluations use machine learning (ML) and generative AI to enhance quality assessments and provide data-driven coaching recommendations.
Key Amazon Connect Customer agent evaluations features include the following:
Generative AI-powered evaluations that automatically answer evaluation questions with rich context and detailed reasoning
Flexible automation options including manual assessment, rule-based automation, and AI-powered automation based on business needs
Multi-evaluator calibration capabilities that enable simultaneous evaluation of contacts to reduce bias and improve accuracy
Custom evaluation forms with ML-powered Conversational Analytics and intelligent contact selection
Coaching insights that provide targeted recommendations for agent development
Native integration built directly into Amazon Connect Customer without requiring external systems
AI-powered quality evaluations help maintain service consistency, ensure compliance through automated monitoring, and support agent development through intelligent feedback and generative AI-driven coaching recommendations.
Workforce optimization cycle
Amazon Connect workforce management components work together as services that consume shared forecast data, with forecasting as the required foundation. These six core components help optimize your contact center experience. The components include forecasting, capacity planning, scheduling, schedule adherence, agent flexibility, and intraday operations. Forecasting publishes predictions that the other five components consume independently—capacity planning and scheduling, for example, each use the published forecast without depending on each other, creating a comprehensive optimization system.
Forecasting
Analyze and predict contact volume and handle time based on historical data. Forecasts provide the foundational data that other components consume. Each downstream component uses published forecasts without depending on the others.
Capacity planning
Determine how many agents you need to handle forecasted contact volumes while meeting service level goals, up to 64 weeks in the future.
Scheduling
Generate agent schedules for day-to-day workloads that are flexible and meet business and compliance requirements.
Schedule adherence
Track whether agents are following published schedules. This helps ensure you achieve your service level targets while improving agent productivity
Agent Flexibility
Manage overtime, voluntary time off, shift exchanges, and time off requests to balance agent supply with customer demand.
Intraday operations
Monitor daily patterns and workload levels to adjust staffing schedules as needed, with forecasts refreshed every 15 minutes throughout the day.
Traditional Workforce Management Challenges
Common challenges
With traditional workforce management, you are relying on various manual processes, communication from various teams, and inaccurate or out of date forecasting and data.
Common challenges include the following:
Inaccurate forecasts leading to overstaffing or understaffing.
Time-consuming manual scheduling processes that can't quickly adapt.
Inconsistent quality evaluation dependent on individual reviewer preferences.
Limited visibility into true agent performance and productivity.
Difficulty balancing business needs with agent preferences.
How AI Enhances Workforce Management
ML powered forecasting
ML transforms workforce forecasting with powerful capabilities and insights that were previously difficult and time consuming. ML-powered forecasting in Amazon Connect Customer uses advanced machine learning to identify complex patterns in your contact data to predict contact volumes and their arrival rates.
By predicting contact volumes and arrival rates, Amazon Connect Customer improves the accuracy and efficiency of forecasts and schedules. Amazon Connect Customer generates these forecasts using an ML model specifically tailored for contact center operations.
It accounts for the following two key factors:
Seasonal trends across multiple years
Recent arrival patterns
Figure 54 ML powered forecastingSelect image to enlarge
The result? Short-term forecasts are computed daily and long-term forecasts weekly. Once you publish a forecast, it drives your scheduling and capacity planning with greater accuracy than traditional methods.
Modern workforce optimization is not about separate tools and processes. Instead, it focuses on creating a seamless system where each component supports the others.
Integrated approach to workforce optimization
Imagine you are the manager at a banking contact center. Your forecast shows an expected increase in calls about a new mortgage product. That information automatically flows into your capacity planning tool, which calculates the needed staffing. The scheduling system then creates efficient schedules, and the quality management system knows to focus evaluations on mortgage-related calls during that period.
Figure 55 Integrated approach to workforce optimizationSelect image to enlarge
Forecasts created
Historical contact data is used to create forecasts that flow into capacity planning.
Staffing requirements
The scheduling system then creates optimal schedules.
Agent schedules
This creates efficient agent schedules based on the configured staff rules and shift patterns.
Targeted evaluations
Contact evaluations can be targeted toward mortgage-related calls, ensuring agents are adhering to company guidelines and policies.
Introduction to ML Driven Forecasting and Capacity Planning
How ML Forecasting Works and Benefits
How ML forecasting works
ML-driven forecasting uses machine learning models tailored for contact center operations to analyze historical contact center data and predict future contact volumes and handle times. Unlike traditional forecasting methods, ML models can identify complex patterns, seasonal trends, and subtle relationships between variables that human forecasters might miss. For example, imagine you are running a retail customer service center. Your ML forecasting system analyzes two years of data and notices patterns of call volume increases outside of obvious holiday periods.
Figure 56 How ML forecasting worksSelect image to enlarge
Data collection
The system gathers historical information including:
Contact volumes
Handling times
Abandon rates
Special events
2: Data preprocessing
AI cleans the data, removing anomalies that could skew predictions.
Pattern recognition
ML algorithms identify recurring patterns and correlations.
Model training
The system creates and refines mathematical models based on the historical data.
Forecasting
The trained models generate predictions for future time periods.
Continuous learning
As new data comes in, the system updates its models to improve accuracy.
Amazon Connect automatically updates short-term forecasts daily and long-term forecasts weekly to provide fresh forecasts based on current information.
Forecasting Types
Contact centers typically use three types of forecasts.
Short term forecasts daily/weekly
Short-term forecasts (daily/weekly) Short-term forecasts predict contact volumes for the immediate future, typically up to 18 weeks ahead. These forecasts are crucial for day-to-day staffing decisions and scheduling. Think of this like checking the weather app before getting dressed in the morning. You need accurate information to make immediate decisions.
Figure 57 Forecasting TypesSelect image to enlarge
Long term forecasts monthly/quarterly/yearly
Long-term forecasts (monthly/quarterly/yearly) Long-term forecasts look further ahead (up to 64 weeks) to help with strategic decisions around hiring, training, and infrastructure planning. This is more like studying climate patterns before deciding where to build a house. You are making bigger decisions that cannot be easily changed.
Figure 58 Forecasting TypesSelect image to enlarge
Intraday forecasts real time
Intraday forecasts (real time) Intraday forecasts are ultra-short-term forecasts that help managers make immediate staffing adjustments throughout the day and update every 15 minutes. Imagine having a weather app that updates every 15 minutes during a storm. That is how intraday forecasting helps contact center supervisors navigate unexpected surges in call volume.
Figure 59 Forecasting TypesSelect image to enlarge
Benefits of ML forecasting
Using ML for contact center forecasting delivers several key advantages as follows:
Improved accuracy – ML models typically reduce forecast errors compared to traditional methods. This translates directly to better staffing decisions.
Automated updates – The system automatically refreshes forecasts (daily for short-term, weekly for long-term), which eliminates manual work and ensures forecasts stay current.
Pattern detection – ML identifies complex relationships that humans might miss, such as how weather conditions in specific regions affect certain types of support calls.
Multi-channel integration – Modern ML forecasting handles forecasts across all customer contact channels (phone, chat, email, social media) allowing for unified planning.
Common forecasting mistakes to avoid
Even with ML assistance, contact centers still make the following common forecasting errors:
Ignoring outliers
While days with unusual spikes or drops in activity should be cleaned from regular training data, they should still be analyzed for what caused the spike or drop.
Over reliance on technology
ML provides recommendations, but human judgment remains valuable for interpreting results and applying business context.
Using insufficient historical data
For best results, provide at least 12 months of historical data to capture seasonal patterns. Amazon Connect Customer models can use a maximum of 156 weeks (approximately 3 years) of data. The minimum requirement is 1,000 contacts per month within the last 6 months.
Not adjusting forknown future events
Always supplement ML forecasts with information about upcoming promotions, product launches, or other events that may affect contact volumes.
Forecasting is where all workforce planning begins. Without accurate forecasts, your scheduling and capacity planning will always be flawed, no matter how sophisticated your other processes are.
Capacity Planning Components and Process
Capacity planning overview
Capacity planning is the process of determining how many agents your contact center needs to handle forecasted contact volumes while achieving your service level targets. It bridges the gap between your forecasts (what you expect to happen) and your scheduling (who works when). Think of capacity planning like determining how many checkout lanes to open at a grocery store. Open too few, and customers wait in long lines. Open too many, and you are paying cashiers to stand around. The goal is to find the right balance between customer wait times and resource utilization.
Successful capacity planning requires the following three key components:
Accurate forecasts
Before you can create a capacity plan, you need reliable forecasts of contact volume and handle time. You must publish long-term forecasts before generating capacity plans.
Planning scenarios
Scenarios let you model different business conditions to see how they affect staffing needs. A scenario typically includes the following:
Maximum occupancy rate – Percentage of time agents spend actively handling contacts
Shrinkage factors – Breaks, training, absences, and after-contact work
Daily attrition rate – Model staff turnover rates
FTE hours per week – Working hours per week for full-time employees
Outsourced contacts – Model outsourcing percentages to third parties
Business operation days – Days per week or month that the contact center operates (for example, 5 days per week, 7 days per week, or specific calendar days)
Maximum OT – Amount of OT allowed
Maximum VTO – Amount of VTO allowed
Service level goals
You must define what success looks like using one of the following two metrics:
Service level – Percentage of contacts answered within a target time (for example, 80 percent in 20 seconds)
Average speed of answer (ASA) – The average time customers wait before speaking with an agent (for example, 30 seconds)
Figure 60 Service level goalsSelect image to enlarge
Capacity planning process
The capacity planning process includes the following steps.
Prepare your inputs
Before starting, you need:
Published long-term forecasts for the relevant time period
At least one planning scenario defining your operational parameters
Clear service level or ASA targets
Generate the capacity plan
Using the forecasts and scenarios, the capacity planning system calculates:
Required FTE employees with and without shrinkage
Forecasted occupancy rates
Gaps between available and required FTEs
Maximum OT and VTO allowances
Analyze the results
Review your capacity plan to identify:
Periods of understaffing where you will need to hire or use OT
Periods of overstaffing where you might offer VTO
Patterns that suggest you need to adjust your resourcing strategy
Create action plans
Based on your analysis, develop plans to address any gaps:
Recruiting and hiring timelines
Training schedules for new hires
OT policies for covering short-term gaps
Cross-training opportunities to increase flexibility
Historical Data Utilization
Using Historical Data and ML Forecasts
Did you know that your contact center is a repository of valuable data? Every call, chat, and email contains clues about future customer behavior. You can use the power of your historical contact data combined with ML forecasts to make smarter business decisions. This is about extracting practical insights that help you deliver better customer experiences while controlling costs.
Historical data is the foundation of all contact center planning. ML forecasts of historical data provide the following:
Baseline understanding of typical patterns
Trend identification for growing or declining contact types
Seasonal pattern recognition
Think of historical data as your contact center's memory. Just as you remember that your coffee shop is busiest on Saturday mornings, your historical data remembers when your contact center is busy or slow.
The following image shows an example of an ML forecast based on historical data in Amazon Connect Customer.
Figure 61 Using Historical Data and ML ForecastsSelect image to enlarge
Key historical metrics for forecasting
When using historical data for ML-powered forecasting, focus on the following essential metrics:
Volume metrics
Contact counts by channel
Interval distribution (daily, 15-minute and 30-minute breakdowns)
Day-of-week patterns
Month-of-year patterns
Handling metrics
Average handle time
Amazon Connect Customer automatically generates forecasts, with the ability for manual adjustments when needed through override capabilities.
While ML excels at finding patterns, human judgment remains crucial for contextual understanding, interpretation of unusual results, strategic decision making, and stakeholder communication.
Introduction to Intelligent Scheduling
AI Optimized Scheduling Process
AI scheduling overview
To run a contact center, you need the right number of agents working at the right times to achieve your operational goals.
Scheduling in Amazon Connect Customer helps with the following:
Generate agent schedules for day-to-day workloads that are flexible and meet business and compliance requirements.
Ensure you have the right number of agents to avoid both overstaffing (which leads to overspending) and understaffing (which impacts service levels).
AI scheduling starts with forecasting to predict how many contacts will arrive at different times. Contact centers used to rely on educated guesses to create schedules. Today, AI-optimized scheduling transforms this process. You can use Amazon Connect Customer to create AI optimized schedules that are per-channel service level or average speed of answer (ASA) targets.
The following image shows an example of an Amazon Connect Customer forecast graph with predicted call volumes by hour and day with peak periods highlighted.
Figure 62 AI scheduling overviewSelect image to enlarge
AI trend detection
Unlike humans who might miss subtle patterns, AI can detect trends across factors like day-of-week patterns, time-of-day variations, seasonal trends, and historical anomalies.
Calculating staffing needs
After the AI system has predicted contact volume, it calculates how many agents are needed for each 15-minute or 30-minute period considering the following:
How long it typically takes to handle each contact
The target service level
Required break times
Staff rules
Figure 63 AI scheduling overviewSelect image to enlarge
Scheduling optimization
The system then creates shift activities that put the right number of agents in the right places, trying several combinations to find the optimal solution.
Figure 64 AI scheduling overviewSelect image to enlarge
Improved accuracy
AI scheduling addresses the following common problems:
Overstaffing – AI optimizes staffing levels to predicted demand, ensuring you're not paying for idle time.
Understaffing – By accurately predicting busy periods, AI ensures adequate coverage during peak times.
Inefficient shift patterns – AI creates shifts that precisely match your contact patterns to maximize efficiency.
Example of the AI optimized schedule
Real-world example: For an online retailer, the AI system might notice that Monday mornings have high call volumes about weekend orders. Additionally, Thursday evenings see increased chat requests about weekend deliveries. The system would schedule more phone agents on Monday mornings and more chat agents on Thursday evenings.
Balancing Agent Preferences and Contact Center Needs
Business and agent needs
Contact center scheduling has traditionally been employer-centric and often leads to high turnover, lower productivity, and increased absenteeism.
Contact center agents and businesses face common needs.
Business needs
The following core operational metrics drive sustainable contact center performance:
Sufficient coverage for all expected contacts
Meeting service levels
Cost efficiency
Skill coverage
Agent needs
The following critical workplace flexibility factors support a stable and engaged workforce:
Consistent schedules
Work-life balance
Preferred shift times
Time off for personal events
Consecutive days off
Balancing solutions
Amazon Connect Customer improves your contact center's approach with scheduling optimization as follows:
Staff rules for agent-level customization – Before creating schedules, contact centers set agent preferences for shift times, preferred days off, maximum consecutive workdays, and specific time-off needs.
Shift exchange – Agents can trade shifts with each other, provided the exchange doesn't create coverage issues.
Flexible time off options – Progressive contact centers offer voluntary time off (VTO) during slow periods, advanced time-off requests, and optional overtime during busy periods.
Staff Rules Configuration
Staff contract rules
You can use Amazon Connect Customer to set customized rules at the agent level. Before creating schedules, contact centers set agent preferences for shift times, preferred days off, maximum consecutive workdays, and specific time-off needs.
Min working time
Minimum working hours/minutes per day/week
Max working time
Maximum working hours/minutes per day/week
Min consecutive working days
Minimum consecutive working days
Max consecutive working days
Maximum consecutive working days
Min time gap
Minimum time gap between shifts (hours)
Min consecutive days off
Minimum consecutive days off
Flexibility with scheduling adjustments
There is flexibility in scheduling adjustments so that contact centers can take a progressive approach when balancing business and agent needs.
Shift exchange options
Flexible time off options
Agents can trade shifts with each other, provided the exchange doesn't create coverage issues.
Real world examples of staff rules configuration
For the holiday season, a retail contact center might do the following:
Ask for volunteers to work key dates with overtime incentives.
Create shorter shifts on major holidays.
Plan far in advance so agents know their holiday schedule.
For agents who are parents, a contact center might do the following:
Create dedicated parent shifts during school hours.
Allow part-time options during school terms.
Create flexible scheduling during school breaks.
Flexibility with scheduling adjustments
Shift exchange options
Agents can trade shifts with each other, provided the exchange doesn't create coverage issues.
Flexible time off options
Progressive contact centers offer the following:
Voluntary time off (VTO) during slow periods
Advanced time-off requests
Optional overtime during busy periods
Real world examples of staff rules configuration
For the holiday season, a retail contact center might do the following:
Ask for volunteers to work key dates with overtime incentives.
Create shorter shifts on major holidays.
Plan far in advance so agents know their holiday schedule.
For agents who are parents, a contact center might do the following:
Create dedicated parent shifts during school hours.
Allow part-time options during school terms.
Create flexible scheduling during school breaks.
Intraday Forecast Performance Dashboard
Dashboard overview
The Intraday forecast performance dashboard is a real-time monitoring tool in Amazon Connect Customer that helps contact center managers optimize daily operations. It tracks and forecasts key metrics including contact volume, handle time, answer speed, and staffing levels over 24-hour periods in 15-minute intervals.
The dashboard combines current data with historical patterns and short-term forecasts so managers can anticipate operational needs. This helps them adjust staffing levels and respond to emerging trends. Its color-coded indicators and comparison features help you quickly assess performance against benchmarks, which leads to immediate, data-driven operational decisions.
The Amazon Connect Customer Intraday Forecast Performance Dashboard provides forecasts for the following:
Contact volume and average handle time for queues that have a minimum of 5,000 unique contacts per week per queue-channel for the last 4 weeks. This threshold is evaluated on a rolling basis, with intraday forecasts refreshed every 15 minutes throughout the day covering future intervals of the current day.
Average queue answer time follows the same threshold of 5,000 unique contacts per week per queue-channel over the trailing 4-week period.
The following image shows the Intraday dashboard in Amazon Connect Customer.
Figure 65 Intraday Forecast Performance DashboardSelect image to enlarge
Performance overview chart
The Intraday trailing performance overview chart provides aggregated metrics based on your filters. Each metric in the chart is compared to your compare to benchmark time range filter.
The following image shows an example Intraday trailing performance overview chart.
Figure 66 Intraday Forecast Performance DashboardSelect image to enlarge
Image features
This image shows the following:
Contact volume during your time range selection was 1,213.
This is down ~13 percent compared to your benchmark number of contacts handled.
Avg. handle time has increased by 3.92 percent.
Avg. speed of answer has increased by 41 seconds.
As average speed of answer increases, service quality decreases. This metric inversely relates to contact center performance, with longer wait times indicating less efficient customer experience.
Abandonment rate shows the current day performance and is not compared to the previous performance.
The colors that appear for the metrics indicate negative (red) compared to the benchmark.
Comparison trend graphs
The Intraday performance dashboard displays the following trend graphs, which cover different metrics:
Contact volume
Average handle time
Average speed of answer
Effective staffing
These graphs include the Intraday forecast that projects up to 24 hours on a 15-minute interval based on the following:
The value of the respective metric
The historical actuals from the current day
The historical actuals from the same time in the past week
These trend graphs provide data only for the next 24 hours and the past 24 hours. There is no option to change the time range.
The following image shows an example of a Contact volume trend graph from the Intraday dashboard.
Figure 67 Intraday Forecast Performance DashboardSelect image to enlarge
Comparison against short term forecasts
You can compare Average handle time and Contact volume against published short-term forecasts.
To select this option, choose the Compare to button and select Short-term published forecast. This automatically picks up the published short-term forecast for the time range selected. You can't select an unpublished forecast or a specific published forecast.
For historical widgets, it compares against the same time range as the widget. For the daily projection widget, it compares against the entire day.
Figure 68 Intraday Forecast Performance DashboardSelect image to enlarge
Daily projection chart
The daily projection chart provides a projection of how the day will end by combining historical metrics for the day so far with Intraday forecasts for the remainder of the day. This is available for the following metrics:
Average handle time
Average queue answer time
Contact volume
Effective staffing
This widget only supports comparing against short-term forecasts for contact volume and average handle time.
Figure 69 Intraday Forecast Performance DashboardSelect image to enlarge
Real Time Schedule Adjustments
Why Adjustments Are Needed
Even with the best AI and forecasting, schedules need adjustment because of the following:
Agents may be absent (illness, family emergencies, transportation problems)
Contact volumes might differ from forecasts (marketing campaigns perform differently than expected)
Handle times might change (new issues might take longer to resolve)
Business priorities might shift suddenly
Real-time schedule adjustment is the process of identifying these mismatches between planned and actual conditions, then making quick, intelligent changes to minimize their impact.
Identifying the mismatch
A supervisor can identify mismatches between planned and actual conditions in Amazon Connect Customer using real-time adherence monitoring and intraday forecasting tools. The real-time metrics dashboard refreshes every 15 seconds, giving supervisors a live view of agent states compared to their published schedules. When an agent deviates from their planned activity, Schedule Adherence Notification Rules automatically alert the supervisor via email, task, or third-party integrations. Configurable adherence thresholds (1–10 minutes per activity) help distinguish genuine non-adherence from minor operational variances, reducing alert fatigue. On the demand side, intraday forecasts operate at 15- or 30-minute intervals, allowing supervisors to compare forecasted contact volume against actual arrival patterns and spot emerging staffing gaps before they impact service levels.
Mitigating the mismatch
Once a mismatch is identified, supervisors have several tools to respond quickly. Schedules can be edited in real time with minute-level precision. Agents with overlapping shifts can swap schedules, and when agents need to be reassigned, supervisors can move them to the correct staff group and regenerate the schedule for that individual without disrupting the rest of the team's optimized schedule. For non-phone activities such as coaching, training, or meetings, the Optimize Activity Placement feature automatically places them at algorithmically optimal times within shifts to minimize service level impact. When forecasts change, supervisors must manually trigger a schedule refresh because published schedules do not auto-update.
Types of adjustments
Understanding what happened and why can help to refine staffing strategies for similar situations. Review the following types of adjustments.
Schedule extensions
When contact volumes exceed forecasts or agent availability falls below required levels, implementing schedule extensions provides essential coverage to maintain service levels.
Asking agents to stay later (overtime)
Bringing agents in earlier
Converting scheduled training to work time
Schedule reductions
During periods of unexpected low contact volume, strategically reducing scheduled agent hours helps align staffing with actual business needs while managing labor costs.
Offering voluntary time off (VTO) during slow periods
Sending agents to training during slow times
Allowing longer breaks during slow periods
Skill adjustments
Flexible modification of agent skill assignments allows for dynamic reallocation of workforce resources to match changing contact patterns and business priorities.
Temporarily reassigning agents to different channels
Enabling backup skills
Moving agents between departments
When deciding how to adjust schedules, consider the following:
Impact on service levels
Impact on agent well-being
Cost implications
Fairness and process
Duration of the need
Insider tip: Create an opportunity to improve future scheduling by documenting each change and the outcome.
Intelligent Scheduling Enhancements
Multi Skill Agent Scheduling
The AI-optimized scheduling process you learned about has received several enhancements that give supervisors more control and produce better results. Previously, scheduling optimization treated agents as interchangeable. Now it accounts for the fact that agents have different specialized skills, and customer needs vary by skill type.
Introduced in November 2025, multi-skill scheduling optimizes based on each agent's specific skill set. If your contact center handles billing, technical support, and sales, the scheduler now ensures the right mix of skills is available at every time slot, not just the right headcount.
How it works
The scheduling algorithm considers:
Each agent's certified skills (billing, tech support, sales, Spanish-speaking, etc.)
Forecasted demand per skill
Service level targets per queue or skill group
The result is a schedule where you have enough billing-skilled agents during billing peaks and enough tech support agents during the hours technical issues spike — even if total headcount remains the same.
Why it matters
Without multi-skill scheduling, a contact center might be fully staffed but still miss service levels because the wrong skills are available. You could have 20 agents on shift but only 2 who can handle Spanish-language billing inquiries during a Spanish-language peak.
Multi-skill scheduling eliminates this mismatch by treating skills as a dimension of the optimization, not an afterthought.
Schedule adherence thresholds and notifications
Supervisors can now define how early or late agents can start or end shifts and activities, with automated notifications when thresholds are exceeded.
Introduced in October 2025, this feature adds proactive compliance monitoring to scheduling:
Define thresholds — set acceptable ranges for how early or late agents begin and end shifts and scheduled activities (breaks, training, lunch)
Automated notifications — when agents exceed these thresholds, supervisors receive email or text alerts automatically
No manual checking required — the system monitors adherence continuously and flags exceptions
This shifts schedule adherence from a reactive review (checking yesterday's data) to a real-time compliance tool. Supervisors know immediately when patterns emerge, allowing coaching conversations while the behavior is fresh.
Individual scheduling and bulk operations
Two operational improvements simplify day-to-day scheduling management for larger teams.
Individual agent scheduling
Managers can now schedule individual agents and merge their schedule with the existing team schedule. This is useful for:
New hires who join mid-cycle and need schedules created outside the normal generation run
Agents returning from leave who need reintegration into the rotation
Special assignments or temporary schedule adjustments for specific individuals
Bulk operations
For larger changes, managers can now:
Copy existing schedule configurations as templates for new schedules
Bulk edit scheduling settings across multiple agents simultaneously
Apply rule changes to groups rather than updating agents one by one
These efficiency improvements reduce the administrative time supervisors spend on scheduling mechanics, freeing them to focus on optimization and coaching.
Introduction to AI Powered Agent Performance Evaluation
Automated Quality Management
AI powered quality management
Traditional quality management approaches limit contact centers to evaluating a small sample of interactions, typically 1–3 percent, leading to inconsistent evaluations and delayed coaching opportunities.
Amazon Connect Customer Conversational Analytics (formerly known as Contact Lens) transforms quality management through generative AI-powered automation as follows:
Use AI to automatically evaluate 100 percent of customer interactions.
Generate performance insights from conversations as they happen.
Consistently apply standardized and objective evaluation criteria.
Flag to supervisors for immediate attention.
Extract performance trends and insights across agent population.
Statistical analysis provides valid insights into trends.
AI-powered analytics in Amazon Connect Customer go deeper by analyzing what actually happens during interactions.
Amazon Connect Customer Conversational Analytics with generative AI can do the following:
Evaluate agent performance based on natural language criteria automatically.
Detect compliance issues or script adherence problems.
Identify successful techniques used by top performers.
Provide objective, consistent evaluations.
Deliver feedback much faster than manual processes.
Instead of a supervisor reviewing 3–5 calls per agent per month, the AI system can evaluate every single interaction. This ensures no problematic interactions slip through while identifying top performers consistently.
Effective Evaluation Forms and Best Practices
Designing effective evaluation forms
Evaluation forms are important for measuring the quality and effectiveness of your contact center. You can use AI to help with this in many ways, but it's important to implement best practices when creating evaluation forms.
To design evaluation forms that work effectively with AI automation, use the following best practices:
Be specific
Use specific, observable behaviors rather than vague criteria. For example, use: "Did the agent address the customer by name at least once during the call?", rather than: "Was the agent professional?"
Use binary criteria when possible
Questions with clear yes/no answers are easier for AI to evaluate accurately.
Include verification points
For compliance items, specify exactly what the agent should say.
Balance process and outcome
Include both process measures (did the agent follow steps?) and outcome measures (was the issue resolved?)
Test and refine
Start with a small set of criteria, validate AI evaluations against human evaluations, and refine as needed.
The following image shows the evaluation form builder in Amazon Connect Customer.
Figure 70 Automated Quality ManagementSelect image to enlarge
Evaluation Forms in Action
Implementing effective evaluation forms and processes creates a foundation for quality improvement across your contact center. The most effective forms combine AI automation with human insight to measure what matters most. Scores on evaluations should do more than grade agent performance. They should guide coaching conversations and highlight development opportunities.
Overview
Review all the key information about the evaluation form such as:
Evaluation score
Evaluation ID
Status
Timestamps
Timezone
History
Evaluation scores
See how this contact evaluation compared to the agent's trailing four-week average and the average of all contacts.
Weighting
See how each section and question is weighted within the evaluation from.
AI generated answer
Receive a generative AI-powered recommendation for the answer, along with context and justification (reference points from the transcript that were used to provide answers).
Did you notice how AI assists with objective scoring in the completed evaluation form example? This technology saves supervisors time while providing consistent feedback to agents. Through smart form design and regular evaluations, agents learn exactly what success looks like. Organizations benefit from better customer experiences and customers receive more consistent service. Your evaluation strategy becomes a powerful tool that benefits everyone involved in the customer service journey.
Performance Evaluation Enhancements
Generative AI email overviews
Beyond evaluating 100 percent of interactions with standardized criteria, automated quality management includes features that improve both the evaluation process and the actions taken based on results.
Agents handling email interactions now receive AI-generated overviews that help them respond faster and more accurately.
This feature provides agents with:
Email overview — a concise summary of the customer's email, highlighting the key issue and any relevant history
Suggested actions — recommended next steps based on the email content and organizational policies
Draft responses — AI-generated response suggestions the agent can review, edit, and send
This reduces the time agents spend reading lengthy email threads and deciding how to respond. The AI handles the analysis; the agent reviews and acts.
Introduction to Audio Enhancement
Two modes of audio enhancement
Audio Enhancement offers two specialized modes. Each targets a different type of audio interference on the agent's side of the call.
Voice isolation
Voice Isolation suppresses all noises and background speech, leaving only the agent's voice audible to the customer.
Best for environments where:
Multiple agents sit close together and cross-talk bleeds into calls
The agent works from home with family or housemates nearby
Confidential conversations from adjacent desks could be overheard by the customer
Voice Isolation is the more aggressive mode. It filters out everything that is not the primary speaker's voice.
Noise suppression
Noise Suppression suppresses only background noises — things like keyboard clicking, air conditioning, traffic, dog barking — while preserving human speech.
Best for environments where:
The primary issue is environmental noise rather than other voices
Agents occasionally need nearby colleagues to contribute to a call
The workspace has mechanical noise (printers, HVAC) but limited cross-talk
Noise Suppression is the lighter-touch mode. It cleans up the audio environment without silencing nearby speech.
Selecting the right mode
Choosing the right mode depends on the agent's physical environment. The following comparison helps.
Environment
Recommended mode
Why
Open-plan office, desks close together
Voice Isolation
Prevents cross-talk from adjacent agents
Work-from-home, shared living space
Voice Isolation
Filters household speech and activity
Private office with street noise
Noise Suppression
No cross-talk risk; just environmental noise
Contact center floor, good desk spacing
Noise Suppression
Background hum without significant voice bleed
Quiet home office with occasional pet noise
Noise Suppression
Intermittent noise, no ongoing speech interference
Configuration and control
Audio Enhancement is managed at two levels: administrators enable and configure it, and agents with proper permissions can adjust their own settings.
Administrator setup
Administrators enable Audio Enhancement through User Management settings in the Amazon Connect Customer console. They can:
Enable or disable Audio Enhancement for specific security profiles
Set the default mode (Voice Isolation or Noise Suppression)
Control whether agents can change their own settings
Agent self service
Agents with proper permissions can adjust Audio Enhancement settings from the CCP. Changes take effect on the agent's next call. They can:
Toggle Audio Enhancement on or off
Switch between Voice Isolation and Noise Suppression
This gives agents flexibility to adapt to changing conditions — for example, switching to Voice Isolation when a noisy meeting starts in a nearby conference room.
Impact on call quality metrics
Audio Enhancement improves several operational metrics:
Reduced "I can't hear you" complaints from customers
Lower repeat-request rates (customers asking agents to repeat themselves)
Improved transcription accuracy for Conversational Analytics
Better sentiment scores when audio quality is no longer a frustration factor
RECALL & APPLY
Practice what you learned
Try answering in your own words, then compare with the suggested answer. These activities are for self-study and are not automatically graded.
01How do forecasting and scheduling answer different questions?
Topic: Workforce optimization components
Reveal suggested answer
Forecasting estimates the workload. Scheduling assigns agents and activities to meet that workload within the applicable rules.
02What would you inspect before offering voluntary time off?
Topic: Agent flexibility and intraday operations
Reveal suggested answer
Check the remaining demand, available skills, staffing position, service targets, and applicable rules for the rest of the day.
03Must scheduling wait for a completed capacity plan?
Topic: The workforce planning relationship
Reveal suggested answer
The source explains that capacity planning and scheduling can each consume published forecasts independently. Distinguish the conceptual planning cycle from a required technical dependency.
04Why might a technically valid forecast still need adjustment?
Topic: ML forecasting and workforce challenges
Reveal suggested answer
It may not reflect a future event absent from historical data, such as a product launch or a planned service change.
05A new mortgage product is expected to increase calls. Which teams need the information?
Topic: An integrated workforce example
Reveal suggested answer
Forecasting, staffing and scheduling teams need the expected workload. Training and quality teams need the new contact types and required handling practices.
06Why should an unusual spike be investigated before it is removed?
Topic: How ML forecasting works
Reveal suggested answer
The spike may represent a data error or a real event likely to recur. Those cases require different treatment.
07Which forecast would support a hiring plan rather than today’s queue response?
Topic: Forecast horizons
Reveal suggested answer
A long term forecast supports hiring and capacity decisions. Intraday information supports immediate operational adjustment.
08What would you do if forecasts repeatedly miss a promotion period?
Topic: Forecast quality and limitations
Reveal suggested answer
Review the historical data and the promotion’s effect, then incorporate the known business event into the planning process and assess the revised result.
09Why would two plans based on the same demand forecast require different staffing?
Topic: Capacity planning inputs
Reveal suggested answer
Different service targets, occupancy assumptions, available hours, attrition, or other scenario inputs can change the staffing requirement.
10What makes a capacity plan useful to a manager?
Topic: Capacity planning process
Reveal suggested answer
It identifies the size and timing of staffing gaps and links them to concrete hiring, training, outsourcing, or scheduling decisions.
11Why can stable contact volume still create a staffing shortfall?
Topic: Historical data for forecasting
Reveal suggested answer
Handling time may rise, changing the workload even if the number of contacts is unchanged.
12Why is daily headcount alone insufficient for scheduling?
Topic: Intelligent scheduling
Reveal suggested answer
Demand changes across intervals and contact types. Coverage must match the required times and skills, including planned non-contact activities.
13How should an organization handle competing requests for the same time off?
Topic: Business needs and agent preferences
Reveal suggested answer
Apply the configured rules and the organization’s process consistently, considering coverage, fairness, and any applicable constraints.
14Why can a schedule fail to meet demand despite enough total agents?
Topic: Staff rules and scheduling constraints
Reveal suggested answer
The combination of skills, availability, shift patterns, and staff rules may prevent enough coverage in particular intervals.
15What should the instructor establish before interpreting the chart?
Topic: Intraday forecast performance
Reveal suggested answer
Identify the selected metric, queue or channel, timeframe, comparison, and which points are actuals versus forecasts.
16How would you investigate higher answer times?
Topic: Trend graphs and comparisons
Reveal suggested answer
Review volume, handle time, and effective staffing together. A change in any of them can affect waiting, so avoid assuming one cause from one metric.
17How does a daily projection differ from actual performance so far?
Topic: Published forecasts and daily projections
Reveal suggested answer
It combines observed results with forecast values for the remaining intervals. It is not a completed-day measurement.
18Does a changed forecast automatically update every published schedule?
Topic: Real time schedule adjustments
Reveal suggested answer
The source says supervisors must trigger the relevant refresh. Verify the published schedule and communicate any approved change.
19A team is fully staffed but lacks the required language skill. What is missing?
Topic: Multiskill scheduling and adherence
Reveal suggested answer
The staffing view needs a skill dimension. Total headcount does not show whether the right agents can handle the waiting work.
20Why compare automated and human evaluations during rollout?
Topic: Automated quality management
Reveal suggested answer
The comparison reveals ambiguous criteria, missing context, and scoring errors that need refinement before the results drive decisions.
21Improve the question “Was the agent good?”
Topic: Evaluation form design
Reveal suggested answer
Use a specific observable criterion, such as whether the agent confirmed the customer’s requested next step, with a clear scoring rule.
22What should an agent verify in a draft email response?
Topic: Email support and evaluation follow up
Reveal suggested answer
Check the actual issue, policy applicability, factual details, tone, and any promised action before sending the response.
23Which mode fits nearby conversations that should not reach the customer?
Topic: Audio enhancement modes
Reveal suggested answer
Voice isolation is the mode described for suppressing background speech. Test it with the actual headset and working environment.
24Why should a configuration change be tested on the next call?
Topic: Audio enhancement configuration
Reveal suggested answer
The source describes changes taking effect on the next call. Confirm the active mode and listen for both noise reduction and clear primary speech.
Capture a key idea, an example, or a question for your instructor.
Transform your customer experience with the AI-powered self-service capabilities of Amazon Connect Customer. This course explores how intelligent automation can reduce call volumes, enhance customer satisfaction, and lower operational costs.
You will learn about natural sounding voice interactions using our text-to-speech service, Amazon Polly, and how to design impactful conversational interfaces using Amazon Lex. You will also discover how to deploy generative AI solutions with Amazon Q in Connect to handle customer requests without agent intervention.
Through practical examples and use cases, you will learn to create seamless self-service experiences that satisfy customer needs while optimizing your contact center resources.
Introduction to AI Powered Self Service
Self Service in Modern Contact Centers
Modern AI self service
Imagine you run a retail contact center. During holiday seasons, your call volume triples, mostly with customers asking about order status. Without strong self-service, you would need to hire seasonal staff, which could increase costs and create training challenges. With the self-service capabilities of Amazon Connect Customer, you could automatically handle these common requests, leaving your agents available to tackle more complex issues.
Review the following to learn how modern AI self-service can help:
Traditional limitations – Self-service capabilities have evolved far beyond basic touch-tone menus (for example, "Press 1 for sales, 2 for support").
Modern capabilities – Today's AI-powered solutions can understand natural language, personalize responses, and complete actions on behalf of customers.
Business benefits – For businesses, this means reduced call volumes, lower operational costs, and more satisfied customers.
Self service or human agents
Self-service can handle the standard queries that your contact center receives daily, which leaves human agents free to handle specialized queries. This provides the following typical benefits:
Continuous 24-hour availability of information and assistance
Immediate access to answers without waiting in queues
Consistent experience regardless of time or day
Reduced call volume for routine inquiries and simple transactions
Lower operational costs per customer interaction
Improved agent satisfaction by focusing on complex, value-added tasks
Figure 71 Amazon Connect AI Self Service CapabilitiesSelect image to enlarge
Adding self-service options to call centers helps everyone. Customers get quick answers whenever they need them without waiting in queues for an agent. The contact center saves money by having self-service capabilities handle simple questions while skilled agents focus on specialized problems. The information collected from self-service also helps companies understand what customers want. As more people prefer solving problems quickly on their own, having good self-service options is necessary. Companies that use technology effectively while still providing human help when needed will benefit.
Evolution from Traditional IVR to AI Powered Self Service
Four Types of IVR
Self-service capabilities have evolved beyond basic touch-tone menus and directed dialog options. Today's AI-powered solutions can understand natural language, personalize responses, and complete actions on behalf of customers. For businesses, this can mean reduced call volumes, lower operational costs, and more satisfied customers.
The following sections explore the four types of Interactive Voice Response (IVR).
Traditional IVR
The following are features of the first generation of automated phone systems:
Limited to touch-tone keypads or directed dialog
Required customers to navigate menus by pressing buttons
Included examples such as "Press 1 for sales, 2 for support, 3 for billing"
Customer experience:
System: "Please listen carefully as our menu options have changed." Customer: Sighs and waits to press the right number
Automatic speech recognition ASR
These systems recognize spoken words but not context as described in the following:
Understands natural speech instead of requiring button presses
Matches spoken words to predefined commands
Capabilities limited to specific phrases and commands
Customer experience:
Customer: "I need to update my billing address." System: Recognizes "billing address" and routes accordingly
Natural language understanding NLU
These systems interpret meaning behind words as described in the following:
Identifies customer intent from conversational language
Understands context and variations in expression
Handles multiple intents in a single statement
Customer experience:
Customer: "My internet is really slow today." System: Identifies intent as "technical support – service issue" and routes to the appropriate department automatically
Intelligent action systems
These advanced systems understand and process decisions such as the following:
Processes complete transactions without human intervention
Handles complex security and verification processes
Performs multi-step processes while maintaining context
Customer experience:
Customer: "I'd like to make a payment." System: Verifies identity, checks account balance, processes payment, and sends confirmation
Customer: Completes entire transaction without agent interaction
Compare technologies
Feature
Traditional IVR
ASR
NLU
Intelligent actions
Input Method
Keypad buttons
Voice commands
Conversational language
Natural conversation
Understanding
None
Word recognition
Intent recognition
Context-aware understanding
Customer Effort
High
Medium
Low
Minimal
Resolution Speed
Slow
Medium
Fast
Fastest
Response Latency
Instant (pre-recorded)
Low
Low
Higher (LLM inference)
Business Value
Basic automation
Improved routing
Better experience
Complete automation
By using the intelligent actions modern IVR approach, Amazon Connect Customer empowers customers to speak naturally and get help quickly, instead of navigating through multiple menu layers.
When implementing self-service, companies may worry that customers will feel they are getting a reduced experience compared to talking with an agent. However, customers may prefer the speed and availability of self-service to efficient inquiry resolution.
Think about how you interact with modern voice assistants like Amazon Alexa. You speak naturally, and they understand your intent. Amazon Connect Customer brings this same technology to contact centers, which creates experiences that are natural and helpful rather than robotic and limiting.
Remember that good self-service should not be more complicated. It should be faster and easier than talking to an agent.
How Amazon Connect Customer AI Enhances Self Service Experience
Three Core Services
Amazon Connect Customer self-service capabilities are built on three underlying AWS services: Amazon Polly, Amazon Lex, and Amazon Q in Connect Customer. These three AWS services are pre-integrated with Amazon Connect Customer, so you can access and utilize them from within the Amazon Connect Customer admin workspace.
Amazon Polly
This service converts text to natural-sounding speech, which gives your self-service system a human-like voice. Gone are the days of robotic, monotone IVR systems.
These three services, along with Amazon Connect Customer, work together to create natural, powerful customer experiences.
Amazon Lex
Amazon Lex, a conversational AI service, powers NLU in your self-service solutions. It uses conversational AI technology similar to that powering Amazon Alexa, which makes it possible for customers to speak or type naturally instead of navigating rigid menus.
Traditional IVR systems followed rigid scripts with limited options. Discover how AI transforms this experience in the following comparison:
Traditional IVR: "For account balance, press 1. For payment information, press 2."
AI-powered self-service: "Hi there! How can I help you today?"
Customer: "I want to check my account balance."
AI-powered self-service: "I'd be happy to help you check your balance. For security, can you please verify the last four digits of your social security number?"
The difference is dramatic. Amazon Lex understands the customer's intent, which makes interactions feel more like talking to a helpful person than navigating a machine.
Amazon Q in Connect Customer
This generative AI assistant, the same assistant that helps agents, can be used for self-service. Providing dynamic responses, step-by-step guidance, and can complete actions on behalf of customers.
Self-service capabilities have evolved beyond basic touch-tone and directed dialog menus. Examples include "Press 1 for sales" and "Press or say 1 for sales, 2 for support."
Imagine a customer is contacting their bank in the following example:
Customer: "I noticed a strange charge on my account from yesterday."
Amazon Q in Connect Customer: "I understand you're concerned about an unfamiliar transaction. I can help you review recent charges and place a temporary hold on your card if needed. Would you like me to show you the most recent transactions on your account?"
This level of understanding and assistance was not possible with traditional self-service systems. Amazon Q in Connect Customer doesn't just follow scripts—it comprehends situations and responds appropriately.
Amazon Q in Connect Customer can do the following:
Answer open-ended customer questions using knowledge from your business.
Recommend personalized next steps based on:
What is understood of the customer's current situation
What is known about the customer in Amazon Connect Customer Profiles (this could include products owned, recent orders or contact history) or other data sources
Hand off to deterministic flows, or human agents seamlessly when needed, with full conversation context.
Introduction to Amazon Polly for Text to Speech
Understanding Amazon Polly
Amazon Polly overview
Amazon Polly is an AWS service that converts text into lifelike speech. Your contact center can use Amazon Polly, integrated with Amazon Connect, to deliver natural-sounding voice prompts to callers. Instead of recording every prompt, you can type what you want your IVR to say, and Amazon Polly says it to your customers.
Imagine you manage a bank's contact center. Traditionally, contact centers hired voice actors to record prompts for every account balance, transaction amount, and address. This made customer experiences rigid and changes costly. With Amazon Polly, your system can dynamically generate speech based on real-time data: "Hello Jane, your current balance is $523.45." This creates a personalized experience without extensive recording sessions.
Figure 72 Introduction to Amazon Polly for Text to SpeechSelect image to enlarge
Using Amazon Polly
Amazon Connect Customer uses the built-in, managed integration with Amazon Polly to provide you with the ability to dynamically create speech-based prompts through text. Amazon Polly Neural and Standard voices are included at no additional charge within Amazon Connect Customer.
To learn more about using Amazon Polly in Amazon Connect Customer, review the following steps.
Step 1
In the Amazon Connect Flow designer, select the Play Prompt block from the Interact section of the Block Library.
Add this block to your canvas by dragging it into position or using keyboard navigation.
This block enables text-to-speech capabilities through Amazon Polly.
Step 2
Click on the play prompt block to open the block settings.
Choose the 'Text-to-speech or chat text' option. A text field will appear.
Select the Interpret as option and select Speech Synthesis Markup Language (SSML) to allow you to add SSML tags to modify the speech output.
Add your message such as 'Welcome to customer service. How may I help you today?'
Then click Save.
Text to speech best practices
Text-to-speech (TTS) allows you to instantly update IVR messages, offer personalized self-service, and support multiple languages without the cost of re-recording audio files. However, poorly implemented TTS can sound robotic and damage customer trust. To ensure your prompts sound natural and professional, apply the following best practices.
Script for the ear
When writing TTS prompts, remember that your customer is listening, not reading:
Keep sentences short – Break down complex information into short, conversational sentences to avoid cognitive overload.
Put actions first – Start instructions with the required action. For example, use "Say billing or press 2" instead of "To hear your billing information, please say 2."
Avoid jargon – Use simple, everyday language that the average caller easily understands.
Use punctuation for natural pauses – Insert commas and periods into the text to force the TTS engine to pause appropriately.
Control pronunciation and pacing with SSML
TTS engines can struggle with acronyms, phone numbers, or custom product names. Use SSML to refine the output:
Control pronunciation – Use SSML tags to specify phonetic pronunciations for specialized terms or to spell out characters individually (for example, case numbers).
Emphasize key actions – When listing menu options, present the descriptive phrase first, then speak the digit. For example: "Sales, press 1."
Adjust speaking rate – Slow the rate for critical information such as account numbers or confirmation codes.
Choose and maintain a consistent voice persona
Your TTS voice represents your brand in every customer interaction:
Align with your brand – Choose a TTS voice that matches your brand identity. A casual brand might use a friendly, energetic voice, while a financial institution should use a reassuring, professional tone.
Maintain consistency – Do not mix pre-recorded studio prompts with default TTS voices. Keep the voice and tone consistent across all menus and sub-menus.
Disclose AI usage – State upfront that the caller is interacting with an automated system to preserve trust and set appropriate expectations.
Handle dynamic content and allow interruptions
Personalization and flexibility improve the customer experience:
Personalize intelligently – Use TTS for dynamic, real-time content (for example, "Your account balance is $50.12" or "Your appointment is confirmed for Tuesday"). Do not ask for information the system already has on file.
Allow barge-in – Configure your system to accept customer input while TTS is playing. This lets returning customers skip prompts they have heard before.
Test and provide fallbacks
Thorough testing prevents poor experiences in production:
Test your prompts – Call an assigned phone number in Amazon Connect and listen as a customer would. Test in simulated noisy environments to ensure the generated voice remains intelligible. You can also test directly through Amazon Polly in the AWS console.
Provide a live agent option – Make sure the option to reach a human agent is available on the main menu and that the system routes there if a customer gets stuck.
Consider the full flow – Listen to the entire conversation path from the caller's perspective. Engage a conversational UI designer for expertise in creating pleasant customer experiences.
Voice Selection and Customization
Voice selection
Amazon Polly offers a variety of voices across many languages. Each voice has its own character and tone, so you can select one that best represents your brand. For example, you might choose a warm, friendly voice for a travel agent, or a more professional, authoritative voice for financial services.
Use the Set voice flow block in Amazon Connect to set the text-to-speech (TTS) language and voice for the flow. You can use this flow block to choose from dozens of voices in a variety of languages and accents.
To learn more about the Set voice flow block, review the following key configuration areas:
Language
Set the Language for your flow.
Voice selection
Select one of the voices available for the selected language.
For customers that want to provide an even more unique and brand focused experience to customers, creating your own Polly voice is possible as well.
Customize options
Amazon Connect offers specialized speaking styles for select voices. Examples include the following:
Conversational Style – Sounds like a friendly conversation rather than a formal announcement. This is typically the style recommended for customers for contact center workloads.
Newscaster Style – Mimics the clear, professional delivery of a news broadcaster.
Using SSML for enhanced speech control
Speech Synthesis Markup Language (SSML) gives you additional control over how Amazon Polly pronounces text.
Using SSML in Amazon Connect
Amazon Connect supported SSML tags
Practical customer support example
Using SSML for enhanced speech control
Speech Synthesis Markup Language (SSML) gives you additional control over how Amazon Polly pronounces text.
Using SSML in Amazon Connect
To use SSML in Amazon Connect do the following:
In any text-to-speech flow block, set the Interpret as field to SSML
Wrap your text in tags.
Add SSML tags to customize pronunciation, pauses, and more.
Amazon Connect supported SSML tags
The following is a selection of SSML tags supported in Amazon Connect:
break: Adds a pause (e.g., adds a one-second pause)
prosody: Controls volume, rate, or pitch (e.g., speaking slowly)
say-as: Specifies how to pronounce characters, words, and numbers
phoneme: Makes a specific phonetic pronunciation
For a full list of tags, see SSML tags supported by Amazon Connect in the Amazon Connect Administrator Guide.
Practical customer support example
Review the practical customer support example:
"I'm sorry you are experiencing an issue with your service. Your case number is CX25791."
In this example Amazon Polly is instructed to speak at 90 percent of normal speed. It then adds a short pause before reading the case as individual characters rather than, "CX twenty-five thousand, seven hundred and ninety-one."
Introduction to Amazon Lex for Intelligent Bots and IVR
Amazon Lex Core Concepts and Terminology
Amazon Lex core concepts
Amazon Lex is a fully managed AI service you can use to build conversational interfaces or chatbots for applications. It uses AI technology to recognize customer speech and text input, understand the meaning behind it, and respond appropriately.
When integrated with Amazon Connect Customer, Amazon Lex transforms traditional automated phone menus from rigid, button-pressing experiences into natural conversations.
For example, customers experience, "How can I help you today?" and "I'd like to check my account balance," instead of, "Press 1 for sales. Press 2 for support."
At the heart of Amazon Lex are the following two capabilities:
Natural language processing (NLP) refers to the overall ability of computers to work with human language.
Natural language understanding (NLU) is a subset of NLP that focuses specifically on comprehending the meaning and intent behind what a person says.
Conversational customer experience using AI is not just about recognizing the words but understanding what the customer wants to accomplish. Amazon Lex goes beyond recognizing words in customer statements. When a customer says, "I need to update my shipping address," it understands their intent to change their shipping information.
Examples of Amazon Lex
Banking example
The transition from button-pressing experiences into natural conversation is illustrated through a banking example.
The example starts with pressing 1 on a keypad, progresses to basic voice commands, and culminates in natural conversation. In the natural conversation stage, customers can say something like, "I'd like to transfer money from my checking to savings account." Each stage shows how customer interaction becomes more natural and conversational.
Figure 73 Introduction to Amazon Lex for Intelligent Bots and IVRSelect image to enlarge
The image shows how Amazon Lex processes customer speech input. The diagram uses an example: "I need to update my shipping address to AnyStreet 100." First, automatic speech recognition (ASR) processes the speech. The text moves through NLP and NLU. These systems identify the intent (UpdateShippingAddress) and extract slot information (AddressSlot: AnyStreet 100). Finally, the system fulfills the intent and sends a confirmation to the customer.
Figure 74 Recognizing a change in customer intentSelect image to enlarge
Amazon Lex terminology
You can use Amazon Lex to build applications (bots) to elicit information from users to accomplish a task. For example, you can create a bot to order flowers, book a hotel room, or order a pizza.
There are some key terms to understand when working with Amazon Lex.
Bot
An Amazon Lex bot is powered by ASR and NLU capabilities.
Amazon Lex bots can understand user input provided with text or speech and converse in natural language.
Language
An Amazon Lex V2 bot can converse in one or more languages.
Each language is independent of the others. You can configure Amazon Lex V2 to converse with a user using native words and phrases.
For more information about slots, see Languages and locales supported by Amazon Lex V2 in the Amazon Lex V2 Developer Guide.
Intent
An intent represents an action that the user wants to perform. You create a bot to support one or more related intents. For example, you might create an intent that orders pizzas and drinks. For each intent, you provide the following required information:
Intent name – A descriptive name for the intent. For example, OrderPizza.
Sample utterances – How a user might convey the intent. For example, a user might say, "Can I order a pizza," or "I want to order a pizza."
How to fulfill the intent – How you want to fulfill the intent after the user provides the necessary information. We recommend that you create an AWS Lambda function to fulfill the intent. You can optionally configure the intent so Amazon Lex V2 returns the information back to the client application for the necessary fulfillment.
In addition to custom intents, Amazon Lex V2 provides built-in intents to quickly set up your bot.
For more information about slots, see Built-in intents in the Amazon Lex V2 Developer Guide.
Amazon Lex V2 always includes a fallback intent for each bot. The fallback intent is used whenever Amazon Lex can't deduce the user's intent. For more information about the fallback intent, see AMAZON.FallbackIntent in the Amazon Lex V2 Developer Guide.
Slot
An intent can require zero or more slots or parameters. You add slots as part of the intent configuration. At runtime, Amazon Lex prompts the user for specific slot values. The user must provide values for all required slots before Amazon Lex can fulfill the intent.
For example, the OrderPizza intent requires slots such as pizza size, crust type, and number of pizzas. In the intent configuration, you add these slots. For each slot, you provide slot type and a prompt for Amazon Lex to send to the client to elicit data from the user. A user can reply with a slot value that includes additional words, such as, "large pizza please," or "let's stick with small." Amazon Lex can still understand the intended slot value.
Slot type
Each slot has a type. You can create your custom slot types or use built-in slot types. Each slot type must have a unique name within your account. For example, you might create and use the following slot types for the OrderPizza intent:
Size – With enumeration values Small, Medium, and Large.
Crust – With enumeration values Thick and Thin.
Amazon Lex also provides built-in slot types. For example, AMAZON.NUMBER is a built-in slot type that you can use for the number of pizzas ordered.
For more information about slot types, see Built-in slot types in the Amazon Lex V2 Developer Guide.
Version
A version is a numbered snapshot of your work publishable for various workflow stages. These stages include development, beta deployment, and production.
Once you create a version, you can use a bot as it existed when the version was made. After you create a version, it stays the same while you continue to work on your application.
Alias
An alias is a pointer to a specific version of a bot. With an alias, you can update the version your client applications are using.
For example, you can point an alias to version 1 of your bot. When you are ready to update the bot, you publish version 2 and change the alias to point to the new version.
Because your applications use the alias instead of a specific version, all of your clients get the new functionality without needing to be updated.
Putting Amazon Lex into Action
Now that you have explored the key terminology of Amazon Lex, review how intents and slots interact with each other in Amazon Connect.
Customer "I want to transfer money."
Amazon Lex determines that the customer is wanting to transfer money by the sample utterances provided in the TransferFunds intent. The sample utterances are used to train the bot to allow it to classify the customer's utterance.
Bot "Sure, I can help with that. Which account would you like to transfer from"
Now that Amazon Lex knows what intent the customer needs it can provide clarifying questions to help gather data needed to fulfill the intent.
Slots provide the ability to gather data to fulfill the recognized intent. The next step in the conversation is determined by prompting the user to provide the required data for intent fulfillment. In this scenario, the slot prompts for the user to provide the account details of where they would like funds to be transferred from.
Customer "My checking account." Bot "And which account would you like to transfer to"
After the customer provides a response, Amazon Lex determines whether it needs to fill any further slots. If needed, it will ask the customer to provide the necessary information.
The second slot asks for the destination account.
Customer "My savings account." Bot "How much would you like to transfer"
Amazon Lex continues to ask the customer to provide the necessary information until all slots are filled.
The third slot is for a dollar amount.
Customer: "Five hundred dollars." Bot: "Just to confirm, you want to transfer 500 dollars from your checking account to your savings account. Is that correct?"
Now that all slot data has been gathered, the intent can be fulfilled. In this scenario, a confirmation prompt has been configured to confirm the provided details before completing.
Seeing it in action in the Amazon Lex test window
You can test your bot in the Amazon Lex test window before using it within a live customer journey.
This gives you the ability to ensure that your slots are capturing the correct information.
Insider tip: Confirmation is not required for all intents and should be used selectively. For transactional intents or actions with significant consequences, implement confirmation to validate the bot's intended action before execution.
When using Amazon Lex with Amazon Connect for phone interactions, it benefits from special optimizations:
Telephony audio sampling – Amazon Lex is trained on telephony audio at an 8 kHz sampling rate, which provides increased speech recognition accuracy specifically for phone-based use cases. Because Amazon Lex is optimized to process 8 kHz audio directly, it avoids the quality loss that occurs when upsampling or downsampling between different rates, providing reliable speech recognition accuracy for phone calls.
Dual tone multi-frequency (DTMF) input – In addition to voice recognition, Amazon Lex can accept numeric input from keypad presses, which gives customers flexibility in how they respond.
Understanding 8 kHz telephony audio
The 8 kHz sampling rate (narrowband audio) is the standard for traditional telephone networks (PSTN). This limits the frequency range to a maximum of 4 kHz, which is sufficient for speech intelligibility but removes higher-frequency consonant detail.
Why this matters for your bot design:
Consonant clarity – Frequencies above 4 kHz are filtered out, which can make it harder to distinguish between similar-sounding consonants (for example, "s" versus "z" or "f" versus "th").
Background noise tolerance – In noisy environments, narrowband audio provides less separation between the caller's voice and ambient sounds.
Speech recognition accuracy – Amazon Lex is specifically trained on 8 kHz audio, which mitigates these limitations for telephony use cases. You can further improve accuracy by configuring a custom vocabulary for specialized terminology.
The evolution - Modern carriers increasingly support wideband audio (16 kHz) through VoLTE and HD Voice, which doubles the available frequency range. Amazon Connect supports these higher-fidelity connections when the caller's network provides them.
Best practice - Use an Amazon Lex custom vocabulary to help the transcription layer recognize specialized terminology (industry jargon, product names, or acronyms) that may be harder to distinguish at 8 kHz.
Building Conversational Interfaces
Key Principles
Building Conversational Interfaces
Building effective conversational interfaces is about creating a natural dialogue that feels helpful and intuitive. Here are some key principles to keep in mind:
Be human-centered – Design with the customer's needs and expectations in mind.
Keep it simple – Use clear, straightforward language.
Provide context – Help customers understand where they are in the conversation and what's happening next.
Handle errors gracefully – Plan for misunderstandings and create friendly ways to get the conversation back on track.
Confirm important information – Confirm details for critical actions before proceeding.
Well-designed conversations anticipate and adapt to customer needs. The image shows this through a typical banking exchange. A customer asks to check their account balance. The bot responds by asking which account type. When the customer changes their mind, the bot seamlessly transitions to help with a money transfer instead.
Figure 75 Building Conversational InterfacesSelect image to enlarge
Example conversational interaction between a customer and bot.
Map customer journeys – Identify the most common reasons customers contact you.
Draft sample dialogues – Write out example conversations for each scenario.
Plan for detours – Customers often change their minds or ask tangential questions.
Conversational interface designers
A conversational designer creates intuitive, engaging interactions between users and AI-powered interfaces. They design natural interactions for chatbots, voice assistants, and other conversational tools. Their work combines elements of user experience (UX) design, linguistics, psychology, and copywriting to craft dialogue flows that align with user needs and brand goals.
Designing conversational interfaces requires shifting from traditional visual UI design to natural human interaction patterns. The goal is to make the interaction feel organic, efficient, and low-effort for the user. The following best practices apply whether you are designing for voice (IVR and voicebots) or text (chatbots and AI assistants).
Voice design
Voice interfaces process information sequentially. Callers cannot scan back through what they heard, so every design decision must account for linear delivery and limited short-term memory.
Structure for cognitive load. Use the "one breath" rule: prompts should take fewer than four seconds to speak aloud. Front-load the critical action or keyword at the beginning of each sentence. For example, use "To book a flight, say flight" rather than "Say flight if you would like to book a flight."
Limit options per decision point. Offer a maximum of three to four choices at any single prompt. Long lists cause callers to forget the first options by the time they hear the last ones. If you need more options, organize them into categories and let callers navigate progressively.
Support barge-in. Allow callers to interrupt a prompt the moment they hear the option they want. Forcing a caller to listen to a 20-second recording before responding wastes their time and increases abandonment.
Handle over-answering. If the system asks "What date would you like to travel?" and the caller says "Next Tuesday with my wife," the system should extract the date and store the extra context (two tickets) rather than returning an error.
Text design
Text interfaces give users control over pacing. They can read at their own speed, scroll back, and see visual anchors. This creates different design opportunities and constraints.
Use visual structure. Break responses into short paragraphs. Use bullet points for lists and bold text for key terms. Callers read on small screens, so concise formatting improves comprehension.
Provide interactive elements. Buttons, quick-reply chips, and carousels reduce typing effort and guide users toward valid responses. These visual aids have no equivalent in voice channels.
Allow longer responses when appropriate. Unlike voice, text interfaces can present a full paragraph or a comparison table without overwhelming the user. However, keep each message focused on a single topic to maintain clarity.
Support rich media. Link to images, documents, or calendar widgets when they help the user complete a task faster than text alone.
Amazon Lex and Amazon Connect Integration
Integration Overview
Amazon Lex and Amazon Connect Integration
Amazon Connect and Amazon Lex work seamlessly together to create a powerful platform for conversational customer service.
With this integration, you can do the following:
Replace rigid IVR menus with natural conversations.
Automate routine customer inquiries and transactions.
Collect information before transferring to agents.
Provide 24/7 self-service options.
Deliver consistent experiences across voice and chat.
Building and Testing an Amazon Lex Bot in Amazon Connect
Transcript Building and Testing an Amazon Lex Bot in Amazon Connect
Create and configure the bot
AWS has streamlined the Amazon Lex bot creation process, so you can build complete conversational experiences without leaving Amazon Connect. First, you will need to log into your Amazon Connect admin workspace by navigating to your instance URL, such as https://yourinstancename.my.connect.aws. You will either need the Admin security profile or make sure your security profile includes the "Channels and Flows - Bots - Create" permission.
Once logged in, navigate to the bot creation area. In the navigation menu, choose on Routing, then select Flows.
Now on the Flows page, you can see a Bots tab. Choose Bots, then select Create bot.
Excellent! The Details dialog box has opened. Now configure our bot with the following information. For bot Name, enter "BankerBot". Remember, this needs to be a unique name within your AWS account. Bot Description is optional, but adding context is recommended. Enter, "Bot to help with banking actions". For compliance with COPPA, or the Child Online Privacy Protection Act, for this banking bot, select No. Since this application will not target children under 13. Then, choose Create to proceed.
Perfect! Amazon Connect has successfully created the bot and directed us to the bot configuration page. As you can see, the BankerBot is ready for configuration. The next step is adding language support. To choose the initial language to use for the bot, choose Add language.
You can add multiple supported languages to your bot, but for this demonstration, choose English (US) as the initial language.
Define the transfer intent
Now that language has been added, you can add your intents, which represent the goals your users want to accomplish. Remember, there are two types of intents. Custom intents represent specific actions your bot should handle. Built-in intents are for common actions. Every bot automatically includes a fallback intent for unrecognized requests.
To create your first custom intent, choose Add intent, then choose Add empty intent.
In the Add intent dialog, enter "TransferMoney" for the Intent Name and "Providing the ability for customers to transfer funds from one account to another" for the Description. Then, choose Add.
Great! Your intent has been created. Now you can configure the utterances. These are phrases users might say to trigger this intent. Choose Add and enter phrases like "Please transfer money from { FromAccount } to { ToAccount } for { Amount }", then choose Add. You can add phrases like "I want to transfer some money" and so on. Any word wrapped in parentheses represents a slot parameter. You can add those next. Once you have entered all of your phrases, choose Save.
Next, you can configure the slots. Slots are parameters required to fulfill the intent. To start defining the slots, choose Add. Remember the slot parameters are the words you added in the utterance phrases that were wrapped in parentheses. Start with Amount, setting the slot Name, then select the slot Type as AMAZON.Number. Also, add the slot Prompt to "How much would you like to transfer?". Finally, choose Add.
Then, repeat the process for the FromAccount slot, and the ToAccount slot. Set the slot types to AMAZON.Number and add the Slot Prompts of, "Which account would you like to transfer funds from?" and "Which account would you like to send funds to?". Then, choose Add for each slot, then Save when finished.
Now you can configure the prompts. To configure the messages your bot will use, choose Edit. For the Confirmation prompt message, enter "Would you like me to proceed in transferring { Amount } from { FromAccount } to { ToAccount }?" For the Decline response, enter "Okay, I won't transfer the money.". When you are finished, choose Save.
The last step in the process of building a bot is to build the bot's language. By building the bot's language, you can use the bot within our Amazon Connect Flow that we will configure. To start the build process, select the Build language button. Shortly after that the build will be complete. That's how you build a complete Amazon Lex bot directly within Amazon Connect. You've successfully created a functional financial services bot with intents, utterances, slots, and conversational prompts - all without leaving the Amazon Connect interface.
Build the contact flow
Now navigate to the Flows management area. In the navigation menu, choose Routing, then choose Flows.
On the Flows page, choose Create flow.
In the flow designer, you can enter the Flow name "Banking bot flow".
In the flow designer, drag the Play prompt block from the Interact section of the Block Library onto the canvas.
Drag the arrow from the Entry point block and connect it to the Play prompt block.
Next, choose the Play prompt block to open the block configuration pane on the side of the screen. Then choose the Text-to-speech or chat text option.
Enter your welcome message into the textbox. In this demonstration you can welcome the customer with, "Welcome to the example Banking Bot," and then choose Save.
Now, from the Block Library, drag a Get customer input block from the Interact section, onto the canvas.
Connect the Success branch of the Play prompt block to the input of the Get customer input block. Then choose the Get customer input block to open the configuration pane. Now choose the Amazon Lex tab under the Config tab.
Next, choose the Select a Lex bot list, and choose your Amazon Lex bot. Select Banker Bot.
Next choose the Alias list and select your bot alias. In this demonstration, you can use the TestBotAlias. Remember that the TestBotAlias should not be used for production traffic.
Next, under the Customer prompt or bot initialization section, choose the Text-to-speech or chat text option and provide the text you want to play to the customer when they meet your Amazon Lex bot. For this demonstration, enter "How can I help you today?".
Under the Intents section, choose the Add an intent link twice to add two empty intent branches. Next you need to type the names of the intents from the BankerBot Amazon Lex bot. You can enter TransferMoney and FallbackIntent into the empty intent branch boxes. Then when complete, choose Save.
You can now see that the Get customer input block has updated with the two intent branches that you added.
From the Block Library, drag a Play prompt block from the Interact section, onto the canvas. Connect the TransferMoney branch of the Get customer input block to the new Play prompt block. Next, choose the Play prompt block to open the block configuration pane on the side of the screen. Select the Text-to-speech or chat text option. In the textbox, enter the message to give the customer once they have completed the TransferMoney intent. In this demonstration, you can say, "Thank you for contacting us." and then choose Save.
Drag another Play prompt block from the Interact section, onto the canvas. Connect the FallbackIntent branch of the Get customer input block to the new Play prompt block. Then choose the Play prompt block to open the block configuration pane on the side of the screen. Select the Text-to-speech or chat text option. In the textbox, enter the message to give the customer if they hit the Fallback intent as you will transfer them through to an agent. In this demonstration, you can say, "I will now transfer you to an agent." and then choose Save.
From the Block Library, under the Set section, drag a Set working queue block onto the canvas. Connect the FallbackIntent linked Play prompt block Success branch to the Set working queue block. Then, choose the Set working queue block to open the block configuration pane on the side of the screen. Select the Search for queue list. Then choose the BasicQueue and choose Save.
From the Block Library, under the Terminate section, drag a Transfer to queue block onto the canvas. Connect the Success branch from the Set working queue block to the Transfer to queue block. Then drag a Disconnect block onto the canvas. Connect the At capacity and Error branches from the Transfer to queue block to the Disconnect block.
Now, connect the Error branch from the Set working queue block to the Disconnect block.
Then, connect the Error branch from the FallbackIntent linked Play prompt block to the Disconnect block. Next, connect the Success branch and the Error branch from the TransferMoney linked Play prompt block to the Disconnect block. Now, connect the Default branch of the Get customer input block to the FallbackIntent linked Play prompt block. And the Get customer input block Error branch to the Disconnect block. Then, connect the Error branch from the welcome message Play prompt block to the Disconnect block.
Publish and test
Finally, you need to publish the flow so that it can be used in a customer journey. Only published flows can be assigned to a phone number or configured in a communications widget. In the corner of the flow designer, choose the Publish button.
At the Publish confirmation dialog, choose Publish to activate the flow immediately.
Now the flow is active, so you can test the complete customer experience. From the navigation menu, choose Home to go to the Configuration guide.
In the Configuration guide, under Step 1. Explore your channels of communication, choose the Test chat link to go to the Test chat page.
On the Test chat page, choose the Test Settings link, which will open the Test Chat Settings dialog. Under System Settings, select the Contact Flow list and find the contact flow that was just published. In this demonstration, you can choose the Banking bot flow that was just built and published. Then, choose Apply.
The test chat widget will load and enter the flow that was selected. In this demonstration, you're presented with the configured welcome message and the message that was configured in the Get customer input block. The flow is now waiting for an input, so you can enter, "I want to transfer some money", one of the sample utterances that was configured in the BankerBot.
The BankerBot has understood the request and recognized that you want to transfer some money by selecting the TransferMoney intent. Now the Slots parameters need to be filled to be able to fulfil the intent. You can provide the account number "123456789" as a source account and send the message back to the bot.
Now the ToAccount slot needs to be filled. You can provide the account number "987654321" as the destination account and send the message back to the bot.
The third slot, Amount, provides the remaining customer prompt. You can enter the amount "12345".
The last prompt back from the bot is the Confirmation prompt, asking to confirm the transfer of the funds. You can send, "yes", back to the bot.
The bot has collected and confirmed the transfer details, and the sample flow has completed. Executing a real transfer requires an authorized backend fulfillment integration.
After implementing your Amazon Lex bot in Amazon Connect, test thoroughly and optimize over time.
Important language considerations
Figure 76 Building Conversational InterfacesSelect image to enlarge
When using an Amazon Lex V2 bot, the language attribute in Amazon Connect must match the language locale used to build your Amazon Lex bot. For example, if your Lex bot uses Australian English (en_AU), you must configure Amazon Connect to use the same language locale using either of the following:
A Set voice block to indicate the Amazon Connect language model
A Set contact attributes block to specify the language
This ensures that speech recognition works correctly for your customers.
Optimizing Self Service with Generative AI
Generative AI Capabilities for Amazon Lex
Applying Generative AI with Lex Bot Creation
Optimize your Amazon Lex bot creation and self-service performance by using generative AI. You can take advantage of the Amazon Bedrock generative AI capabilities to automate and speed up your Amazon Lex bot building process. Amazon Bedrock generative AI capabilities for Amazon Lex can generate intents, slots, and sample utterances automatically.
Amazon Lex currently provides the following generative AI capabilities to optimize Amazon Lex V2 bots:
Create new bots and populate them with relevant intents and slot types efficiently using natural language description.
Generate sample utterances for your bot intents automatically.
Improve your bots' slot resolution performance.
Create an intent to help answer your customer questions.
Use Amazon Bedrock Agents and Amazon Bedrock knowledge bases to help answer your customer's questions.
Improve intent classification and slot resolution.
Note: These features use generative AI. As you use Amazon Lex, remember that it may give inaccurate or inappropriate responses. For more information, see AWS Responsible AI Policy.
Powered by Amazon Bedrock: AWS implements automated abuse detection. Amazon Lex V2 generative AI features are built on Amazon Bedrock. As a result, users inherit Amazon Bedrock controls for enforcing safety, security, and responsible AI use.
You can activate generative AI capabilities for Amazon Lex V2 either through the console or API.
Using the console
Sign in to the AWS Management Console and open the Amazon Lex V2 console at https://console.aws.amazon.com/lexv2/home.
Select the bot and the locale in the bot for which you want to turn on generative AI capabilities.
In the Generative AI configurations section, select Configure.
Toggle the Enabled button for each feature that you want to activate. Select the model and version that you want to use for that feature. Enabling a feature may incur additional charges.
Select Save after you turn on the features that you want to activate. A green success banner appears to confirm that the capabilities are turned on.
Using the API
To enable generative AI capabilities for a new bot, use the CreateBot operation to create a new bot.
Send a CreateBotLocale request, modifying the generativeAISettings object as necessary. If you are enabling the capabilities for an existing bot, send an UpdateBotLocale request instead.
To enable usage of the descriptive bot builder, modify the descriptiveBotBuilder object. Specify the foundation model to use in the modelArn field and set the enabled value to True.
To enable slot resolution improvement, modify the slotResolutionImprovement object. Specify the foundation model to use in the modelArn field and set the enabled value to True.
To enable sample utterance generation, modify the sampleUtteranceGeneration object. Specify the foundation model to use in the modelArn field and set the enabled value to True.
Generative AI Capabilities in Detail
Review each of the capabilities to learn how generative AI can help you build effective customer experience journeys using Amazon Lex V2.
Descriptive bot builder
Create a bot by using natural language to describe what the bot should be able to do. Amazon Lex V2 invokes Amazon Bedrock models to generate intents and slot types that fit your bot's use case.
Here are some helpful example bot descriptions you can use with descriptive bot builder in Amazon Lex V2:
Industry
Example prompt
Financial services
"Our financial card service assists users with essential tasks for new cards. These tasks include card activation, PIN delivery through email or mail, and card verification using a zip code. We help customers with tasks associated with their existing credit cards. These tasks include inquiring about benefits, reporting lost cards, requesting new cards, resetting PINs, and paying bills."
Food services
"I want a bot to help customers order food (using item ID, quantity, size), check order status, and cancel an order. Use Order ID for indexing orders."
Airline
"We are an airline domain that helps users book flight tickets and manage their reservations. Our services include checking reservation details, obtaining receipts, inquiring about flight status, rescheduling, eliciting flight details, and canceling booked flights. You can also generate additional intents if they help support functions in the domain description."
Insurance
"We are an insurance company that sells car, home, and annuity insurance policies. I want a bot that can check claim status, file a claim, make policy payments and cancel a policy. We use policy_id and last four digits of the Social Security Number (SSN) for account identification and validation."
Vehicle management
"We are building a Towed Cars Lookup bot that helps drivers whose car has been towed to find where the car is located. This bot should ask for the address or location where the automobile was towed from. It should also gather details about the vehicle, including its license plate, make, model, and year."
Travel
"I am a travel agent, and I want a bot to help my customers book a trip to a AnyCompany theme park. AnyCompany has several parks all over the world to choose from and also has hotels, dining, and special entertainment that can be reserved."
For more information, see Use a description to build a bot in Lex V2 with the descriptive bot builder in the Amazon Lex V2 Developer Guide.
Utterance generation
Use utterance generation to automate the creation of sample utterances for your intent.
Amazon Lex V2 generates sample utterances for you based on the intent name, description, and existing examples. This feature reduces the time and effort you spend in discovering and writing your own sample utterances.
For more information, see Use utterance generation to generate sample utterances for intent recognition in the Amazon Lex V2 Developer Guide.
Assisted slot resolution
You can improve the accuracy of some built-in slots in your bot's conversation flow by using assisted slot resolution.
Assisted slot resolution uses Amazon Bedrock large language models (LLMs) to improve recognition of some built-in slots, which results in an improved interpretation of customer responses during slot elicitation. For utterances that could not be resolved normally, you can attempt to resolve them a second time using Amazon Bedrock.
With assisted slot resolution, you can use the power of Amazon Bedrock foundation models to improve the accuracy of the following built-in slots:
AMAZON.Alphanumeric without regex support
AMAZON.City
AMAZON.Country
AMAZON.Date
AMAZON.Number
AMAZON.PhoneNumber
AMAZON.Confirmation
For more information, see Using assisted slot resolution to clarify slot values in Amazon Lex V2 in the Amazon Lex V2 Developer Guide.
QnAIntent
Amazon Lex V2 offers a built-in AMAZON.QnAIntent that you can add to your bot. This intent harnesses generative AI capabilities from Amazon Bedrock by recognizing customer questions and searching for an answer in a knowledge store you have set up.
For example, "Can you provide me details on the baggage limits for my international flight?"
This feature reduces the need to configure questions and answers using task-oriented dialogue within Amazon Lex V2 intents. This intent also recognizes follow-up questions such as, "What about domestic flights?" based on the conversation history and provides the answer accordingly.
For more information, see AMAZON.QnAIntent in the Amazon Lex V2 Developer Guide.
Amazon Bedrock Agents
Use Amazon Bedrock Agents to handle complex workloads requested by customers without having to go through a comprehensive task definition process. Amazon Lex V2 offers a built-in AMAZON.BedrockAgentIntent that you can add to your bot.
This intent harnesses generative AI capabilities from Amazon Bedrock by recognizing customer requests, analyzing them, reasoning them, and responding. It also has the capability to ask any follow-up questions to achieve the task needed.
For example, imagine you defined a retail agent that can check customer's order status. When the customer asks for order status, agent first requests customerId or associated emailId to retrieve the details and responds with correct order status. You can also decide to integrate your AMAZON.BedrockAgentIntent with an Amazon Bedrock knowledge base to directly answer any customer queries.
For more information, see Using BedrockAgentIntent to use a Amazon Bedrock Agent in Amazon Lex V2 in the Amazon Lex V2 Developer Guide.
Improve intent classification and slot resolution
Assisted NLU is a feature that uses LLMs to improve Amazon Lex V2 intent classification and slot resolution capabilities. It enhances accuracy while staying within your bot's configured intents and slots.
The feature does not generate or modify any bot content. This feature helps to improve the overall accuracy of the NLU system, resulting in a more seamless and effective conversational experience for users.
For more information, see Improve intent classification and slot resolution in Lex V2 with assisted NLU in the Amazon Lex V2 Developer Guide.
By using the generative AI capabilities for Amazon Lex, you can significantly transform and enhance your self-service customer experience. This streamlines bot development, automatically generates comprehensive dialogue components, and continuously improves accuracy through advanced NLU. This powerful combination reduces operational costs and accelerates time-to-market for new self-service solutions. Most importantly, it delivers more intuitive and effective customer interactions. As you implement these generative AI features, you can create responsive bots that better understand customer intent and provide more accurate, contextual responses. This leads to higher customer satisfaction and reduced support workload for your organization.
Amazon Lex Next Generation NLU and Voice Options
LLMs as Primary NLU
Three major advances in Amazon Lex change how you design and build conversational self-service experiences: LLMs now serve as the primary engine for understanding customer intent, third-party voice AI models can be integrated alongside native AWS services, and language support has expanded considerably.
Earlier in this course, you learned how Amazon Lex uses natural language understanding (NLU) to identify customer intent and extract slot values. Traditionally, this required defining intents with sample utterances and configuring slots manually. That approach still works, but a newer and more powerful option is now available.
Amazon Lex now offers large language models as the primary option for understanding customer intent. This changes the development experience in a few important ways:
What changes for developers
With LLM-powered NLU, you no longer need to provide exhaustive lists of sample utterances for each intent. The model generalizes from fewer examples and handles unexpected phrasing more gracefully.
A customer might say, "Hey, my internet has been acting weird since Tuesday and I'm pretty frustrated." Traditional slot-based NLU might struggle to map this to the correct intent. LLM-powered NLU interprets the meaning naturally and routes it appropriately.
What changes for customers
Customers can speak naturally without adapting their language to match what the system expects. Multi-part requests work without requiring separate conversation turns. The system maintains context better across longer exchanges.
When to use each approach
LLM-powered NLU is the recommended default for new bots. Traditional slot-based NLU remains appropriate when:
You need highly predictable, deterministic routing
Your use case has a small number of well-defined intents
Amazon Lex integrates with Amazon Bedrock through built-in native features and custom AWS Lambda hooks. This integration allows you to enhance Lex with large language models (LLMs) to handle complex, unscripted customer conversations.
Runtime integration answering users
When a user speaks or types to Amazon Lex, Bedrock handles the open-ended conversation layers through two native built-in intents:
AMAZON.QnAIntent (Retrieval-Augmented Generation) – You can link a Lex intent directly to an Amazon Bedrock Knowledge Base (which typically points to company documents vectorized in Amazon OpenSearch). When a user asks a question, Lex intercepts it, queries Bedrock, retrieves the document context, runs it through an LLM, and returns the summarized answer to the customer without requiring custom backend code.
AMAZON.BedrockAgentIntent (complex orchestration) – If you require your bot to execute multi-step tasks (for example, "Book a flight and email me the confirmation"), Lex can delegate the conversation to an Amazon Bedrock Agent. The agent breaks down the request, determines what APIs to call, and prompts the user for missing details dynamically.
Guardrails and fallback using AWS Lambda
If you require strict control over when the LLM is allowed to respond to your customer, you can route the integration through an AWS Lambda function:
Intent shield – Lex handles deterministic tasks first (for example, collecting account numbers, capturing alphanumeric codes).
Bedrock fallback – If the user says something unexpected, Lex triggers its FallbackIntent. The Lambda function forwards the transcript to Amazon Bedrock, validates the response, and passes it back to Lex for delivery.
Automated bot building developer layer
Amazon Bedrock also assists within the AWS Console to reduce bot development time:
Descriptive bot builder – You describe your ideal bot using natural language, and Bedrock automatically generates the required Lex intents, slots, and utterances.
Assisted slot resolution – If a customer inputs a variation of a slot that your bot does not strictly recognize, Bedrock uses an LLM to infer the meaning. For example, if the slot expects "Large Pizza" and the caller says, "Give me the biggest size you have," the LLM maps that phrase to the "Large" slot value.
How the voice pipeline works
When a user calls your IVR or talks to your voicebot, the interaction flows through a three-step pipeline:
Speech-to-text (STT) – Amazon Lex takes the incoming audio from the caller and passes it through its built-in speech recognition engine to convert spoken words into text.
LLM processing – Lex takes that text string, along with any prompt context or instructions, and passes it to Amazon Bedrock via API. The LLM analyzes the text and generates a text-based response.
Text-to-speech (TTS) – The LLM's text output is sent back to Lex, which pushes it to Amazon Polly to convert back into audio for the caller.
Important: In the traditional pipeline, Amazon Lex transcribes audio to text before any data is sent to an LLM in Amazon Bedrock. This means the LLM cannot hear background noise, detect tone or sarcasm, or sense urgency directly from the audio. It receives only the transcribed words. With Nova Sonic integration, speech-to-speech processing bypasses this step.
Third Party STT and TTS Integration
Amazon Connect no longer requires you to use only AWS-native models for speech processing. You can now integrate third-party AI providers for both speech-to-text and text-to-speech in self-service interactions.
Deepgram for speech to text
Deepgram can serve as an alternative to Amazon Transcribe for converting customer speech into text. Organizations might choose Deepgram for:
Specific accuracy advantages in certain accents or domains
Faster processing speed for particular use cases
Existing enterprise agreements or familiarity
ElevenLabs for text to speech
ElevenLabs can serve as an alternative to Amazon Polly for generating spoken responses. Organizations might choose ElevenLabs for:
Highly natural, expressive voice synthesis
Custom voice cloning for brand consistency
Specific voice characteristics not available in native options
Choosing between native and third party
Expanded Language Support
The wait and continue functionality in Amazon Lex, which keeps customers engaged during processing delays, now supports 10 additional languages.
The following languages are now supported for Lex wait and continue interactions:
Chinese (Mandarin)
Japanese
Korean
Cantonese
Spanish
French
Italian
Portuguese
Catalan
German
This expansion enables organizations to build multilingual self-service experiences that maintain engagement across a broader range of customer populations without requiring separate bot configurations for each language.
Introduction to Generative AI Self Service with Amazon Q in Connect
The Power of a Unified AI Assistant
Rather than building separate generative AI tools for agents and customers, Amazon Q in Connect offers a unified approach. The same AI assistant that helps your agents find information and resolve customer issues can now interact directly with your customers through your IVR and digital channels. Your carefully curated knowledge base, established integrations, and refined configurations for agent assistance are valuable assets. These resources can be used to create exceptional self-service experiences for your customers.
Think about your current contact center operations. When a customer calls, they might start in your IVR system, potentially transfer to a virtual agent, and sometimes need assistance from a human agent. Traditionally, each of these touch points might use different technologies and access different knowledge sources. With Amazon Q in Connect, you are creating a seamless experience powered by a single, intelligent AI assistant that maintains context and consistency throughout the customer journey.
Customer Self Service Interactions with Amazon Q in Connect
How Amazon Q in Connect Works for Self Service
At its core, Amazon Q in Connect uses large language models (LLMs) to process and understand customer inquiries. When implemented for self-service, Amazon Q in Connect works much like it does for your agents. Amazon Q in Connect provides a customer-facing interface through your IVR or digital experience for self-service. Amazon Lex with Amazon Q in Connect transforms how these interactions happen by using the knowledge articles stored in your knowledge base. The same knowledge base that is established for your agents.
Watch the following video to see how Amazon Lex and Amazon Q in Connect create natural, helpful conversations for your customers.
Video: Video - Customer Self-service with QiC.mp4
Transcript Customer Self Service Interactions with Amazon Q in Connect
Let's follow Sofía, as she explores home insurance options with AnyCompany insurance.
She starts on the AnyCompany website, and opens the embedded chat interface. Because Sophia visited before and provided her information, she's greeted by her name.
Sofía starts by saying that she recently bought a home and wants to know how she can get the right coverage.
The answers provided are powered by a generative AI assistant for customer service, Amazon Q in Connect. Amazon Q in Connect is powered by your knowledge articles to provide real-time assistance to both customers and agents across chat interactions like you see here, as well as voice calls.
Amazon Q in Connect can be turned on in your self-service experience in just a few steps and can provide the right answers while carrying on a conversation.
Here, Amazon Q in Connect can provide an overview about the available home insurance coverages, while being able to dive into follow-up questions about the differences.
Amazon Q in Connect can be personalized, to give you control over its tone and behavior.
You can also incorporate data to augment your knowledge articles and tailor the conversation in real time.
Amazon Q in Connect also has the ability to implement guardrails. Amazon Q in Connect can acknowledge when a licensed agent is needed to provide a definitive answer and will provide that option to Sofía.
Sofía isn't quite ready to talk to someone, but she does want to make sure that when she is ready, she doesn't have to go through this process again. So Sofía fills out some key information through this self-service experience provided by Amazon Connect step-by-step Guides.
AnyCompany insurance can securely collect and process sensitive information such as Social Security numbers, credit card details, addresses, and more. This ensures that the agent doesn't have to be exposed to this kind of data.
Details that have already been provided through this conversation are prefilled but editable. And Sofía just has to add the additional information about either herself or her property.
Sofía gets to the last section, but she gets interrupted by an incoming call. Distracted, Sofía doesn't fully complete the form.
But that's not a problem. This information will be stored and ready to continue when Sofía comes back to get the quote that she's looking for.
The conversation flows naturally because Amazon Q in Connect understands context and maintains conversation history. It can handle follow-up questions and even proactively offer related information about delivery options or return policies.
When to use Amazon Lex versus Amazon Q in Connect
Amazon Lex and Amazon Q in Connect serve different purposes in your self-service architecture. Understanding this distinction helps you select the right tool for each customer need.
Q&A and knowledge assistance (searching documents, generating answers)
How it operates
Intent-based: relies on designed flows to gather specific data slots
Search-based: reads unstructured data and summarizes answers dynamically
Primary user
The end-customer (acting as IVR voicebot or web chatbot)
The human agent (populating their screen) or the customer for open FAQs
Strengths
Strict transactional operations, data validation, deterministic routing
Open-ended questions, synthesizing information from multiple sources
Limitations
Requires a developer to code intents for each scenario
Not designed for strict transactional workflows requiring validation
AMAZON.QinConnectIntent, which passes the question to Amazon Q in Connect
Think of Amazon Lex as an automated front desk clerk that excels at executing specific tasks in sequence. Think of Amazon Q in Connect as an expert research assistant that excels at parsing extensive documentation and synthesizing natural language answers.
Key benefits of using Amazon Q in Connect for self-service:
Improved customer experience: Customers receive immediate, accurate responses without waiting for agents to become available.
24/7 availability: Self-service options remain available around the clock.
Reduced contact center load: By handling routine inquiries automatically, Amazon Q in Connect frees up human agents for more high-value interactions.
Consistent information: Using the same knowledge base that your agents use, customers receive the same accurate information at all times.
Cost efficiency: Automated handling of routine inquiries can significantly reduce operational costs.
Scalability: Build your knowledge base once and use it across your entire contact center environment.
By implementing Amazon Q in Connect for customer self-service interactions, you are creating a seamless, intelligent experience that mirrors the quality of human agent assistance. This powerful solution uses your existing knowledge base to deliver consistent responses 24 hours a day, while maintaining natural conversation flow and context awareness. This results in improved customer satisfaction through instant access to information, reduced wait times, and personalized interactions that can seamlessly transition to human agents when needed. As demonstrated in Sofía's journey, customers can engage naturally, receive comprehensive answers, and even pause and resume their experience. Your organization benefits from reduced operational costs, increased scalability, and the ability to handle more complex inquiries with your human agents. Amazon Q in Connect transforms self-service from a basic FAQ system into an intelligent, conversational experience.
Understanding when to use Amazon Lex versus Amazon Q in Connect
Amazon Lex and Amazon Q in Connect serve different primary purposes within Amazon Connect self-service. Understanding the distinction helps you select the right tool for each customer scenario and avoid building duplicate solutions.
Amazon Lex
Amazon Q in Connect
Amazon Lex is a conversational engine for structured, intent-driven actions. Think of it as an automated front desk clerk that follows a defined sequence of steps to complete a specific task.
Primary role: Execute deterministic self-service transactions such as booking appointments, resetting passwords, checking account balances, and routing calls.
How it operates: Lex relies on configured intents, slots, and fulfillment logic. A customer says "I want to check my balance." Lex matches the CheckBalance intent, prompts for the account number slot, verifies the PIN slot, calls a backend system via AWS Lambda, and reads the result back to the caller.
Strengths: Strict transactional workflows, data validation, slot capture, and predictable routing logic.
Limitation: If a customer asks an open-ended question like "What is your return policy for damaged goods bought during a holiday sale?" Lex cannot answer unless a developer specifically coded an intent for that scenario.
Amazon Q in Connect
Amazon Q in Connect is a generative AI knowledge assistant that retrieves information and resolves issues through search and reasoning. Amazon Q in Connect directly helps customers resolve their issues through automated self-service interactions.
Primary role: Answer open-ended questions, search unstructured knowledge sources, and provide contextual guidance without requiring pre-built intent flows.
How it operates: Amazon Q in Connect reads your connected knowledge sources (documentation, policy files, product catalogs) and synthesizes natural language answers in real time. It can also perform actions defined through tools, such as checking order status, processing returns, or scheduling callbacks.
Strengths: Handling ambiguous questions, synthesizing information from multiple documents, maintaining conversational context across follow-up questions, and adapting responses based on customer profiles.
Customer-facing capabilities: When exposed directly to customers, Amazon Q in Connect handles unscripted scenarios. A customer can ask "Can I return a laptop I bought three weeks ago if I lost the receipt?" and receive a contextual answer drawn from your return policy documentation.
When to use which, and how they work together. You do not have to choose one over the other. Amazon Connect supports both in a unified experience. Use Amazon Lex for structured transactions where you need predictable data capture and deterministic fulfillment (booking, payments, password resets, and call routing). Use Amazon Q in Connect for knowledge retrieval, open-ended questions, and scenarios where the customer's request does not fit a predefined intent flow.
In practice, the two services complement each other. Lex handles the structured front door: identify the customer, capture account details, and route to the correct workflow. When a customer asks something that falls outside configured intents, Lex can seamlessly hand off to Amazon Q in Connect, which searches your knowledge base and provides the answer. Throughout this process, full conversation context transfers between services, so the customer never repeats information. When the self-service experience cannot fully resolve an issue, both Lex and Amazon Q in Connect support seamless handoff to a human agent with the complete interaction history preserved.
Understanding Natural Customer Conversations
Natural Language Understanding
Think about how customers naturally ask questions. Someone might ask, "How do I return my new laptop?" and another might say, "I need to send back an electronic item that I bought." Though phrased differently, both customers need the same information. Amazon Q in Connect understands these variations in human communication, and interprets the true intent behind customer questions rather than simply matching keywords.
Amazon Q in Connect processes these questions using advanced natural language understanding and generates responses that feel natural and helpful. Instead of receiving a rigid, prewritten script, customers get conversational responses that directly address their specific concerns. For example, when asked about electronics returns, Amazon Q in Connect might respond like this:
"I can help you with returning your electronic item. Our electronics return window is 30 days from purchase, and you will need the original packaging and receipt. Would you like me to guide you through the return process or explain our testing requirements for electronic returns?"
Example Customer Interaction
Example customer interaction with Amazon Q in Connect
Review the following diagram to see how a customer might interact with Amazon Q in Connect when needing to change a previously booked flight.
Intent recognition
Amazon Q in Connect understands variations in how customers phrase their needs.
Natural responses
Responses feel conversational rather than scripted.
Conversational memory
The system maintains context throughout the interaction.
Follow up capability
The system can handle follow-up questions and clarifications.
Understanding how customers naturally communicate is fundamental to creating effective self-service experiences. Amazon Q in Connect excels at bridging the gap between human expression and automated assistance. By interpreting intent rather than relying on exact keyword matches, Amazon Q in Connect delivers truly conversational interactions that feel intuitive and helpful to your customers. The result is a self-service experience that customers find genuinely useful and engaging. It understands not just what they are asking, but what they really need to accomplish. This leads to faster resolutions and higher satisfaction rates across all customer touchpoints.
Optimizing Your Knowledge Foundation
Knowledge Base Optimization
For these conversations to work effectively, Amazon Q in Connect draws information from your organization's knowledge base. The same knowledge base of carefully curated information that helps your agents can now drive customer self-service interactions. This might include your internal documentation, policy details, product specifications, and lessons learned from successful customer interactions.
The key is to ensure that this information is well-organized, up to date, and relevant for the type of interactions that you receive. When you update a policy or add new product information, both your agents and the self-service system immediately have access to the latest details.
Confirm up to date content
Ensure that your knowledge base contains accurate, up-to-date information.
Structure your content
Well-structured content helps guide Amazon Q in Connect to provide your agents with more refined responses.
Consider breaking up long sections to shorter focused paragraphs centered on one idea.
Break down complex processes into clear steps by using bulleted or numbered lists when appropriate.
Simplify language and be specific
Be sure to write clearly and simply without complex vocabulary or sentence structures and steer away from making broad, vague statements.
Plain, straightforward writing is easier for Amazon Q in Connect and your agents to comprehend
Supplement visuals with descriptions
Amazon Q in Connect can analyze and extract meaningful information from unstructured or semistructured documents with advanced parsing enabled by default.
When using PDF files, it involves breaking down the document into its constituent parts. Such as text, tables, images, and metadata, and identifying the relationships between these elements.
All other files types you should provide descriptive text to supplement visuals like images, charts, and graphs that are referenced in your documentation.
It's important to add captions and summaries directly to your documents to help Amazon Q in Connect better interpret this content.
Define terms and reference unknown terminology
Include definitions for acronyms, abbreviations, and uncommon terms unique to your business on first use in each article.
Connect these terms to the phrases that your customers use to aid the understanding of Amazon Q in Connect.
Use examples
Always illustrate concepts with specific use cases by using examples.
By providing these details liberally, you ground how well Amazon Q in Connect understands your business, as well as the agent's expected responses and actions.
Stay agile, stay informed, and let your knowledge base be a dynamic evolving repository that your agents and generative AI-powered capabilities like Amazon Q in Connect can use
Responsible AI and Continuous Improvement
Responsible AI in self-service:
Use AI guardrails: Configure guardrails to prevent inappropriate responses and ensure accuracy.
Enable citations: Configure your AI assistant to cite sources when providing information.
Personalize responses: Pass customer context to Amazon Q in Connect to provide personalized responses.
Add conversation bridges: Create smooth transitions when the system needs to hand off to a human agent.
Monitor and refine: Regularly review transcripts of Q&A interactions to identify improvement opportunities.
Continuous improvement
Success with Q&A capabilities comes from continuous refinement. By reviewing conversation transcripts, you can identify areas where the system excels and where it needs improvement. Look for patterns in customer questions that might indicate gaps in your knowledge base or opportunities to make responses more helpful. Remember, the goal is to create conversations that feel natural and helpful while providing accurate information. As you implement and refine these capabilities, focus on how well the system understands customer intent and how naturally it can engage in back-and-forth dialogue. This attention to conversational quality helps ensure that your customers receive the kind of experience that builds trust and satisfaction with your self-service options.
Actions for Amazon Q in Connect Self Service
What Are Actions
What are Actions for Amazon Q in Connect Self service
Actions in Amazon Q in Connect allow the AI assistant to do more than just answer questions. They facilitate self-service experiences for customers. Actions in Amazon Q in Connect are defined through tools. Actions can intelligently gather information and orchestrate the customer journey through your contact flow.
Think of actions as intelligent instructions that help guide customers through your self-service experiences. When a customer expresses an intent like "I need to change my appointment," Amazon Q in Connect can collect and validate the necessary information. It then routes to the appropriate self-service flow or Amazon Lex bot that handles rescheduling.
The power lies in how Amazon Q in Connect uses natural conversation to do the following:
Understand customer intent more accurately.
Gather required information naturally.
Make smart routing decisions.
Hand off seamlessly to existing self-service workflows.
Tool Types in Amazon Q in Connect
Actions are defined through tools. Amazon Q in Connect comes with the default system tools that you can use immediately. For more specialized tasks, Amazon Q in Connect default system tools can also be extended to create custom tools.
Default System Tools
Amazon Q in Connect comes with the following built-in tools:
QUESTION: Provides answers and gathers relevant information when no other tool can directly address the query
ESCALATION: Automatically transfers to an agent when customers request human assistance
CONVERSATION: Engages in basic dialog when there's no specific customer intent
COMPLETE: Concludes the interaction when customer needs are met
FOLLOW_UP_QUESTION: Enables more interactive and information-gathering conversations with customers. For more information about using this tool, see the Follow Up Question Tool tab.
To learn more about default system tools for Amazon Q in Connect, see the Use Generative AI-Powered Self-Service with Amazon Q in Connect section in the Amazon Connect Administrator Guide.
Follow Up Question Tool
The FOLLOW_UP_QUESTION tool enhances Amazon Q in Connect self-service capabilities by enabling more interactive and information-gathering conversations with customers. This tool works alongside the default and custom tools. It helps collect necessary information before determining which action to take.
The FOLLOW_UP_QUESTION tool complements your defined tools by enabling Amazon Q in Connect to gather necessary information before deciding which action to take. It's particularly useful for the following:
Intent disambiguation — When the customer's intent is unclear, use this tool to ask clarifying questions before selecting the appropriate action.
Information gathering — Collect required details for completing a task or answering a question.
To learn more about custom actions for Amazon Q in Connect, visit the FOLLOW_UP_QUESTION Tool section in the Amazon Connect Administrator Guide.
Custom Tools
Custom tools allow you to extend the Amazon Q in Connect capabilities by creating tools that handle specific tasks and can hand off to Amazon Lex bots relevant to your business.
You can customize a default system tool to be specific for checking order status, to ensure that it is configured to know what details you might need to accomplish this action.
For example, a customer wants to check their order status. You might want to collect specific information before finding their order status or passing to a deterministic bot that will continue to handle that self-service interaction. Amazon Q in Connect might give this response:
"I can help you check your order status. Could you please provide your order number or the email address used for the purchase?"
Typical Actions that custom tools are created for are as follows:
Appointment scheduling and rescheduling: Allow customers to book, view, or change appointments.
Order management: Enable checking order status, modifying orders, or initiating returns.
Account updates: Permit customers to update contact information or preferences.
Payment processing: Facilitate making payments or checking payment history.
To learn more about custom actions for Amazon Q in Connect, see the Custom Actions for Self-Service section in the Amazon Connect Administrator Guide.
Implementation Strategies
Implementation Strategy Best Practices
Successful implementation of Amazon Q in Connect requires thoughtful planning and a strategic approach. Consider the experience of a telecommunications provider who started their journey with a focused implementation strategy. They began by identifying their highest-volume customer inquiries: technical support for internet connectivity issues. By starting with this specific use case, they could refine their approach before expanding to more complex scenarios.
Complex Scenarios
When implementing Amazon Q in Connect for more sophisticated use cases, organizations should follow several key strategies to ensure success:
Start small and expand: Begin with low-complexity interactions or with deterministic self-service experiences that have higher volume to allow expansion.
Focus on customer journey integration: Ensure that Amazon Q in Connect is part of a cohesive customer experience.
Prioritize knowledge management: Invest time in organizing and optimizing your knowledge base.
Personalize when possible: Use customer attributes in flows with Amazon Q in Connect to personalize generated responses.
Monitor, learn, and adapt: Use analytics to identify areas for improvement.
By applying these strategies systematically, you can gradually expand the capabilities of Amazon Q in Connect while maintaining quality and reliability. The most successful implementations typically begin with specific use cases and grow organically based on performance data and customer feedback.
Key Business Impact Metrics
To truly understand the value that Amazon Q in Connect brings to your organization, you need to measure its performance. This measurement should use targeted metrics that align with your business objectives. Organizations that successfully implement these AI-powered solutions typically track the following indicators:
Containment rate: This is the percentage of inquiries fully resolved through self-service.
Cost per contact: Compare the cost of self-service interactions against agent-handled contacts.
Customer satisfaction: Measure satisfaction specifically for self-service interactions.
Agent efficiency: Track how agent handling times and job satisfaction improve.
Resolution speed: Compare how quickly issues are resolved through self-service instead of traditional methods.
A methodical approach to these metrics provides a clear picture of your return on investment. By establishing these measurements early in your implementation process, you can quantify success and identify opportunities for continuous optimization.
Agentic Self Service with Amazon Q in Connect
From Scripted Flows to Autonomous Agents
The self-service capabilities you have learned about so far, Amazon Polly for voice, Amazon Lex for intent recognition, and Amazon Q in Connect for generative AI responses, have evolved into something more powerful. Amazon Q in Connect now supports agentic self-service: AI agents that can reason through complex problems, remember context across a conversation, and take action to resolve customer requests end-to-end.
Previously, self-service systems followed a linear path: recognize intent, gather information, provide a response or transfer to an agent. Agentic self-service breaks this pattern. The AI agent can dynamically decide what steps to take, in what order, based on the customer's actual situation.
How agentic self-service works
Powered by Amazon Nova Sonic, AI agents interact with customers through natural voice conversations. The agent performs the following tasks:
Listens and understands the customer's request (including complex, multi-part problems)
Reasons about what information it needs and what actions to take
Uses tools via Model Context Protocol (MCP) to retrieve data and complete transactions
Responds naturally, adapting its approach based on what it discovers
This happens in real time, with the customer experiencing a fluid conversation rather than a series of prompts.
What this means for customers
Customers can describe their problem in their own words, without navigating menus or simplifying their request. The AI agent handles the complexity:
"I got charged twice for my order last week, and I also need to change my delivery address for the replacement." A single AI agent interaction resolves both issues.
"My internet has been slow since the storm. I've already rebooted the router." The agent skips troubleshooting steps the customer has already tried.
The role of Nova Sonic
Amazon Nova Sonic is the voice-native LLM that powers these conversations. Because it processes speech directly (rather than converting speech to text and back), interactions feel natural and responsive. Customers hear an expressive, conversational voice, not a robotic one.
Enhanced Capabilities
AI agents need access to information and systems to resolve customer requests. Several enhancements make this more powerful and flexible.
Multiple knowledge bases
You can now bring your own Amazon Bedrock Knowledge Bases and connect multiple knowledge bases to a single AI agent. This means an agent can draw from:
Product documentation
Policy and procedure guides
FAQ databases
Troubleshooting knowledge
Each knowledge base can be maintained independently, and the agent determines which source is most relevant for each query.
Model Context Protocol MCP tools
MCP provides a standardized way for AI agents to use tools, such as retrieving information, calling APIs, and completing actions in external systems. When a customer asks to reschedule a delivery, the agent uses MCP to connect to the scheduling system, check availability, and make the change.
This standardized protocol means new tools and integrations can be added without custom development for each one.
Stream messages
AI agent responses are now shown to customers as they are generated, rather than waiting for the complete response before displaying anything. This reduces perceived wait times, particularly for complex queries that require the agent to reason through multiple steps.
For voice interactions, this means the agent begins speaking as soon as it has enough context to respond, creating a more natural conversational cadence.
Automated self service evaluations
The system can automatically evaluate the quality of self-service interactions. It checks whether the AI agent resolved the request correctly, maintained appropriate tone, and followed organizational guidelines. This provides continuous quality monitoring without requiring supervisors to manually review every interaction.
AI Agent Performance Metrics
The Eight Metrics
The following eight metrics provide a complete picture of AI agent performance. They are grouped into three categories: outcome metrics, quality metrics, and customer feedback metrics.
Outcome metrics
These measure whether the AI agent accomplished what it set out to do:
Goal success rate — Did the AI agent successfully resolve the customer's request? This is the primary measure of agent effectiveness. A high goal success rate means customers are getting their problems solved without needing to escalate to a human agent.
Hand-off rate — How often does the AI agent transfer the interaction to a human? Some hand-offs are appropriate (complex situations requiring human judgment), but a high rate may indicate the agent needs tuning.
Quality metrics
Customer feedback metrics
These measure how well the AI agent performs its work:
Faithfulness score — Are the agent's responses grounded in factual, source-backed information? This metric detects contextual hallucinations — cases where the agent generates plausible but unsupported answers. A low faithfulness score is a serious quality issue.
Tool selection accuracy — Did the agent choose the right tools to accomplish the task? When an agent has access to multiple tools via MCP (order lookup, scheduling, refund processing), this metric evaluates whether it selected the appropriate one for each situation.
Conversation turns — How many exchanges does it take to resolve an issue? Fewer turns generally indicate a more efficient agent, though complex issues naturally require more.
Tool use accuracy — Did the agent use the selected tools correctly with proper parameters? This measures execution quality after the correct tool is chosen.
Quality metrics
These measure how well the AI agent performs its work:
Customer feedback metrics
These capture the customer's direct experience:
Invocation success rate — The percentage of AI agent invocations that complete without error.
Completeness score — Did the AI agent fully address all aspects of the customer's request?
Faithfulness Scoring in Depth
Faithfulness scoring deserves closer attention because it addresses one of the most significant risks of autonomous AI agents: hallucination.
When a customer asks, "What's your return policy for electronics?" the AI agent should respond with information from the organization's actual return policy documentation — not generate a plausible-sounding policy from its training data.
Faithfulness scoring evaluates each response against the source material available to the agent. It answers the question: "Is this response supported by the agent's knowledge bases and retrieved data?"
How it works
The system compares the AI agent's response against:
The documents retrieved from connected knowledge bases
Data returned by tool calls (order details, account information)
The conversation context established by the customer
Each response receives a score indicating how well it is grounded in these sources. Responses that contain claims not present in any source material score lower.
What to do with low scores
Low faithfulness scores indicate the agent is generating unsupported information. Actions to take:
Review the specific interactions flagged with low scores
Check whether the knowledge base contains the information the agent should have used
Identify gaps in documentation that force the agent to extrapolate
Adjust agent guardrails to reduce hallucination in specific topic areas
Setting thresholds
Organizations can define minimum acceptable faithfulness scores. Interactions falling below the threshold can be:
Flagged for human review
Automatically escalated to a live agent
Logged for agent improvement training
Accessing the Metrics
These metrics are accessible through multiple channels depending on your role and needs.
Access method
Best for
Capabilities
AI Agent Performance dashboard
Supervisors, day-to-day monitoring
Visual trends, real-time scores, alerts,filtering by agent name, escalation status, or intent
Contact details page
Investigators, issue-level analysis
Step-by-step traces showing each reasoning step, model used, latency, tool parameter inputs, and model output
GetMetricDataV2 API
Developers, custom reporting
Programmatic access, integration with external tools
The dashboard provides the quickest path to understanding agent performance. You can filter by agent name, escalation status, or intent to narrow down specific performance patterns. For deeper investigation of individual interactions, the contact details page allows you to drill into step-by-step traces that show how the AI agent reasoned, which tools it selected, and the latency of each step during self-service voice interactions. The API enables automation — for example, triggering alerts when faithfulness scores drop below a threshold. The data lake supports deeper analysis for teams optimizing agent behavior over time.
RECALL & APPLY
Practice what you learned
Try answering in your own words, then compare with the suggested answer. These activities are for self-study and are not automatically graded.
01What makes a request suitable for an initial self service deployment?
Topic: AI self service in the contact center
Reveal suggested answer
A clear customer need, dependable information or action, manageable exceptions, and a measurable definition of success.
02What is the difference between recognizing words and completing a request?
Topic: IVR and conversational self service
Reveal suggested answer
Recognition produces an interpretation. Completion also requires data collection, validation, fulfillment, and handling of the result.
03Which capability answers an open ended policy question?
Topic: Polly, Lex, and Amazon Q roles
Reveal suggested answer
Knowledge retrieval and generation can support the answer. The workflow must use the approved policy content and handle uncertainty appropriately.
04Why would an account balance prompt use text to speech?
Topic: Amazon Polly prompts
Reveal suggested answer
The value changes by customer and time. Dynamic text can provide the current value without recording every possible message.
05Why can a long written paragraph make a poor voice prompt?
Topic: Designing prompts for listening
Reveal suggested answer
Listeners cannot scan the text or easily revisit an earlier point. Spoken prompts must manage sequence, memory, and interruptions.
06What would you test when a product code sounds wrong?
Topic: Voice selection and SSML
Reveal suggested answer
Test the text and supported SSML interpretation, including whether characters should be spoken individually and whether the selected voice supports the chosen tags.
07In an appointment bot, what is the difference between the intent and a slot?
Topic: Amazon Lex core concepts
Reveal suggested answer
The intent is booking an appointment. Slots hold details such as date, time, location, or appointment type.
08Why use an alias instead of hardcoding a bot version?
Topic: Bot structure and deployment
Reveal suggested answer
An alias lets the integration select a published version through a stable reference. Version changes can then follow a controlled deployment process.
09Does collecting transfer details move money by itself?
Topic: Intents, slots, and confirmation
Reveal suggested answer
No. The conversation captures and confirms data. An authorized backend fulfillment integration must perform and confirm any real transfer.
10Why should a voice bot be tested through a real call path?
Topic: Bot testing and telephony input
Reveal suggested answer
Telephony audio, noise, timing, and caller behavior can expose issues that do not appear in typed tests or clean microphone recordings.
11A customer changes from a balance inquiry to a transfer. What should the design preserve?
Topic: Conversational interface principles
Reveal suggested answer
Preserve useful context, recognize the new goal, collect the required details, and confirm the consequential action.
12Should the same long answer be used unchanged in voice and chat?
Topic: Voice and text design
Reveal suggested answer
Usually not. Voice needs concise sequential delivery, while text can use short paragraphs, lists, and links where appropriate.
13What should be checked before adding the bot to a flow?
Topic: Building a Lex bot in Amazon Connect
Reveal suggested answer
Confirm that the language builds successfully, intent names are consistent, slots capture the intended data, and fallback behavior works.
14What happens if the flow’s intent name does not match the bot?
Topic: Connecting the bot to a flow
Reveal suggested answer
The intended routing can fail. Use the exact configured intent name and test the corresponding branch before deployment.
15What evidence proves that the conversational test succeeded?
Topic: Testing the complete experience
Reveal suggested answer
Correct intent recognition, captured values, confirmation behavior, and expected flow routing. A successful conversation does not prove a backend transaction occurred.
16Why should generated utterances still be reviewed?
Topic: Generative AI for Lex development
Reveal suggested answer
They may be irrelevant, ambiguous, or inconsistent with the intended task. Evaluate them against real customer language and neighboring intents.
17How does assisted slot resolution differ from building a new intent?
Topic: Generative Lex capabilities
Reveal suggested answer
It helps interpret a value for an existing slot. It does not by itself define a new business goal or fulfillment workflow.
18Which path fits retrieving a policy answer versus coordinating several tools?
Topic: Knowledge and agent integrations
Reveal suggested answer
Knowledge question answering fits retrieving and presenting approved information. Agent orchestration fits a task that requires multiple operations and context gathering.
19What does a text based model receive in a speech to text pipeline?
Topic: NLU and voice architecture choices
Reveal suggested answer
It receives the transcribed words and supplied context. It does not directly receive all acoustic information unless the architecture provides it.
20What would you compare when evaluating a speech integration?
Topic: Speech providers and language support
Reveal suggested answer
Use representative customer audio, relevant languages, latency, error recovery, operational requirements, and the supported integration path.
21What should transfer with the customer when self service cannot resolve the issue?
Topic: Amazon Q for customer self service
Reveal suggested answer
The relevant issue, information already gathered, actions attempted, and current state, according to the implemented handoff design.
22Why might one customer interaction use both services?
Topic: Lex and Amazon Q in a shared journey
Reveal suggested answer
A structured workflow can collect required details, while a knowledge assistant answers an unexpected question before the customer returns to the task.
23A customer says “send it back.” What should the assistant establish?
Topic: Natural customer conversations
Reveal suggested answer
Determine the item and intended action from context or a clarifying question. Do not assume the relevant order or policy.
24Why might an accurate document still retrieve poorly?
Topic: The knowledge foundation
Reveal suggested answer
Its structure, terminology, or missing explanatory text may make the relevant information hard to find or interpret.
25What should you do when the assistant repeatedly fails on the same question?
Topic: Responsible self service improvement
Reveal suggested answer
Review the affected conversations, locate the knowledge or configuration gap, make a targeted change, and retest the behavior.
26Why is asking a follow up question sometimes the best action?
Topic: Actions and system tools
Reveal suggested answer
The assistant may lack the information needed to select or execute a tool correctly. Clarification can prevent an incorrect operation.
27What information might an order status tool need?
Topic: Follow up questions and custom tools
Reveal suggested answer
An appropriate order or customer identifier, plus the access and validation required by the organization before returning customer specific information.
28What should a first deployment prove before it expands?
Topic: Implementation strategy and outcomes
Reveal suggested answer
It should reliably address the intended requests, handle exceptions, and meet agreed quality and operational goals.
29A customer reports a duplicate charge and an address change. What makes this more complex than a single FAQ?
Topic: Agentic self service
Reveal suggested answer
It involves separate goals, customer specific data, consequential actions, and a need to track which parts have actually completed.
30Does starting a response quickly prove the task is complete?
Topic: Knowledge, tools, and streamed responses
Reveal suggested answer
No. Response timing and successful task completion are separate. Confirm the tool result and whether the full request was addressed.
31What is the difference between selecting the right tool and using it correctly?
Topic: AI agent outcome and quality metrics
Reveal suggested answer
Selection chooses the appropriate capability. Correct use supplies valid parameters and handles its result properly.
32Can a faithful answer still be incomplete?
Topic: Reliability, completeness, and faithfulness
Reveal suggested answer
Yes. It can accurately reflect a source while failing to answer part of the customer’s request.
33What should a team avoid when a score falls below its threshold?
Topic: Investigating low faithfulness
Reveal suggested answer
Avoid treating the score alone as a complete diagnosis. Inspect the interaction, evidence, and configuration before choosing the correction.
34Choose one use case and describe its rollout plan.
Topic: Performance review and course application
Reveal suggested answer
A strong answer defines the customer need, data and knowledge, tools, human handoff, test cases, outcome metrics, and ongoing review owner.
Capture a key idea, an example, or a question for your instructor.
REFERENCE
A–Z topic index
Choose a topic to jump directly to its explanation in the eBook.
Imagine waking up to a coffee maker that predicts your needs, a self-driving car, and a phone that writes essays for you. This AI-powered world exists today.
In this course, you will explore artificial intelligence concepts that are provided in simple terms for everyone, regardless of technical background. You will discover how machines think and how AI is transforming the world around us.
By course end, you will grasp AI and machine learning basics, understand large language models like Amazon Nova, and recognize their opportunities and ethical questions. Whether you are exploring AI for career growth or simply to understand the world-changing technology, this course provides your essential foundation.
Course objectives
In this course, you will learn to do the following:
Define key artificial intelligence and machine learning (AI/ML) concepts using everyday language.
Identify different types of AI systems and explain how they function.
Recognize how large language models (LLMs) work and their practical applications.
Discover real-world AI applications across various industries.
Understand the ethical considerations that guide responsible AI development.
Introduction to the Foundations of AI and ML
Section objectives
In this section, you will learn to do the following:
Identify basic concepts of artificial intelligence and machine learning.
Identify the relationship between AI, machine learning (ML), and data science.
Have you ever wondered how Amazon Prime Video seems to know exactly what show you might want to watch next? Or how your email sorts of spam without you having to check each message? These everyday technologies use AI, ML, deep learning, and data science to make your life easier.
History and evolution of artificial intelligence
AI has had quite a journey over the decades. The following are some key moments.
The Birth of AI
The term, artificial intelligence, was coined in 1956 at a conference at Dartmouth College. Early researchers were optimistic they could build machines with human-like intelligence within a generation.
Reality Sets In
Researchers discovered that creating true intelligence was much harder than expected. Funding decreased during what became known as the AI Winter.
The Return
AI made a comeback with practical applications in specific domains. IBM's Deep Blue chess computer defeated world champion Garry Kasparov in 1997.
Big Data and Deep Learning
The explosion of internet data and increased computing power led to breakthroughs. Machine learning algorithms improved dramatically, especially deep learning methods.
AI Goes Mainstream
AI has become part of everyday life, from digital assistants like Alexa to recommendation systems on streaming services. Language models can write human-like text, and computer vision systems can recognize objects better than humans in some cases.
Differences Between AI and ML
The following sections distinguish artificial intelligence from machine learning.
Artificial Intelligence
Artificial Intelligence (AI) is teaching computers to perform tasks that typically require human intelligence.
Imagine if you could train your computer to recognize your friends in photos, understand questions when spoken aloud, or drive a car. That is AI in action.
At its core, AI is about creating computer systems that can perform the following:
Solve problems
Learn from experience
Understand natural language
Recognize patterns
Make decisions
Machine learning
Machine Learning (ML) is a specific approach to creating AI. Instead of programming explicit instructions for every situation, you give the computer examples and let it figure out the patterns on its own.
For example, teaching a computer to recognize dogs in pictures. With traditional programming, you would have to write specific rules like, if it has four legs, fur, and a tail, it is a dog. That approach falls apart quickly because there are dogs without tails. The scenario also didn't consider cats, which share features like four legs, fur, and a tail.
With machine learning, the computer analyzes thousands of pictures labelled dog and not dog. The computer then identifies patterns in these images and builds its own rules for recognizing dogs. When shown a new image, it can apply these learned patterns to decide if it's looking at a dog.
Relationship Between AI, ML, Deep Learning, and Data Science
The following definitions explain the relationship between AI, machine learning, deep learning, and data science.
Figure 1 Relationship Between AI, ML, Deep Learning, and Data ScienceSelect image to enlarge
Artificial intelligence
AI is a technology with human-like problem-solving capabilities. Organizations use AI to analyze data to enhance operations.
Machine learning
Machine learning is a type of artificial intelligence that performs data analysis tasks without explicit instructions. Machine learning technology can process large quantities of historical data, identify patterns, and predict new relationships between previously unknown data.
Deep learning
Deep learning is an AI method that teaches computers to process data in a way inspired by the human brain. You can use deep learning methods to automate tasks that typically require human intelligence, such as describing images or transcribing a sound file into text.
Data science
Data science is the study of data to extract meaningful insights for business. This multidisciplinary approach combines principles and practices from mathematics, statistics, and AI. It also incorporates computer engineering to analyze large amounts of data. This helps data scientists to ask and answer questions like what happened, why things happened, and what will happen in the future.
RECALL & APPLY
Practice what you learned
Try answering in your own words, then compare with the suggested answer. These activities are for self-study and are not automatically graded.
01Which part of your customer journey would benefit most from AI support?
Topic: Course scope and learning outcomes
Reveal suggested answer
Accept a specific operational problem and a measurable outcome. Revisit the example as the course progresses.
02Why did larger datasets matter to AI development?
Topic: AI history and development
Reveal suggested answer
More varied examples helped models learn patterns that were difficult to encode as individual rules.
03Is every AI application a machine learning application?
Topic: AI and machine learning
Reveal suggested answer
No. A system can use explicitly programmed rules. Machine learning is one approach within the broader AI field.
Capture a key idea, an example, or a question for your instructor.
Before exploring AI solutions for Amazon Connect Customer, review the fundamental AI terminology. These concepts form foundational understanding of AI systems that enhance customer service platforms. This knowledge helps you understand how artificial intelligence works in contact center environments.
Similar to how customer service representatives need to learn specific job skills, AI systems require certain components to function effectively.
Figure 2 AlgorithmSelect image to enlarge
Algorithm
An algorithm is simply a set of steps to solve a problem, like a recipe for computers. We all use algorithms in everyday life without calling them that.
Figure 3 ModelSelect image to enlarge
Model
A model is what you get after an AI system learns from data. Think of it like a student's brain after studying for a test. The student (or AI) has created an internal understanding that can now be applied to new situations.
Figure 4 TrainingSelect image to enlarge
Training
Training is the process of teaching an AI system by showing it examples. The training is similar to how you might learn to identify birds by looking at different types with a guidebook.
Figure 5 InferenceSelect image to enlarge
Inference
Inference is when an AI system applies what it has learned to new situations. The inference is like taking what you learned in driving lessons and using it to drive on a road you have never been on before.
Figure 6 DatasetSelect image to enlarge
Dataset
A dataset is simply a collection of information used to train or test an AI. Think of the dataset like a textbook full of examples for the AI to learn from.
By understanding these basic AI terms, you can develop the foundation needed to understand AI capabilities. These fundamentals work together to create systems that can enhance customer interactions and streamline contact center operations. With this terminology in mind, you are better equipped to understand how AI solutions are built and deployed.
Role of Data in Machine Learning
What Is the Role of Data in Machine Learning
In machine learning, data is everything. It is the foundation that makes learning possible. Without data, AI systems would have nothing to learn from.
Clean, diverse data teaches AI better than massive amounts of similar examples. AI systems need accurate information, just like a student needs reliable textbooks for learning.
Every AI task requires specific data types. Speech recognition needs audio samples. Image recognition demands diverse pictures. Each job has unique data requirements.
Understanding how machines learn from information
Although the advanced math can be complex, the basic process is something everyone can understand. The following explains the AI learning process:
Start with a task: First, define what the AI must learn. This might be recognizing faces in photos, translating languages, or predicting which movies someone might like.
Gather example data: Next, collect examples related to this task. If you are teaching an AI to recognize apples, gather thousands of apple images and images of things that aren't apples.
Feature recognition: The AI examines the examples to identify features or patterns. For apple recognition, it might notice characteristics like round shapes, red or green colors, and particular textures.
Pattern detection: As the AI looks at more examples, it starts to recognize patterns that differentiate categories. It might learn that apples are typically round, but so are oranges, so roundness alone isn't enough to identify an apple.
Trial and error: The AI makes predictions based on what it's learned so far and checks if it's right. If it misidentifies a red ball as an apple, it adjusts its understanding.
Refinement: With each attempt, the AI adjusts its understanding to improve accuracy. This might involve giving more weight to certain features (color and texture) and less to others (exact size).
Insider tip: Garbage in, garbage out is a fundamental principle in machine learning. Low quality data causes low quality results.
Remember: Machines do not truly understand what they are learning in the way humans do.
When an AI learns to identify pictures of cats, it does not know what a cat actually is, it just recognizes patterns of pixels that humans have labelled as cats.
RECALL & APPLY
Practice what you learned
Try answering in your own words, then compare with the suggested answer. These activities are for self-study and are not automatically graded.
01A contact classifier labels a new chat. Is this training or inference?
Topic: Core AI vocabulary
Reveal suggested answer
Inference. The application is applying an existing trained model to a new example.
02What happens if training examples exclude a common customer accent?
Topic: Data quality and model learning
Reveal suggested answer
The model may perform less well for that group. The dataset and evaluation examples need suitable representation.
Capture a key idea, an example, or a question for your instructor.
Have you ever wondered why a virtual assistant can tell you the weather but struggles to understand sarcasm in your voice? Or why your photo app can recognize your friends' faces but cannot tell you what they are thinking? The answer lies in understanding the different categories of AI systems and their capabilities.
In this section, you will explore the various types of artificial intelligence. From simple systems that follow preset rules to sophisticated learning systems that can improve over time. You will also explore the current state of AI technology and what might be possible in the future. Understanding these categories will help you recognize the strengths and limitations of the AI you encounter every day.
Section objectives
In this section, you will learn to do the following:
Identify narrow AI and general AI systems.
Compare rule-based AI systems with learning-based systems.
Identify different machine learning approaches and their real-world applications.
Introduction
When discussing AI systems, you should understand narrow and general AI. This distinction explains why AI in Amazon Connect Customer has specific capabilities and limitations. Narrow AI excels at specialized tasks like customer interactions. General AI, seen in science fiction, would reason broadly like humans but does not exist yet.
By understanding this fundamental difference, you can set realistic expectations for what AI can accomplish in your contact center operations.
Narrow AI
Narrow AI performs specific tasks extremely well but only those particular jobs. These systems excel at their assigned functions but cannot apply their skills elsewhere.
Figure 7 Narrow AISelect image to enlarge
Narrow AI examples
Spam filters protect your inbox but can't summarize your important messages.
Translation tools convert your sentences between languages but miss cultural nuances.
Voice assistants tell you weather forecasts but get confused by philosophical questions.
Narrow AI is a specialized tool in your digital toolkit. Excellent for specific tasks, yet ineffective for everything else.
General AI
General AI refers to systems with human-like intelligence, such as being able to understand, learn, and apply knowledge across different domains.
Figure 8 General AISelect image to enlarge
General AI examples
Transfers knowledge between different fields.
Reasons about abstract concepts.
Understands context and nuance.
Sets its own goals and pursue them.
Responds with something resembling consciousness or self-awareness.
Here is the important part: true general AI doesn't exist yet.
RECALL & APPLY
Practice what you learned
Try answering in your own words, then compare with the suggested answer. These activities are for self-study and are not automatically graded.
01Why can a strong spam classifier still fail at a different task?
Topic: Narrow AI and general AI
Reveal suggested answer
Its learned patterns and interface address spam classification, not every possible reasoning or language task.
Capture a key idea, an example, or a question for your instructor.
Another important way to categorize AI systems is by how they make decisions. Do they follow preset rules, or do they learn patterns from data?
Rule Based Systems
Rule-based systems follow explicit if-then rules created by human programmers. These systems have the following characteristics:
Follow logical rules created by experts
Have predictable behavior
Cannot improve without humans changing their rules
Work well for problems with clear rules and limited variables
A classic example is a tax preparation program. The program doesn't learn new tax rules on its own. Humans must program new rules when tax laws change.
Learning Systems
Learning systems use data to develop their own rules and improve over time. These systems have the following characteristics:
Identify patterns in data without explicit programming
Can improve with more experience and data
Might discover unexpected connections
Require training instead of rule-writing
Example: Modern spam detection learns from user behavior. When you mark an email as spam, the system learns from that example and gets better at identifying similar emails in the future.
Compare the systems
Neither approach is universally superior, but each has strengths for different situations. The following table compares the strengths of each system.
Rule-based systems are better
Learning systems are better
Precision is crucial (like medical diagnosis or legal applications).
The problem is too complex for humans to define all the rules.
The problem domain has clear, unchanging rules.
The environment changes frequently.
Available data is limited.
Large amounts of data are available.
Mistakes would be costly or dangerous.
The system needs to personalize to individual users.
Both approaches have valuable applications. Rule-based systems are suitable for handling structured workflows with clear rules. Learning systems are more suited for adapting to evolving customer needs. Understanding these differences helps you choose the right AI approach for each contact center challenge.
Learning approaches
Figure 9 Learning ApproachesSelect image to enlarge
Introduction to learning approaches
Machine learning teaches computers to make predictions based on data. You train a model using an algorithm and example data. Then, your application uses this model to generate real-time predictions at scale..
Three main learning methods
Explore the fundamental approaches that power AI systems in contact centers and beyond, each with unique strengths for different customer service challenges.
Supervised learning
In supervised learning, you provide the AI system with labeled data. Labeled data is information that comes with the correct answers already attached, like flashcards where questions have answers on the back.
The system then learns to predict the correct output for new, unseen inputs. It's like learning with a teacher who provides practice problems and the answers. Real-world examples of supervised learning include the following:
Email spam filters
Medical diagnosis from images
Price prediction for homes
Facial recognition systems
Unsupervised learning
Unlike supervised learning, unsupervised learning does not start with labeled data. The system identifies patterns, structures, or relationships within the data on its own. Real-world examples of unsupervised learning include the following:
Customer segmentation for marketing
Recommendation systems, such as recommendations based on customers who bought similar items
Anomaly detection for fraud prevention
Topic discovery in document collections
Reinforcement learning
Unlike supervised learning, reinforcement learning works through trial and error. The AI agent learns by taking actions and receiving feedback through rewards or penalties. Through this process, it gradually improves its strategy to maximize long-term rewards.
Real-world examples of reinforcement learning include the following:
Game playing (like AWS DeepRacer)
Autonomous vehicles learning to navigate traffic
Industrial robotics learning optimal movement patterns
Energy management systems optimizing resource usage
Choosing a learning approach
To create successful AI implementation, it is crucial to understand which learning approach fits your specific contact center needs. Each method addresses different types of business challenges.
Supervised learning
Unsupervised learning
Reinforcement learning
Use supervised learning for the following situations:
You have labeled training data available.
Your problem has clear input-output pairs.
You need specific, predictable responses.
Your task involves classification or regression.
Supervised learning is best suited for customer sentiment analysis and call categorization. It requires historical data with known outcomes.
All three approaches solve different challenges. Choosing the right approach depends on your data and business goals.
Learning approach selection in practice
Supervised learning
Use supervised learning for the following situations:
You have labeled training data available.
Your problem has clear input-output pairs.
You need specific, predictable responses.
Your task involves classification or regression.
Unsupervised learning
Use unsupervised learning for the following situations:
You lack labeled data.
You want to discover hidden patterns.
You need to group similar items together.
You are exploring data without specific predictions in mind.
Unsupervised learning helps identify emerging customer concerns and conversation themes. It finds valuable insights when you don't know what you're looking for.
Reinforcement learning
Use reinforcement learning for the following situations:
Your problem involves sequential decision-making.
You can define clear reward signals.
Your agent needs to learn through trial and error.
You are building systems that interact with environments.
Reinforcement learning optimizes dynamic call routing and conversational flows. It improves through ongoing interactions with customers and systems.
All three approaches solve different challenges. Choosing the right approach depends on your data and business goals.
RECALL & APPLY
Practice what you learned
Try answering in your own words, then compare with the suggested answer. These activities are for self-study and are not automatically graded.
01Which approach suits a fixed eligibility rule?
Topic: Rules and learned patterns
Reveal suggested answer
An explicit rule is usually easier to inspect and enforce when the policy is precise and stable.
02Which method fits grouping similar contacts without predefined categories?
Topic: Machine learning approaches
Reveal suggested answer
Unsupervised learning, because the aim is to discover groups or structures in unlabeled data.
03How would you approach sentiment classification with labeled historical calls?
Topic: Selecting a learning approach
Reveal suggested answer
Start with a supervised learning formulation and evaluate performance on representative held out examples.
Capture a key idea, an example, or a question for your instructor.
AI systems vary dramatically in their complexity and capabilities. Some can only perform simple tasks within narrow domains, whereas others demonstrate impressive versatility across multiple applications. Understanding this spectrum helps set realistic expectations about what different AI technologies can do.
Simple AI Systems
Simple AI systems typically do the following:
Focus on a single, well-defined task.
Follow straightforward logic or learned patterns.
Have limited adaptability to new situations.
Require human intervention when facing unexpected scenarios.
The following are simple AI system examples:
Rule-based bots that follow predefined conversation flows
Basic recommendation systems that suggest products based on past purchases
Document classification systems that sort text into predefined categories
Simple voice recognition systems that respond to specific commands
These systems are like specialized tools. They do one job well but aren't flexible.
Complex AI Systems
Complex AI systems typically do the following:
Handle multiple related tasks or domains.
Combine multiple AI techniques and models.
Adapt to new situations within their domain.
Process and integrate different types of information.
The following are complex AI system examples:
Virtual assistants that understand context across conversations
Autonomous vehicles that navigate unpredictable environments
Advanced medical diagnostic systems that consider numerous factors
Large language models that can generate text across various topics and styles
These systems are more like versatile assistants. They can handle a range of situations within their area of expertise.
AI System Complexity
The following factors can contribute to AI system complexity:
Multimodal processing: Is the system effective at working with different types of data, such as text, images, and audio, simultaneously?
Memory and context: Does the system remember previous interactions and maintain context over time?
Adaptation capability: How well can it adjust to new situations without requiring retraining?
Integration of multiple models: Does the system combine several specialized AI models to achieve its goals?
Learning approach: Does it use multiple learning approaches together, such as supervised, unsupervised, reinforcement?
Consider that even the most complex AI systems today are still fundamentally narrow AI. They are just very sophisticated within their domains. The apparent intelligence of these systems comes from their specialized capabilities rather than general human-like understanding.
RECALL & APPLY
Practice what you learned
Try answering in your own words, then compare with the suggested answer. These activities are for self-study and are not automatically graded.
01Does adding several models remove the need for defined scope?
Topic: Simple and complex AI systems
Reveal suggested answer
No. The integrated system still needs a clear purpose, evaluation criteria, and an understood operating boundary.
Capture a key idea, an example, or a question for your instructor.
So far, you have explored narrow AI, general AI, rule-based systems, and learning systems. But a new category of AI has emerged that changes how we think about what AI can do in the workplace: agentic AI.
Unlike traditional AI systems that wait for instructions and respond to specific queries, agentic AI systems can reason about problems, remember context across interactions, and take independent action to achieve goals. Think of the difference between a calculator (you ask, it answers) and a skilled colleague (they understand your goal, figure out the steps, and get it done).
What Is Agentic AI
Agentic AI represents a fundamental shift in how AI systems operate. These systems go beyond simple input-output responses to demonstrate autonomous goal-directed behavior.
Reasoning
Agentic AI systems can break down complex problems into steps, evaluate options, and determine the best course of action — much like a human problem-solver would approach an unfamiliar situation.
For example, when a customer contacts support with a billing discrepancy, an agentic AI system doesn't just retrieve account information. It reasons through the issue: checking recent transactions, identifying the anomaly, determining the cause, and deciding on the appropriate resolution.
Memory
Agentic AI systems maintain context across interactions. They remember what happened earlier in a conversation, what was discussed in previous sessions, and what actions have already been taken.
This means a customer doesn't have to repeat their issue every time they interact with the system. The AI agent remembers the full history and picks up where things left off.
Action
Perhaps most importantly, agentic AI systems can take independent action. They can use tools, call APIs, update records, send messages, and complete tasks. They can not just suggest next steps, but actually run them.
An agentic AI system might look up a customer's order, check the shipping status, initiate a refund, send a confirmation email, and update the case record as part of resolving a single customer request.
Tool Use
Agentic AI systems interact with external tools and services through standardized protocols. The Model Context Protocol (MCP) is one such standard that enables AI agents to retrieve information from knowledge bases, query databases, and complete actions in external systems.
This tool-use capability is what transforms a language model from a conversational interface into a capable agent that can accomplish real work.
Humorphism
AWS uses the term "humorphism" for this design philosophy, where AI is designed to collaborate like a human teammate rather than operate like a conventional software tool.
In a traditional software model, you tell a tool exactly what to do, step by step. In the humorphism model, you describe what you need and the AI figures out how to accomplish it, similar to how you would delegate a task to a capable team member.
This distinction is central to understanding how Amazon Connect Customer has evolved. Rather than providing a set of tools for contact center agents to operate, Amazon Connect Customer now provides AI agents that work alongside human agents. These agents help with reasoning through problems, remembering context, and taking action independently.
How Agentic AI relates to other categories
Agentic AI sits alongside the categories you have already learned about. The following comparison shows where it fits.
Category
Characteristics
Example
Narrow AI
Excels at one specific task
A spam filter
General AI
Human-like reasoning across all domains
Does not yet exist
Rule-based
Follows explicit if-then rules
A tax preparation program
Learning-based
Improves from data over time
A recommendation engine
Agentic AI
Reasons, remembers, and acts toward goals
An AI agent that resolves customer issues end-to-end
Agentic AI systems are still narrow AI. They operate within specific domains. However, within those domains, they demonstrate a level of autonomy and capability that earlier narrow AI systems did not possess. They combine learning-based approaches (using LLMs for reasoning) with tool use and memory to achieve goal-directed behavior.
Understanding agentic AI is essential because it underpins the entire Amazon Connect portfolio of solutions you will explore throughout this badge readiness path.
RECALL & APPLY
Practice what you learned
Try answering in your own words, then compare with the suggested answer. These activities are for self-study and are not automatically graded.
01What turns a language response into an action?
Topic: Agentic AI capabilities
Reveal suggested answer
A tool or API performs an operation outside the model. The surrounding application controls permissions and execution.
02A refund assistant can explain policy and issue a refund. What must the design distinguish?
Topic: Agentic AI in the wider AI landscape
Reveal suggested answer
It must distinguish answering a question from executing a transaction, including authority, confirmation, and failure handling.
Capture a key idea, an example, or a question for your instructor.
Have you ever asked Amazon Alexa a question, used Amazon Bedrock to generate content, or seen Amazon Lex power a chatbot conversation? These advanced conversational AI tools are increasingly powered by large language models (LLMs), one of the most rapidly evolving AI technologies today. These systems can write essays, translate languages, create poetry, answer complex questions, and even write computer code. These AI tools can do all of this by predicting which text should come next in a sequence.
In this section, you will explore what makes LLMs work, such as the special terminology used to describe them, their capabilities, and important limitations. You will learn how these systems differ from other types of AI. You will also discover how phrasing your requests (prompts) dramatically affects the results.
Section objectives
In this section, you will learn to do the following:
Compare LLMs and traditional AI.
Identify the key terminology associated with LLMs and their functions.
Identify the capabilities and limitations of modern language models.
Identify how context and prompt engineering affect LLM performance.
Key Terminology, Capabilities, and Limitations
Key Terminology
When reading about LLMs, you will encounter several specialized terms. Understanding this vocabulary will help you make sense of how these systems work and how to use them effectively.
Prompt and prompt engineering
A prompt is the input text given to an LLM to elicit a response. It can be a question, instruction, partial text, or any other text that tells the model what kind of output you want. Examples include the following:
"What are three ways to improve sleep quality?"
"Write a poem about autumn leaves in the style of Robert Frost."
"Translate this sentence to Spanish: 'I would like to order dinner.'"
Prompt engineering is the process you use to guide generative AI solutions to generate your desired outputs using techniques such as the following:
Chain-of-thought prompting
Few-shot learning examples
System prompts compared to user prompts
Temperature and other generation parameters
Why it matters: How you phrase your prompt dramatically affects the quality and nature of the response you get from an LLM.
To learn more about prompt engineering, see What is Prompt Engineering?
Tokens
Tokens are the basic units that LLMs process. They are essentially the pieces the model breaks text into. A token might be the following:
A full word (sunshine)
Part of a word (sun + shine)
A character (!)
A space between words
For example, the sentence, "I love artificial intelligence!" might be broken into tokens like ["I", "love", "artificial", "intelligence", "!"].
Why it matters: When using LLMs, you will often see limits expressed in tokens rather than words or characters. For English text, one token is roughly three fourths of a word on average.
Embeddings
Embeddings are numerical representations of words or tokens. They capture meaning by positioning similar concepts near each other in a mathematical space. For example, in this mathematical space:
Dog would be close to puppy and canine.
King minus man plus woman might equal something close to queen.
Paris would be near France just as Tokyo is near Japan.
Why it matters: Embeddings allow LLMs to understand relationships between words and concepts, enabling them to generate coherent and contextually appropriate text.
Context Window
The context window refers to how much text the model can see and consider when generating a response. It is the maximum amount of text (in tokens) that can be processed at one time. Examples include the following:
Early LLMs might have had a context window of 512 tokens (about a page of text).
Modern LLMs might have context windows of 8,000, 32,000, or even over 100,000 tokens.
Why it matters: A larger context window allows the model to reference information from much earlier in a conversation or document. This makes it more consistent and capable of handling longer texts.
Parameters
Parameters are the adjustable values that determine an LLM's behavior, learned during training. The number of parameters often indicates a model's capability and complexity.
Examples include the following:
GPT-3: 175 billion parameters
GPT-4: Estimated 1.76 trillion parameters (though the exact number isn't public)
Why it matters: Understanding parameter counts helps compare models and understand their capabilities and limitations.
Comparing Training and Inference
Training is the process of creating the model by exposing it to vast amounts of text data, allowing it to learn patterns. Inference is when the trained model is actually used to generate text or provide responses to user inputs.
Why it matters: The training process requires enormous computing resources but happens once. Inference (using the model) requires far less computing power. It occurs each time someone interacts with the system.
Fine tuning
Fine-tuning is the process of additional training on specific data to adapt a general LLM for particular purposes or to follow certain guidelines. For example, a general LLM might be fine-tuned to do the following:
Specialize in medical information.
Follow a company's brand voice.
Provide more concise responses.
Avoid certain topics or phrasings.
Why it matters: Fine-tuning is how general-purpose LLMs become specialized tools for specific applications, improving their performance for particular use cases.
Hallucinations
Hallucinations occur when LLMs generate plausible sounding but false or inaccurate information. This is one of the most significant challenges with current LLM technology.
Examples include the following:
Inventing fake citations or research papers
Creating fictional historical events
Generating incorrect technical specifications
Why it matters: Understanding hallucinations is crucial for responsible LLM use, especially in professional or educational contexts where accuracy is essential. Preventing hallucinations requires guardrails like fact-checking, confidence thresholds, human oversight, and regular model evaluations.
Capabilities of LLMs
LLMs are capable of producing large increases in productivity when applied correctly to a use case. Knowing the capabilities of your LLM will help determine suitability for the intended actions.
Generate human like text
LLMs can produce coherent, fluent text in various styles, from academic papers to poetry to business emails. They can adapt their writing style based on instructions.
Summarize information
LLMs can condense long documents or articles into shorter summaries while preserving key points.
Answer factual questions
For many common questions, especially about well-documented topics, LLMs can provide accurate information they have encountered in their training data.
Follow complex instructions
Modern LLMs can understand and execute multi-step instructions or tasks described in natural language.
Creative writing
LLMs can generate creative content like stories, poems, songs, and scripts based on provided prompts or themes.
Code generation and explanation
Many LLMs can write computer code in various programming languages and explain how existing code works.
Language translation
LLMs can translate text between languages they have been trained on, often with good accuracy for common language pairs.
Analyze sentiment and tone
LLMs can analyze text to determine whether it expresses positive, negative, or neutral sentiments.
Limitations of LLMs
The way you ask a question to LLMs can dramatically change the quality of the answer you receive. Although AI assistants seem incredibly knowledgeable, they sometimes hallucinate information that is not true. Understanding limitations becomes a key consideration when deciding if an LLM is applicable for your use case.
Hallucinations
LLMs can confidently generate information that appears plausible but is actually incorrect or made up. They can create false references, invent statistics, or fabricate facts that sound convincing.
Limited knowledge cutoff
LLMs only know information included in their training data up to a certain date. They do not have access to real-time information or events that occurred after their training cutoff date unless they are connected to search tools.
No true understanding
Despite appearances, LLMs do not truly understand text the way humans do. They predict patterns rather than comprehend meaning, which leads to limitations in reasoning and common sense.
Difficulty with complex reasoning
Although they are improving, LLMs still struggle with multi-step logical reasoning, especially for math problems or situations requiring cause-and-effect analysis.
Bias and stereotyping
LLMs can reflect and amplify biases present in their training data, potentially producing outputs that contain stereotypes or unfair representations.
Contextual limitations
LLMs have a limited context window. They can only see a certain amount of text at once, causing them to forget or lose track of information in very long conversations.
No external verification
LLMs typically cannot check external sources to verify information unless specifically connected to other tools or the internet.
Lack of Self awareness
LLMs don't have true self-awareness or consciousness. They do not know what they know or recognize when they are making mistakes.
Hallucination example
This example shows a possible hallucination could occur in an LLM response.
User Query: "Tell me about the 2021 Mars landing of the Harmony spacecraft and its discovery of microbial fossils."
LLM Hallucinated Response: "The Harmony spacecraft landed on Mars in March 2021 as part of a joint NASA-ESA mission. Its rover, Concordia, made headlines when it discovered fossilized microorganisms in the Arcadia Planitia region. These 2.8 billion-year-old microfossils showed cellular structures similar to Earth's cyanobacteria but with unique Martian adaptations. Dr. Elena Matsuoka, the mission's lead scientist, called it 'definitive evidence of ancient life on Mars,' prompting a mission extension and the planned Harmony II follow-up mission set for 2026."
Explanation of the Hallucination: This response is entirely fabricated. There was no "Harmony spacecraft" that landed on Mars in 2021, no "Concordia rover," and no discovery of microbial fossils. The scientist mentioned and all details about the mission are fictional. This demonstrates how an LLM can confidently generate plausible but completely false information when responding to queries about non-existent events. To aid in the identification and reduction of hallucinations, Amazon Bedrock Guardrails may be used to reduce possible hallucinations. To learn more, visit the Amazon Bedrock Guardrails page.
RECALL & APPLY
Practice what you learned
Try answering in your own words, then compare with the suggested answer. These activities are for self-study and are not automatically graded.
01Why can a long conversation lose useful earlier details?
Topic: Large language model terminology
Reveal suggested answer
Earlier information may fall outside the context supplied to the model or receive insufficient attention. Applications must manage context deliberately.
02Does adding a document to a prompt retrain the model?
Topic: Context, parameters, and fine tuning
Reveal suggested answer
No. It supplies context for inference. Fine tuning changes model parameters through a separate training process.
03A fluent answer cites a policy that does not exist. What is the problem?
Topic: LLM capabilities and hallucinations
Reveal suggested answer
The answer is unsupported even though it sounds convincing. Check the actual policy source and correct the response.
04How should a support workflow handle an uncertain policy answer?
Topic: LLM limitations
Reveal suggested answer
Retrieve the approved source or involve a qualified agent. Do not treat confidence of wording as evidence of accuracy.
Capture a key idea, an example, or a question for your instructor.
The LLM landscape is evolving rapidly. Since you learned the fundamentals of large language models, several developments have changed how LLMs are deployed in real-world systems, particularly in contact centers. In this lesson, you learn about four key advances: voice-native models, standardized tool-use protocols, LLM-enhanced NLU, and the emergence of third-party model integration.
Amazon Nova Sonic
Traditional LLMs process text. Amazon Nova Sonic is different. It is a voice-native LLM designed from the ground up for natural, expressive voice interactions. Rather than converting speech to text, processing it, and converting back (which introduces latency and loses vocal nuance), Nova Sonic works directly with audio.
What makes it voice native
Nova Sonic processes speech directly as its input and output modality. This means it understands not just what someone says, but how they say it. Tone, pacing, emphasis, and emotion are all part of the model's understanding.
This eliminates the latency of traditional speech-to-text-to-LLM-to-text-to-speech pipelines and produces more natural conversational experiences.
Role in Amazon Connect Customer
Nova Sonic powers agentic self-service in Amazon Connect Customer. When a customer calls and interacts with an AI agent, Nova Sonic enables that agent to:
Understand spoken requests naturally, including complex or ambiguous phrasing
Respond with expressive, human-like speech in real time
Maintain conversational flow without awkward pauses or robotic tone
This is what makes the agentic self-service experience feel like talking to a capable colleague rather than navigating a menu system.
Compared to traditional approaches
Aspect
Traditional pipeline
Nova Sonic
Processing
Speech → text → LLM → text → speech
Speech → model → speech
Latency
Higher (multiple conversion steps)
Lower (direct processing)
Nuance
Lost in text conversion
Preserved in audio
Conversation feel
Robotic, transactional
Natural, expressive
Model Context Protocol
In the previous lesson on agentic AI, you learned that AI agents can use tools to take action. But how does an AI agent know which tools are available, what they do, and how to use them? The Model Context Protocol (MCP) solves this problem.
MCP is a standardized protocol that enables AI agents to interact with external tools and data sources. Think of it as a universal adapter. Just as USB provides a standard way to connect devices to a computer, MCP provides a standard way to connect AI agents to the tools they need.
How MCP works
MCP defines a structured way for AI agents to:
Discover what tools are available (for example, a CRM lookup tool, a refund processing tool, or a knowledge base search tool)
Understand what each tool does and what inputs it requires
Invoke tools with the correct parameters
Interpret the results returned by those tools
This standardization means the same AI agent can work with many different tools without needing custom integration code for each one.
Why MCP matters for contact centers
In Amazon Connect Customer, MCP enables AI agents to:
Retrieve customer information from knowledge bases
Look up order status in backend systems
Process returns or schedule appointments
Access multiple data sources in a single interaction
Without MCP, each tool connection would require custom development. With MCP, new tools can be added and immediately used by AI agents through the standard protocol.
MCP in the broader environment
MCP is not proprietary to Amazon Connect Customer. It is an open standard gaining adoption across the AI industry. This means skills and patterns learned here apply broadly, and third-party tools built to the MCP standard can be integrated into Amazon Connect Customer AI agents.
Amazon Lex assisted NLU overview
Traditionally, Amazon Lex used slot-based natural language understanding (NLU) to interpret customer intent. This approach required defining specific intents, sample utterances, and slots for each conversation path. Now, Amazon Lex offers LLM-enhanced NLU through a feature called Assisted NLU.
Amazon Lex now offers large language models as an enhanced option for understanding customer intent through Assisted NLU. In Primary mode, the LLM serves as the default means of processing user input. In Fallback mode, the LLM activates only when the traditional NLU confidence score is below threshold. Both modes coexist with the slot-based system rather than replacing it. This represents a significant evolution in how conversational AI systems are built.
Traditional Slot-Based NLU
Assisted NLU (LLM-Powered)
Impact on development
Traditional Slot-Based NLU With traditional NLU, developers must:
Define each possible intent manually (for example, "CheckBalance", "TransferFunds", "ReportLostCard")
Provide sample utterances for each intent (dozens of example phrases)
Define slots for each piece of information needed (account number, amount, date)
Handle edge cases with explicit fallback logic
This approach works well for predictable interactions but struggles with unexpected phrasing, complex requests, or conversations that span multiple intents.
Third party AI integration overview
Amazon Connect Customer no longer relies exclusively on AWS-native AI models. Third-party speech and voice AI providers can now be integrated alongside native services, giving organizations more choice in how they build voice experiences.
Two notable integrations are now available for self-service interactions:
Deepgram for speech-to-text (STT) — known for high accuracy and speed in transcribing spoken language
ElevenLabs for text-to-speech (TTS) — known for highly natural, expressive voice synthesis
This multi-model approach means organizations can select the best AI model for each specific task rather than being limited to a single provider. For example, a contact center might use ElevenLabs for its premium voice quality in customer-facing interactions while using Amazon Polly for internal notifications where cost efficiency is prioritized.
The ability to mix native AWS models with third-party providers reflects the broader industry trend toward composable AI architectures. These are systems built from best-of-breed components rather than monolithic platforms.
Amazon Lex Assisted NLU
Traditional Slot Based NLU
Traditional Slot-Based NLU With traditional NLU, developers must:
Define each possible intent manually (for example, "CheckBalance", "TransferFunds", "ReportLostCard")
Provide sample utterances for each intent (dozens of example phrases)
Define slots for each piece of information needed (account number, amount, date)
Handle edge cases with explicit fallback logic
Assisted NLU LLM Powered
Assisted NLU (LLM-Powered) With Assisted NLU enabled in Amazon Lex:
The LLM understands intent from natural language without needing exhaustive sample utterances
Complex, multi-part requests are interpreted correctly ("I want to check my balance and also dispute the charge from yesterday")
Unexpected phrasing is handled gracefully
Context from earlier in the conversation informs understanding
This dramatically reduces the development effort required to build conversational experiences and improves accuracy for real-world customer language. Amazon Lex remains the conversational AI service — the LLM is the engine powering its understanding.
Impact on development
Aspect
Slot-based NLU
LLM-powered NLU
Setup effort
High (define all intents, utterances, slots)
Lower (model generalizes from fewer examples)
Handling unexpected input
Poor (falls back to error)
Strong (interprets from context)
Multi-intent requests
Requires complex flow design
Handled naturally
Maintenance
Manual updates for new phrases
Adapts with minimal tuning
Third party AI model integration
Two notable integrations are now available for self-service interactions:
Deepgram for speech-to-text (STT) — known for high accuracy and speed in transcribing spoken language
ElevenLabs for text-to-speech (TTS) — known for highly natural, expressive voice synthesis
RECALL & APPLY
Practice what you learned
Try answering in your own words, then compare with the suggested answer. These activities are for self-study and are not automatically graded.
01What should you assess beyond the sound of a generated voice?
Topic: Amazon Nova Sonic and voice interaction
Reveal suggested answer
Assess response relevance, latency, interruption handling, task completion, and the behavior when the system cannot help.
02Does a standard tool interface make every action appropriate?
Topic: Model Context Protocol and tool access
Reveal suggested answer
No. The application still needs authorization rules, suitable input validation, and review of consequential actions.
03What should a test set include beyond the designer’s sample utterances?
Topic: Assisted NLU and model integration
Reveal suggested answer
Paraphrases, ambiguous requests, corrections, incomplete information, and unsupported requests.
04What stays important when an LLM improves intent recognition?
Topic: Traditional and assisted NLU
Reveal suggested answer
Required data collection, confirmation, business validation, and the action that fulfills the request.
Capture a key idea, an example, or a question for your instructor.
An LLM is an AI system trained on massive amounts of text data to recognize and produce human language. Think of an LLM as an extremely sophisticated text prediction system. It's similar to the autocomplete feature on your phone, but at a much larger scale and with far greater capabilities.
Why do they call them large
The large in the name refers to the following:
The vast amount of text data used to train them (trillions of words)
The number of parameters they contain (billions or trillions)
The extensive computing power required to build them
Understanding the purpose of LLMs
The primary purpose of LLMs is to understand and generate human language, allowing them to do the following:
Answer questions and provide information.
Create written content in various formats and styles.
Translate between languages.
Summarize long documents.
Engage in conversational interactions.
Assist with writing and editing tasks.
Generate creative content like stories or poetry.
How LLMs work
At a basic level, LLMs work by predicting the next word or token in a sequence based on all the previous words. Through their training, they have learned patterns and relationships between words and concepts across billions of examples. For example, if you type "The capital of Spain is," an LLM would predict "Madrid" as the next word because it has seen this pattern many times in its training data. Modern LLMs become useful by maintaining prediction capabilities over extended text, enabling coherent generation of paragraphs and longer content.
Figure 11 How LLMs workSelect image to enlarge
How LLMs differ from traditional AI systems
The AI landscape has transformed dramatically with the emergence of LLMs. These powerful systems represent a fundamental shift from traditional AI approaches in several key ways. LLMs process information differently, learning from vast amounts of text rather than following explicit programming. Their ability to understand context and generate human-like responses sets them apart from their predecessors.
General language ability
Traditional AI systems are typically designed for specific, narrow tasks with predetermined functions and limited flexibility. These specialized systems excel at their designated purposes but cannot efficiently adapt to different contexts or tasks.
Examples of narrow-task traditional AI include the following:
Image recognition systems that identify objects in photos but cannot explain the content
Recommendation engines that suggest products based on user data but cannot justify their selections
Fraud detection systems that flag suspicious transactions without explaining their reasoning
In contrast, LLMs possess general language capabilities that enable them to perform across diverse domains without task-specific programming. A single LLM can do the following:
Craft creative content like poems, stories, or marketing copy.
Explain complex scientific concepts in accessible terms.
Generate professional communications, such as emails or reports.
Translate between languages while preserving meaning.
Answer questions across numerous knowledge domains.
This versatility stems from LLMs' foundational understanding of language patterns and contextual relationships rather than from specialized programming for each individual task.
Learning approach
Traditional AI systems often require task-specific training data and features explicitly defined by humans. LLMs use a different approach with pre-training on vast general text data followed by fine-tuning. Traditional AI systems often require the following:
Task-specific training data
Features explicitly defined by humans
Separate models for different tasks
LLMs use a different approach, such as the following:
Pre-training on vast general text data followed by fine-tuning
Self-learning of features and patterns from the data
Adapting one model to many tasks
Input and output flexibility
Traditional AI systems typically have structured inputs and outputs. LLMs are more flexible, accepting natural language instructions and generating varied outputs based on prompts.
Traditional AI systems typically have structured inputs and outputs, such as the following:
Predefined categories (for classification tasks)
Numerical predictions (for regression tasks)
Specific data formats required
LLMs are much more flexible because they can do the following:
Accept natural language instructions.
Generate varied outputs based on how they are prompted.
Adapt to different formats and styles of communication.
Context processing
Traditional AI often processes each input independently in the following ways:
Each image is classified separately.
Each transaction is evaluated on its own.
LLMs can maintain context over extended interactions by doing the following:
Remembering earlier parts of a conversation
Building on previously mentioned information
Maintaining consistency across a generated text
Trade offs Between Generative AI, LLMs, and Traditional AI
Although generative AI and LLMs offer impressive capabilities beyond traditional AI systems, important trade-offs exist.
Costs
LLMs require substantial computing resources for training and inference, making them more expensive to develop and deploy than many traditional AI solutions.
Task suitability
Traditional AI systems remain more efficient and reliable for specific structured tasks, especially those requiring precise, deterministic outcomes or operating in data-constrained environments.
Possibility of hallucinations
LLMs are prone to hallucinations and can generate plausible-sounding but incorrect information. This creates risks traditional rule-based systems typically don't face.
Interpretability
Traditional AI often offers clearer decision paths, whereas LLMs operate as complex black boxes that are difficult to audit.
Streamlined implementation
Existing traditional AI technologies offer proven, lightweight solutions for many business needs without the complexity of implementing cutting-edge LLMs.
Despite their challenges, LLMs offer tremendous value through unmatched natural language processing (NLP) and creative problem-solving capabilities. Their cross-domain adaptability replaces multiple specialized systems, and human-like interactions enhance user experiences. Ongoing advancements promise improved efficiency and reliability, making these trade-offs increasingly worthwhile for many applications.
RECALL & APPLY
Practice what you learned
Try answering in your own words, then compare with the suggested answer. These activities are for self-study and are not automatically graded.
01Why can one language model support both summarization and translation?
Topic: What an LLM does
Reveal suggested answer
Training captures language patterns that can be applied through different prompts, though quality still varies by task and language.
02When might a specialized model be preferable?
Topic: LLMs and traditional AI
Reveal suggested answer
When the task has a narrow output, strong accuracy requirements, or resource constraints that a specialized solution can meet more directly.
03What evidence should support choosing an LLM for a workflow?
Topic: AI solution tradeoffs
Reveal suggested answer
A relevant evaluation comparing output quality, operational cost, latency, and failure behavior against the available alternatives.
Capture a key idea, an example, or a question for your instructor.
Context in LLMs refers to all the text the model can see when generating a response, including the following:
Your current query or instruction
Previous messages in a conversation
Any additional information you have provided
System instructions (often invisible to users)
The model can only respond based on this visible context. It cannot recall previous conversations unless you explicitly refer to them or they are still within the context window.
Understanding prompt engineering
Prompt engineering is the practice of crafting inputs to get the best possible outputs from language models. It is essentially learning how to talk to AI systems effectively. The way you phrase your requests to an LLM, known as prompt engineering, dramatically affects the quality, accuracy, and usefulness of the responses you receive.
The following are basic prompt engineering techniques.
Be specific and clear
Specific, clear prompts eliminate ambiguity, producing desired AI outputs efficiently while reducing the need for repeated refinements. Review the following example:
Vague: "Tell me about planets."
Specific: "Explain the key differences between rocky and gas-giant planets in our solar system in basic terms for a 10-year-old."
Provide examples Few shot learning
Showing examples helps the model understand the pattern you want. Review the following example:
Convert these sentences to past tense:
"I walk to the store" becomes "I walked to the store."
"She runs quickly" becomes "She ran quickly."
"They build a house" becomes "They built a house."
Specify format and length
Tell the model how you want information structured. Review the following example:
Create a three-column table comparing apples, oranges, and bananas, based on the following:
1) Nutritional benefits
2) Growing conditions
3) Common varieties Keep each cell to 15 words or fewer.
Use role prompting
Asking the model to adopt a perspective or role can shape its response. Review the following example:
As an experienced math teacher helping a struggling eighth grader, explain how to solve for x in the equation 3x + 7 = 22.
Advanced prompt engineering techniques
Advanced techniques move beyond basic instructions to use deeper understanding of how language models process information and generate responses. These techniques can help you achieve more nuanced, accurate, and contextually appropriate outputs while reducing common issues like hallucinations or inconsistencies.
Chain-of-thought prompting
System instructions
Controlling creativity and precision
Handling hallucinations
Common prompt engineering mistakes
Even experienced users can fall into common pitfalls when crafting prompts for AI systems. Understanding these frequent mistakes is crucial for improving your prompt engineering skills and getting more reliable results. By learning to recognize and avoid these common errors, you can save time, reduce frustration, and achieve more consistent outcomes in your interactions with AI.
The following are the most prevalent mistakes and how to address them:
Being too vague: Vague prompts lead to generic responses. Add specificity about audience, purpose, tone, and format.
Overloading with instructions: Too many contradictory or complex requirements can confuse the model. Focus on the most important instructions.
Not iterating: Prompt engineering often requires refinement. Adjust your prompt and try again when responses are unsatisfactory.
Forgetting about context limitations: Remember that models can only see a limited amount of text. Very long conversations might lose important context from earlier messages.
Advanced prompt engineering examples
Stepwise problem solving
Encourage the model to show its reasoning process in the following way:
Question: A shirt costs $25. If it's discounted by 20 percent and then there's an additional 10 percent off, what is the final price?
Think through this step by step.
System instructions
Many LLM applications allow setting system-level instructions that shape all responses. The following is an example:
You are a helpful assistant that specializes in explaining scientific concepts using everyday analogies. Keep explanations under 100 words.
Controlling creativity and precision
Some systems allow adjusting temperature settings, such as the following:
Higher temperature: More creative and varied, but potentially less accurate responses
Lower temperature: More predictable and conservative, and often more factual responses
Handling hallucinations
When accuracy is critical, you can add instructions like the following:
If you're unsure of any information, please explicitly state that you don't know rather than guessing.
RECALL & APPLY
Practice what you learned
Try answering in your own words, then compare with the suggested answer. These activities are for self-study and are not automatically graded.
01Why does “help this customer” produce inconsistent answers?
Topic: Context and clear prompts
Reveal suggested answer
The instruction omits the customer’s issue, approved information, the assistant’s role, and the required outcome.
02How would you improve “summarize this call”?
Topic: Examples and output requirements
Reveal suggested answer
Specify the audience and fields, such as customer issue, actions taken, outcome, and next step. Provide the transcript and any required length limit.
Have you ever wondered how Amazon Music seems to know exactly what podcast you might enjoy listening to next? Or how recipes are shown on your Alexa that you might want to cook? These everyday conveniences are powered by AI and ML technologies that have quietly transformed how businesses operate and how we interact with digital services.
Section objectives
In this section, you will learn to do the following:
Identify how AI and ML are transforming business decision-making processes.
Describe natural language processing applications in everyday use.
Describe how AI enables creative content generation across different media.
Identify automation and virtual assistant applications across industries.
Business Decision Making, Virtual Assistants, and Recommendations
AI Powered Business Intelligence
Businesses today face a flood of data, such as sales figures, customer behavior, market trends, and inventory levels. Making sense of all this information would be difficult for humans alone. That is where AI-powered business intelligence comes in.
AI transforms business decisions by analyzing vast datasets to reveal hidden patterns and trends. Companies use these insights to optimize operations and predict market changes with unprecedented accuracy. AI systems help businesses turn raw data into actionable insights through several key capabilities.
Data analysis at scale
AI systems help businesses turn raw data into actionable insights by analyzing massive datasets far beyond what human analysts could process manually.
These systems can identify patterns across millions of transactions or customer interactions, detecting trends that would be impossible to spot through manual analysis.
Predictive analytics
AI does not just analyze past data; it can predict future trends. The following are examples of predictive analytics:
Sales forecasting: Based on current patterns, we expect a 12 percent increase in winter coat sales in the Northeast region.
Customer behavior: Customers who purchase this product typically return within 45 days to buy these complementary items.
Risk assessment: This loan applicant has an 85 percent probability of repayment based on our model.
Automated reporting and visualization
AI systems can automatically generate reports and visual representations that make complex data understandable.
These tools can create custom dashboards that update in real-time, highlighting the most relevant information for different stakeholders and adapting to changing business conditions.
Real world examples
The following industry examples illustrate how AI systems help business:
Retail: Retailers use AI to optimize inventory levels, preventing both shortages and excess stock. This is done by analyzing historical sales data, seasonal trends, local events, weather forecasts, and social media buzz.
Healthcare: Hospitals use AI systems to improve operations and patient care by predicting patient admission rates to staff appropriately, optimizing surgery schedules, and analyzing treatment outcomes.
Financial services: Banks and investment firms rely on AI for numerous decision-support functions including algorithmic trading, credit scoring models, and anti-fraud systems.
Virtual Assistants and Conversational AI
Customer service bots are AI systems that handle inquiries through advanced language processing algorithms. These digital assistants analyze context and access databases to provide relevant responses to common questions. When faced with complex issues, they escalate to human agents while continuously learning from interactions.
Voice assistant request processing cycle
Audio input
The voice assistant activates and performs the following actions to capture the user's speech through a microphone:
The system records the raw audio waveform of your voice.
Background noise is filtered out using noise cancellation algorithms.
Voice Activity Detection (VAD) identifies when the user starts and stops speaking, so the system knows when a complete utterance has been captured.
Example: User says, "What's the weather like today?"
Speech to text
The audio recording is converted into written text with the following actions:
The audio is broken into small segments (usually 10-20 milliseconds each).
These segments are analyzed to identify phonemes (speech sounds).
Phoneme sequences are compared against language models to recognize words.
The system produces a text transcript of what was said.
Example: The audio is input as text, "What's the weather like today?"
Natural Language Processing
The system uses the following actions to analyze the text to understand its meaning and context:
The text is broken down into tokens (words and punctuation).
Part-of-speech tagging identifies nouns, verbs, adjectives, and so forth.
Dependency parsing determines relationships between words.
Named entity recognition identifies specific objects, places, or concepts.
The system builds a semantic representation of the user's request.
Example: The system recognizes that weather and today are the key entities.
Intent classification
The system uses the following actions to determine what the user is trying to accomplish:
Machine learning models classify the request into predefined categories.
The system identifies the user's goal or intent.
Confidence scores are calculated for different possible intents.
The highest-scoring intent is selected.
Example: The intent is classified as weather inquiry with location as current and time as today.
Response generation
The system creates an appropriate answer to the user's request with the following actions.
Based on the identified intent, the system retrieves necessary information.
For a weather request, it connects to a weather service API.
Data is gathered, such as temperature, conditions, and forecast.
A natural-sounding response is constructed following response templates.
Example: System generates, "The current temperature is 72 degrees with partly cloudy skies."
Text to speech
The text response is converted into spoken audio with the following actions.
Text is broken down into phonetic representations.
A voice synthesis model generates natural-sounding speech.
Prosody (rhythm, stress, intonation) is applied to sound more human.
The audio response is played through speakers.
Example: "The current temperature is 72 degrees with partly cloudy skies" is spoken aloud.
Speech to speech
The spoken input is converted directly into spoken output with the following actions:
Audio input is captured and converted into phonetic representations.
A voice conversion model processes the speech patterns and applies target voice characteristics.
Prosody (rhythm, stress, intonation) from the original speech is preserved or modified as needed.
The transformed audio is generated and played through speakers.
Example: A person says "Hello, how are you today?" in English, and the response is converted to sound like a different speaker. The response may be in a different accent while maintaining the original meaning and emotional tone.
Recommendation systems and personalization
If you have ever been pleasantly surprised by a Recommended for You suggestion that perfectly matched your tastes, you've experienced an AI recommendation system in action.
These systems work by using the following:
Collaborative filtering: This approach finds patterns based on user similarity, such as "Users who liked the same movies also enjoyed this one."
Content-based filtering: This approach analyzes the characteristics of items, such as "Because you enjoyed action movies with strong female leads, you might like this film."
Hybrid approaches: Most modern recommendation systems combine multiple methods for better results.
Some real-world applications of these systems include the following:
Streaming media: Streaming provider customers rely on content recommendations for their next binge set.
Ecommerce: The Frequently bought together suggestions on Amazon.com significantly impact purchasing decisions.
Social media: Platforms use AI to personalize content feeds based on your interactions.
News and content: News sites and apps personalize what stories appear first based on your interests.
Speech to speech
Speech-to-speech AI represents the next evolutionary leap, eliminating the text intermediary altogether. Instead of converting text to speech, these advanced systems can take spoken input in one language and convert it directly to spoken output in another language. The technology enables natural conversations across language barriers without requiring text intermediaries or causing significant delays.
For example, an executive can now speak to an international team in real time, preserving voice inflections while delivering the message in the team's language.
RECALL & APPLY
Practice what you learned
Try answering in your own words, then compare with the suggested answer. These activities are for self-study and are not automatically graded.
01How does a forecast differ from a guaranteed outcome?
Topic: AI in business decisions
Reveal suggested answer
A forecast is an estimate based on data and assumptions. Teams should inspect uncertainty and update their plans as evidence changes.
02At which step would a weather assistant obtain current weather data?
Topic: The voice assistant processing cycle
Reveal suggested answer
The application retrieves data after identifying the request. Speech recognition alone does not supply the weather information.
03Why might a recommendation system combine several methods?
Topic: Speech interaction and recommendations
Reveal suggested answer
Each method captures different evidence. Combining them can address missing data or improve relevance for different users and items.
Capture a key idea, an example, or a question for your instructor.
Earlier in this section, you explored how AI and ML are transforming business decision-making, powering virtual assistants, and enabling creative applications. Now you will learn how agentic AI is being applied across entire industries through the Amazon Connect portfolio.
In the first half of 2026, AWS announced the expansion of Amazon Connect Customer from a single contact center product into four agentic AI solutions, each purpose-built for a specific industry. These solutions demonstrate how the same underlying AI capabilities (reasoning, memory, action, and tool use) take different forms depending on the domain.
Four industry solutions
The following four solutions show how agentic AI adapts to specific industry needs. Each solution uses specialized AI agents trained for its domain.
Amazon Connect Health
Amazon Connect Health deploys five specialized AI agents for healthcare organizations:
Patient verification — confirms identity securely across channels
Appointment management — schedules, reschedules, and sends reminders
Patient insights — surfaces relevant medical history during interactions
Medical coding — assigns appropriate codes from clinical notes
Key characteristics:
HIPAA-eligible by design
Integrates with Electronic Health Record (EHR) systems
AI agents operate within strict compliance boundaries
This demonstrates how agentic AI can function in highly regulated environments where accuracy and privacy are non-negotiable.
Amazon Connect Decisions
Amazon Connect Decisions applies agentic AI to supply chain operations, combining:
30 years of Amazon's operational science
25+ specialized AI tools for logistics and planning
SCOT (Supply Chain Optimization Technologies) foundation models
AI agents in this solution can reason about complex supply chain scenarios such as inventory positioning, demand forecasting, and route optimization. They can recommend or take action based on real-time data.
This demonstrates how agentic AI handles problems with many variables, constraints, and interdependencies that would overwhelm traditional rule-based systems.
Amazon Connect Talent
Amazon Connect Talent applies agentic AI to the hiring process:
Every candidate receives the same fair, standardized experience
Scoring is based on competencies, not subjective impressions
This demonstrates how agentic AI can bring consistency and scale to processes that traditionally depended on individual human judgment. At the same time, it maintains the human-like quality of voice interaction through Nova Sonic.
Amazon Connect Customer
Amazon Connect Customer is the evolution of the original Amazon Connect contact center product, now rebranded and enhanced for agentic self-service.
At its core, Amazon Connect Customer enables AI agents powered by Nova Sonic to handle complex customer requests end-to-end — reasoning through multi-step problems, accessing tools, and resolving issues autonomously. To make building these experiences accessible, the platform includes a no-code conversational AI canvas (via NLX) that lets designers visually compose self-service flows without writing code. This combination of powerful AI reasoning and visual design tooling means organizations can deploy sophisticated conversational experiences rapidly.
This is the solution most directly relevant to the contact center use cases you will explore in detail throughout the remaining courses in this badge path.
Common patterns across solutions
Although each solution targets a different industry, they share common agentic AI characteristics.
Pattern
Description
Example
Specialized AI agents
Purpose-built agents for specific tasks
Health's five distinct agents
Domain knowledge
Models trained on industry-specific data
Decisions' SCOT foundation models
Compliance by design
AI operates within regulatory boundaries
Health's HIPAA eligibility
Human-like interaction
Natural voice via Nova Sonic
Talent's structured interviews
Autonomous action
Agents complete tasks end-to-end
Customer's self-service resolution
The pattern you should recognize is this: agentic AI is not one solution applied generically. It is a design philosophy built on reasoning, memory, and action. It takes different shapes depending on the domain's unique requirements, constraints, and data.
RECALL & APPLY
Practice what you learned
Try answering in your own words, then compare with the suggested answer. These activities are for self-study and are not automatically graded.
01Why is a generic agent insufficient for every industry?
Topic: Agentic AI across industries
Reveal suggested answer
Domain workflows, data access, acceptable errors, and evaluation requirements differ. The agent must be configured for its intended use.
02What do these examples have in common?
Topic: Amazon Connect portfolio examples
Reveal suggested answer
They apply agentic capabilities to a defined domain using relevant information, tools, and operating constraints. Discuss these as the course’s portfolio examples.
03What would you change when moving an agent from support to supply chain planning?
Topic: Shared patterns across agentic solutions
Reveal suggested answer
Change the domain information, tools, permissions, objectives, and test cases. Reusing the general architecture does not validate the new use case.
Capture a key idea, an example, or a question for your instructor.
AI's growing ability to generate creative content represents one of the most fascinating developments in modern technology. These systems now write stories, compose music, and create visual art that rivals human-made work. The boundaries between human and machine creativity continue to blur as AI design tools become more sophisticated and accessible.
Text generation
AI systems can now produce various types of written content, including the following:
Creative writing, such as stories, poetry, scripts, and lyrics
Practical writing, such as marketing copy, blog posts, and educational materials
Prompt: "Create me an example of text generation to showcase creative writing with a maximum of 100 words"
Output: In Eldergrove, clockmaker Orion crafted timepieces infused with emotions. A mysterious woman sought a clock to rewind a cherished moment, paying with a memory. He created a moonlit marvel. She vanished, leaving jasmine's scent. Though heartbroken, Orion found joy in his work. Tales of his magical clocks spread, immortalizing love's fleeting beauty. Time, in Eldergrove, was rewritten, not just measured.
Image generation
AI can create images based on text descriptions using technologies. These systems can produce photorealistic images of scenes that don't exist, artistic renderings in specific styles, product visualizations, and character designs.
Prompt: "Create me an artist impression of an Eldergrove from the provided story."
Video generation
AI video generation tools like Amazon Nova Reel can now create short video clips or even longer sequences based on text descriptions or image inputs. These systems can produce animated scenes, transform still images into motion, generate realistic human movements, and even create entire short films. Although still developing, this technology is rapidly advancing, offering new possibilities for filmmakers, marketers, and content creators.
Prompt: "Create me a video of the story from the attached image."
Conversational AI using speech to speech
AI now engages in natural voice conversations with real-time processing, maintaining context across multi-turn dialogues, and adapting speaking style to match user preferences. Voice-based systems handle interruptions, emotional nuances, and conversational flow while preserving the immediacy of spoken communication. The technology creates natural human-machine conversation through instantaneous voice processing, dynamic emotional expression, and fluid dialogue that mirrors human speech patterns.
This audio is a conversation between a fictitious user and Amazon Nova Sonic. Nova Sonic is a conversational AI model that processes speech-to-speech. The first voice represents the user asking a question and the second voice is Nova Sonic providing a response.
Transcript Output
User: Hi Nova Sonic, please tell me about yourself.
Nova Sonic: Hey there, I'm an AI system here to assist you with any questions or tasks you have. I think of me as your friendly tech buddy.
RECALL & APPLY
Practice what you learned
Try answering in your own words, then compare with the suggested answer. These activities are for self-study and are not automatically graded.
01What makes a useful creative generation prompt?
Topic: Creative AI applications
Reveal suggested answer
A clear purpose, audience, medium, style direction, and constraints. Evaluate the output against those requirements.
Capture a key idea, an example, or a question for your instructor.
AI is transforming physical systems and automation, enabling robots and machines to perform increasingly complex tasks with greater flexibility and less human intervention.
Review the following to learn more about the various ways robotics and automation are being applied.
Industrial robotics and manufacturing
Modern manufacturing uses AI-enhanced robots for the following:
Adaptive assembly: AI-powered robots can recognize part variations and adjust their movements.
Quality control: AI vision systems inspect products with greater accuracy than human inspectors.
Predictive maintenance: AI analyzes sensor data to predict equipment failures before they happen.
Example: Modern automotive manufacturing plants use AI-powered robots that can work collaboratively alongside human workers, adapting to different car models without reprogramming.
Warehouse and logistics automation
Ecommerce and logistics companies use AI-powered systems for the following:
Autonomous mobile robots (AMRs): Warehouse robots that navigate dynamically around obstacles and people
Intelligent sorting systems: AI-powered conveyor systems that identify and route packages
Last-mile delivery: Autonomous vehicles, drones, and sidewalk robots for final delivery
Example: Amazon employs more than half a million mobile robots in its fulfilment centers to help store, sort, and retrieve products.
Consumer and service robotics
The following are examples of how service robots aid humans in a variety of ways:
Home robots: Robot vacuums that map homes and optimize cleaning paths
Healthcare robotics: Surgical assistance robots with enhanced precision
Customer service robots: Hotel and retail robots that assist customers
Example: The da Vinci Surgical System provides surgeons with enhanced precision and control. It translates the surgeon's hand movements into smaller, more precise actions performed by miniature instruments inside the patient's body.
The future of robotics
Emerging developments in robotics include the following:
Advanced dexterity: Robots that can manipulate delicate or irregular objects
Social robotics: Robots designed for natural human-robot interaction
Field robotics: Autonomous systems for agriculture, construction, and exploration
Multi-robot coordination: Swarms of robots working together on complex tasks
Example: Modern autonomous robots demonstrate advanced mobility and adaptability, navigating complex terrain and performing dynamic tasks that were impossible for earlier robotics systems.
AI is revolutionizing robotics and automation across industries, from manufacturing floors to warehouses, homes, and hospitals. These applications represent much more than technological novelties. They are transforming business operations, enhancing human capabilities, and creating new possibilities for customer experiences.
RECALL & APPLY
Practice what you learned
Try answering in your own words, then compare with the suggested answer. These activities are for self-study and are not automatically graded.
01How does a physical robot’s failure differ from an incorrect text answer?
Topic: Robotics and automation
Reveal suggested answer
Physical actions can affect people and equipment directly. The design needs appropriate sensing, operating limits, and failure handling.
Capture a key idea, an example, or a question for your instructor.
Have you ever wondered if the algorithms that recommend products to you online might treat some customers differently? Or whether an AI making hiring suggestions might favor certain types of candidates? As AI becomes more powerful and widespread, important ethical questions arise about how these systems work and their impact on society.
In this section, you will explore the ethical considerations that come with AI development and deployment. You will examine issues of bias and fairness, and look at how AI systems affect our privacy. You will become familiar with concerns about LLMs and consider questions about aligning AI with human values. Understanding these ethical dimensions is crucial for everyone living in a world increasingly shaped by this technology.
Section objectives
In this section, you will learn how to:
Identify key ethical challenges related to bias and fairness in AI systems.
Identify fundamental data privacy concerns in AI development and use.
Identify specific ethical issues associated with LLMs.
Describe concepts of AI alignment, transparency, and social impact.
Bias and fairness
Bias and fairness
When a bank's AI approves loans unevenly across similar groups, is it perpetuating bias or detecting real patterns? When your voice assistant records conversations for improvement, who can access this data and where is it stored?
These ethical dilemmas emerge as AI becomes increasingly woven into our everyday experiences.
Bias and fairness concepts
AI systems can reflect, amplify, or even introduce biases that affect different groups of people unfairly. Biases include the following:
Training data bias: AI systems learn from historical data, which often contains existing societal biases. For example, a hiring algorithm trained on past hiring decisions might perpetuate gender or racial disparities.
Selection bias: The way data is selected or collected can skew results. For example, customer feedback collected only through smartphone apps excludes people without smartphones.
Measurement bias: The method you choose to measure success can create bias. For example, a content recommendation system optimized for engagement might promote controversial or inflammatory content.
Algorithm design bias: The choices made when designing AI systems can introduce bias, such as including certain features while excluding others, which can impact different groups differently. For example, an algorithm trained mostly on male patient data might miss certain diagnoses in women due to their different symptom patterns.
RECALL & APPLY
Practice what you learned
Try answering in your own words, then compare with the suggested answer. These activities are for self-study and are not automatically graded.
01A survey reaches only mobile app users. What bias could follow?
Topic: Ethics, bias, and fairness
Reveal suggested answer
Selection bias. Customers who do not use the app are missing, so the findings may not represent the full customer population.
Capture a key idea, an example, or a question for your instructor.
Elevate your customer engagement strategy with the power of AI in Amazon Connect. This comprehensive course explores innovative AI capabilities that transform how businesses interact with customers across their contact center operations. Discover how to create unified customer profiles with automated data mapping, resolve identity concerns, and design targeted outbound campaigns using natural language segment creation. You will learn to use AI for more personalized customer interactions, improved first-contact resolution, and more efficient campaign management.
In this course, you will learn to do the following:
Implement AI capabilities in Amazon Connect to enhance customer engagement across contact center operations.
Configure AI-powered Amazon Connect Customer Profiles and Identity Resolution to create comprehensive unified customer views.
Design targeted outbound campaigns using AI-driven segmentation and natural language tools.
Measure the business impact of AI-powered customer engagement solutions on operational efficiency and customer experience.
Introduction to AI Powered Customer Engagement
Section objectives
In this section, you will learn to do the following:
Compare traditional contact centers with AI-enhanced engagement models to identify key opportunities for transformation.
Implement AI-powered features to recognize customers across channels and deliver personalized experiences.
Evaluate how the partnership between AI and human agents improves key business metrics and customer satisfaction.
Section introduction
Have you ever called customer service and repeated your information multiple times to different agents? Or have you received generic marketing messages that had nothing to do with your needs?
In this section, you will explore how Amazon Connect AI capabilities are reshaping customer engagement and why these capabilities matter for your business. You will discover how AI helps businesses recognize customers across different channels. This helps businesses understand customer history and needs, so they can engage with customers in more relevant ways. Whether you are managing a support team or designing customer journeys, understanding these AI foundations is valuable. They will help you deliver experiences that feel personalized and efficient instead of robotic and frustrating.
Customer Data in the Contact Center
Traditional contact centers operate with a major constraint: fragmented customer information. Agents often need to juggle multiple systems to gather basic information about who they are talking to and why.
This can create a number of problems with the contact center experience, including the following:
Disjointed customer experience –When customer data lives in separate systems, agents spend valuable time switching between applications instead of helping customers. By the time the agent has gathered the necessary information, the customer is already frustrated.
Repetitive customer identification –One of the most frustrating experiences for customers is having to repeatedly identify themselves as they are transferred between departments or agents. This happens when contact centers lack a unified customer profile that travels with the customer throughout their journey.
Generic service instead of personalized engagement –Without a complete picture of the customer, agents default to generic service scripts instead of tailored interactions. They miss opportunities to reference previous purchases, anticipate needs, or acknowledge loyalty.
Real-world example: You recently moved and updated your address with a company's billing department. When you call customer services about a different issue, the agent asks you toconfirm your address. However, the agent still sees your old address because the billing system and customer services system are not sharing information.
Contact center evolution
Contact centers have evolved through several stages over the years as explained in the following:
1970s 1990s
Basic Call Centers The Phone Only Era
Phone-based support with minimal customer data.
Basic call centers revolutionized customer service by moving interactions from stores to centralized telephone support. This transformation standardized service delivery across entire organizations. Companies achieved significant operational savings while maintaining consistent customer communication for the first time.
Customer viewpoint: Why am I on hold for 20 minutes?
1990s 2010s
Multichannel Contact Centers Breaking Down the Walls
Added email, chat, and social media, but often as separate operations.
Multichannel centers fueled ecommerce growth by supporting customers through emerging digital channels. New service expectations developed as consumers demanded help beyond traditional business hours. Companies built specialized teams for each channel, which created new career paths in digital customer engagement.
Customer viewpoint: Can I just email you instead of calling?
2010 2020
Omnichannel Contact Centers Making Connections
Connected channels but still with fragmented customer data.
Omnichannel approaches elevated customer experience as a primary business differentiator beyond price and product. Organizations began measuring success through customer satisfaction metrics instead of operational efficiency alone. Seamless experiences across touchpoints became essential as consumer loyalty increasingly depended on effortless service interactions.
Customer viewpoint: Why do I need to repeat myself every time I contact you?
2020+
AI Powered Contact Hubs The Intelligent Revolution
Intelligent AI-powered systems that unify customer data and provide AI-assisted insights.
AI-powered hubs transform service operations from reactive cost centers into proactive revenue generators. Predictive capabilities empower companies to address customer needs before problems fully develop. Human agents now focus on complex problem-solving while technology handles routine interactions. Service insights drive product development and marketing strategies across the entire organization.
Customer viewpoint: The system already knew what I needed before I explained it.
Unified customer data
The foundation of effective customer engagement is having a complete, unified view of each customer by bringing all customer information together. When properly implemented, this comprehensive customer profile becomes the central hub from which meaningful interactions flow.
A unified customer profile includes data such as the following:
Basic identity information – Names, contact details, account numbers, and authentication history
Purchase and billing history – Complete transaction records, subscription status, and payment preferences
Previous interactions across all channels – Conversation history from calls, chats, emails, and social media engagements
Service cases and their resolutions – Past issues, how they were resolved, and follow-up outcomes
Preferences and behavioral patterns – Communication preferences, product usage habits, and engagement tendencies
Without this unified foundation, customer data remains trapped in isolated systems, which creates fragmented experiences. When agents lack access to complete information, customers face repetitive questions and inconsistent service.
Even the most sophisticated AI tools will struggle to deliver meaningful improvements in customer engagement when working with disjointed data. Personalization begins with this consolidated view that empowers both human agents and AI systems to understand the full customer context.
Avoid these critical engagement pitfalls
Before revolutionizing your customer experience strategy, be aware of the following crucial pitfalls that even seasoned organizations could encounter:
Never launch new engagement channels without establishing proper data infrastructure first.
Do not rush into automation solutions while your customer data remains fragmented across systems.
Remember that without a comprehensive view of the entire customer journey, your personalization efforts will fall short.
By addressing these fundamental challenges before implementing advanced technologies, you can build a solid foundation that maximizes the return on your customer experience investments.
Personalizing Customer Interactions
AI is transforming customer interactions from generic scripts to personalized conversations. Amazon Connect AI capabilities help businesses recognize customers instantly, understand their needs more deeply, and deliver more relevant experiences.
Beyond basic recognition
Identity Resolution in Amazon Connect uses both rule-based and machine learning (ML) approaches to identify and consolidate matching customer profiles. This helps create a unified view of the customer across channels and systems as follows:
Rule-based matching – Uses exact matching on key identifiers such as phone numbers, email addresses, and account numbers
ML matching – Goes beyond basic rules to detect patterns and relationships that indicate the same customer across different records
With Amazon Connect Identity Resolution, businesses can maintain conversation continuity across channels. This helps honor preferences consistently regardless of how customers identify themselves. Companies can deliver truly personalized experiences without requiring customers to repeatedly explain who they are.
Identity Resolution in action
Customer Profile Alejandro "Ale" Rosalez
Alejandro has used both his personnel email and work email when contacting your business. He has also contacted your business using both the name Alejandro and the name Ale.
Alejandro "Ale" Rosalez
Traditional systems would create separate records for Alejandro and name Ale and each email used. This fragments the customer's history across multiple profiles.
Identity Resolution in Amazon Connect can recognize these are likely the same person based on multiple data points. The system analyzes behavior patterns, device information, and location data.
The numbered markers in the following figure show the identity resolution process.
Figure 12 Personalizing Customer InteractionsSelect image to enlarge
1 - Ingest data
Collect customer data from all sources, with real-time and batch data transfer through software as a service (SaaS), database, and data warehouse connectors.
2 - Map to a profile
Build and update customer profiles in real time as data is ingested.
3 - Resolve identities
Use ML-powered identity resolution.
4 - Merge and enrich profiles
Merge duplicate records using configurable exact and probabilistic matching.
Use out-of-the-box insights and segments from customer interactions.
AI enhanced context for better first contact resolution
AI capabilities of Amazon Connect systematically analyze multiple data points to provide agents with comprehensive context at the moment of customer contact. This proactive intelligence increases the likelihood of resolving issues during the first interaction.
Figure 13 Personalizing Customer InteractionsSelect image to enlarge
As illustrated in the diagram, Amazon Connect places personalization at the center of the customer experience by intelligently connecting the following:
Contact information – Identifying the customer across channels
Contact history – Maintaining continuity between interactions
Case information – Bringing relevant support details forward
Purchase history – Understanding what products and services the customer uses
Customer insights – Using predictive analytics to anticipate needs
This unified approach transforms typical fragmented support experiences into seamless interactions where customers feel recognized and understood from the first moment of contact.
Calculated attributes
Calculated attributes represent a key AI capability that elevates customer service beyond basic identification to truly intelligent engagement. These dynamically generated insights transform raw behavioral data into predictive customer understanding that drives meaningful personalization.
Calculated attributes create actionable intelligence as follows:
Identifying a customer's preferred communication channel – Analyzing patterns to determine whether a customer responds best to chat, email, SMS, or phone contact
Calculating average order frequency – Determining typical purchase cycles to anticipate needs and identify potential upsell opportunities
Recognizing past behavior such as frequency of contact – Establishing baselines for normal customer behavior that help identify potential satisfaction issues
Detecting rising frustration based on recent interactions – Analyzing tone, word choice, and contact patterns to predict customer sentiment
For example, a customer typically contacts support through chat during business hours. If the customer unexpectedly calls after hours, the system might flag this as an urgent issue requiring special attention. This pattern deviation intelligence makes it possible for support teams to prioritize appropriately and respond with the appropriate level of urgency.
Figure 14 Personalizing Customer InteractionsSelect image to enlarge
The system adapts with each interaction, building a deeper understanding of customer needs to guide automated systems and human agents.
Human AI Partnership
The most effective implementations of AI for customer engagement augment human agents instead of replacing them. AI can augment agents as follows:
AI handles data gathering and analysis.
AI suggests possible solutions and provides context.
Human agents apply judgment, empathy, and creativity.
Together, they deliver better outcomes than either could alone.
For example, when a customer contacts support about a product issue, Amazon Connect can immediately recognize them. It can pull up their purchase history and identify the likely product they are calling about. Amazon Connect can recommend solutions based on similar cases before the agent even greets the customer. This leaves the agent free to focus on the human elements of the interaction that AI cannot replicate.
Benefits of AI powered customer engagement
Organizations implementing AI-powered customer engagement solutions see improvements across multiple dimensions of their business. The following tangible benefits demonstrate why forward-thinking companies are adopting these technologies to transform their customer experience operations. Explore each benefit below to understand how AI creates value for both customers and businesses.
Improved first contact resolution
When customers get their issues resolved on the first contact, it delivers the following benefits:
Lower support costs (fewer repeat contacts)
Higher customer satisfaction
Reduced customer effort
"More than 80% of companies cite improved CSAT as the top driver for proactive outreach." AI for Business Success, Jan 2024. Metrigy.
Reduced handle time without sacrificing quality
AI helps agents work more efficiently without cutting corners by doing the following:
Pre-populating customer information from unified profiles
Capturing and categorizing interaction details
Suggesting appropriate next steps
Enhanced personalization at scale
AI makes personalization scalable as follows:
Recognizing customers across channels instantly
Providing relevant context at the right moment
Suggesting personalized next-best actions
Improved operational efficiency
Beyond customer experience benefits, AI delivers significant operational improvements such as the following:
More efficient agent onboarding
Optimized workforce management
Reduced technical debt
Measurable business outcomes
The benefits translate to concrete business results:
Cost Savings
Reduced average handle time
Higher first-contact resolution
More efficient agent training
Revenue Protection and Growth
Improved customer retention
Increased cross-sell/upsell
Higher customer lifetime value
Customer Experience Improvements
Higher customer satisfaction scores
Reduced customer effort
Improved Net Promoter Score (NPS)
As customer expectations continue to rise, the gap between companies that embrace AI-powered engagement strategies and those that continue to use fragmented, reactive approaches widens. Leading organizations will view AI not merely as automation, but as a tool for understanding customers at a deeper level.
RECALL & APPLY
Practice what you learned
Try answering in your own words, then compare with the suggested answer. These activities are for self-study and are not automatically graded.
01What should be improved before adding more automation to disconnected customer systems?
Topic: AI powered customer engagement
Reveal suggested answer
Establish the data foundation and integration needed to understand the customer and the current issue.
02Why is adding chat different from connecting the customer journey?
Topic: Contact center evolution
Reveal suggested answer
Adding a channel provides another interface. Continuity requires sharing relevant identity, history, and case context across those interfaces.
03Two records share a nickname but have different account details. Should they merge automatically?
Topic: Identity and personalization
Reveal suggested answer
A nickname alone is insufficient. Review the configured matching and merging criteria and the consequences of a false match.
04How can a calculated attribute be more useful than a long activity list?
Topic: Customer context and calculated attributes
Reveal suggested answer
It condenses relevant behavior, such as contact frequency, into a value the workflow can use, while the underlying history remains useful for investigation.
05Why should AHT be reviewed together with FCR?
Topic: The human and AI partnership
Reveal suggested answer
A shorter contact is not an improvement if it creates unresolved issues and repeat contacts. Measure both speed and resolution quality.
Capture a key idea, an example, or a question for your instructor.
In this section, you will learn to do the following:
Configure Amazon Connect Customer Profiles with generative AI data mapping to unify customer data from multiple sources.
Implement Identity Resolution using both rule-based and ML approaches to create consolidated customer profiles.
Apply best practices for profile merging to improve agent efficiency and customer experience.
Section introduction
Imagine answering a customer call and immediately having access to your customer's complete history, preferences, and account details. You have their data organized in one place without switching between multiple systems.
With Amazon Connect Customer Profiles, you can combine contact history from Amazon Connect with information from external applications to provide this improved experience.
In this section, you will learn how generative AI streamlines data mapping and how identity resolution automatically finds and merges duplicate customer records.
Components of Amazon Connect Customer Profiles
Amazon Connect Customer Profiles combine customer contact history with important information such as their account number, personal details, contact information, and interaction preferences. After it is enabled, Amazon Connect Customer Profiles automatically creates a unique profile for every customer who contacts your organization.
When a customer contacts your organization, Amazon Connect can identify the customer. Typically, it uses a phone number to identify the customer for a voice call or email address for chat. Amazon Connect Customer Profiles then retrieves the customer's existing profile or creates a new one. The profile may contain the following:
Personal information (name, date of birth, email addresses)
Account details (account numbers, status)
Contact history (previous calls, chats, and interactions)
Case information (open issues, resolved problems)
Data from external systems (customer relationship management records, order history)
Benefits for agents and customers
Benefits for Agents
Having all customer information in one place benefits human support agents in the following ways:
Reduced handle time – Agents no longer need to switch between systems.
Personalized service – Agents can greet customers by name and acknowledge their contact history.
Figure 15 Components of Amazon Connect Customer ProfilesSelect image to enlarge
Benefits for Customers
The benefits for customers are as follows:
Reduced repetition – Customer details are stored, which reduces the need to repeat the same information.
Faster resolution – Customers experience faster issue resolution because agents have customer data when they connect.
Personalized service – Agents can greet customers by name or route them to an agent faster due to account status.
Consistent experience – Customers experience is more consistent across different communication channels.
Figure 16 Components of Amazon Connect Customer ProfilesSelect image to enlargeFigure 17 Components of Amazon Connect Customer ProfilesSelect image to enlarge
1 - Customer profile
This section includes information such as account number, additional information, birth date, email, multiple addresses, name, and party type.
2 - Cases
This section includes status, reference ID, title, source, updated date, and more information related to cases ingested from third-party applications. This is in addition to cases created and managed using Amazon Connect Cases.
3 - More information
This section includes customer-defined attributes and might contain information like cell phone number and shipping address.
This information is sorted alphabetically to help an agent quickly locate the information they need.
4 - Contact history
This section includes dates, times, and duration when this customer contacted your contact center in the past.
5 - Product purchase history
All the assets purchased by a customer can be populated here. The data is ingested from an external application that you have integrated with Amazon Connect Customer Profiles.
RECALL & APPLY
Practice what you learned
Try answering in your own words, then compare with the suggested answer. These activities are for self-study and are not automatically graded.
01Which data would help an agent handle a repeat delivery complaint?
Topic: Amazon Connect Customer Profiles
Reveal suggested answer
The customer’s identity, order and shipping history, previous contacts, open case, and any promised follow up.
02What should the instructor point out in the profile screenshot?
Topic: The customer profile workspace
Reveal suggested answer
Locate the profile, cases, additional attributes, contact history, and purchase history. Explain how each area supports a different part of the conversation.
Capture a key idea, an example, or a question for your instructor.
Modern contact centers handle massive volumes of customer data from multiple sources. Understanding how to effectively integrate this data is crucial for delivering exceptional customer experiences.
Comparing traditional and AI powered data integrations
The following sections compare traditional data integration with generative AI data mapping in Amazon Connect.
Traditional Data Integration Methods
Traditional data integration relies on manual processes and predefined rules.
Organizations typically use extract, transform, load (ETL) pipelines where data engineers manually map fields between systems.
For example, mapping a customer's phone_number field from a customer relationship management (CRM) system to the contact_number field in Amazon Connect requires explicit programming.
Key characteristics of traditional methods include the following:
Manual field mapping – Engineers identify and code relationships between data fields.
Static rule-based transformations – Predefined logic handles data format conversions.
Batch processing – Data moves in scheduled intervals rather than in real time.
High maintenance overhead –Changes require developer intervention and testing.
The next section explains generative AI data mapping.
Generative AI-Powered Data Mapping
Generative AI-powered data mapping in Amazon Connect Customer Profiles is an intelligent system that analyzes your customer data. The data mapping feature uses data from various sources and automatically determines how to organize and combine it into unified profiles. This capability reduces the time needed to create unified customer profiles.
For example, an insurance company can use the generative AI-powered data mapping in Amazon Connect to automatically consolidate customer information. This makes it possible for the company to map customer data from their claims database, CRM system, and payment portal into unified profiles. Agents can view a comprehensive customer history on a single screen and reduce call handling time.
Key characteristics of generative AI-powered features include the following:
Intelligent pattern recognition – AI identifies similar data types across systems automatically.
Semantic understanding – Natural language processing recognizes that customer_phone and contact_number represent the same concept.
Real-time adaptation – Continuous learning improves mapping accuracy over time.
Self-healing integrations – Automatic adjustments occur when source systems change.
The following sections describe the mapping process and configuration.
Although traditional data integration methods provide proven reliability, generative AI-powered automated mapping represents the future of contact center data management. It offers unprecedented speed, accuracy, and scalability for modern customer service operations.
Generative AI Powered Data Mapping Process
The data mapping process consists of the four phases described below.
How does data mapping with generative AI work
Generative AI transforms how you map customer data in Amazon Connect Customer Profiles. The system analyzes your incoming data from various sources and intelligently determines relationship patterns between fields.
Generative AI-powered data mapping in Amazon Connect Customer Profiles automatically determines how to organize and combine it into unified profiles.
This significantly reduces the time needed for unified customer profiles, which empowers you to deliver more personalized customer experiences efficiently.
Let's see how it works in action.
Phase 1 Data Collection and Analysis
Figure 18 Phase 1 Data Collection and AnalysisSelect image to enlarge
The system fetches source attributes and, if available, sample data from your data source. For an Amazon Simple Storage Service (Amazon S3) data source, the first CSV file found in the selected Amazon S3 bucket and prefix will be used as the sample data.
For other data sources such as Salesforce or Zendesk, it fetches attributes through Amazon AppFlow.
Phase 2 Smart Field Mapping
Figure 19 Phase 2 Smart Field MappingSelect image to enlarge
A large language model (LLM) processes each custom attribute from your data source and maps them to standard customer profile attributes.
For example, if your system has fields labeled CustomerEmailAddress, Email_Contact, and EmailID, the AI recognizes these are all referring to email addresses and maps them appropriately.
Phase 3 Key Attribute Selection
Figure 20 Phase 3 Key Attribute SelectionSelect image to enlarge
After mapping fields, the LLM selects suitable attributes that can serve as keys (standard identifiers) for customer profiles. These include the following:
Unique identifier – You must have a unique identifier for your data to avoid ingestion errors. This identifier, also known as the unique key, distinguishes data and enables indexing for search and updates. There can be only one unique identifier.
Customer identifier – You must have at least one customer identifier to avoid ingestion errors. Also known as the profile key, Amazon Connect Customer Profiles uses it to determine if data can be associated to existing profiles or create new ones by searching other profiles for this identifier. You can have multiple customer identifiers.
Product identifier – You must have at least one product identifier to avoid ingestion errors. Also known as the asset key, Amazon Connect Customer Profiles uses it to distinguish from other customer product purchase data and determine profile associations by searching other profiles for this identifier. You can have multiple product identifiers.
Case identifier – You must have at least one case identifier to avoid ingestion errors. Also known as the case key, Amazon Connect Customer Profiles uses it to distinguish from other customer case data and determine profile associations by searching other profiles for this identifier. You can have multiple case identifiers.
Order identifier – You must have at least one order identifier to avoid ingestion errors. Also known as the order key, Customer Profiles uses it to distinguish from other customer order data and determine profile associations by searching other profiles for this identifier. You can have multiple order identifiers.
Additional search attributes (optional) – You can choose attributes in your data source object that you want to index to be searchable. By default, all your identifiers are indexed.
After mapping fields, the LLM selects suitable attributes that can serve as keys (standard identifiers) for customer profiles. These include the following:
Unique identifier – You must have a unique identifier for your data to avoid ingestion errors. This identifier, also known as the unique key, distinguishes data and enables indexing for search and updates. There can be only one unique identifier.
Customer identifier – You must have at least one customer identifier to avoid ingestion errors. Also known as the profile key, Amazon Connect Customer Profiles uses it to determine if data can be associated to existing profiles or create new ones by searching other profiles for this identifier. You can have multiple customer identifiers.
Product identifier – You must have at least one product identifier to avoid ingestion errors. Also known as the asset key, Amazon Connect Customer Profiles uses it to distinguish from other customer product purchase data and determine profile associations by searching other profiles for this identifier. You can have multiple product identifiers.
Case identifier – You must have at least one case identifier to avoid ingestion errors. Also known as the case key, Amazon Connect Customer Profiles uses it to distinguish from other customer case data and determine profile associations by searching other profiles for this identifier. You can have multiple case identifiers.
Order identifier – You must have at least one order identifier to avoid ingestion errors. Also known as the order key, Customer Profiles uses it to distinguish from other customer order data and determine profile associations by searching other profiles for this identifier. You can have multiple order identifiers.
Additional search attributes (optional) – You can choose attributes in your data source object that you want to index to be searchable. By default, all your identifiers are indexed.
Phase 4 Timestamp Processing
Figure 21 Phase 4 Timestamp ProcessingSelect image to enlarge
Finally, the system parses timestamps to maintain the correct chronological order of records. This ensures that customer interactions and updates appear in the right sequence.
The entire process typically takes minutes. This improves speed and accuracy compared to manual data mapping, which could take hours or days.
Setting up Generative AI Powered Data Mapping in Amazon Connect Customer Profiles
Contact center administrators can review and complete the setup of customer profiles for data mapping to occur. This will provide agents with relevant customer information and dynamically personalized interactive voice response (IVR) and self-service assistants to improve customer satisfaction and agent productivity.
Welcome to this demonstration on setting up generative AI-powered data mapping in Amazon Connect Customer Profiles.
To start, sign in to the AWS Management Console.
In the search box at the top of the page, enter Amazon Connect. In the displayed list of services, choose Amazon Connect.
On the Amazon Connect page, the instances in your account are displayed. Choose the Instance alias of the instance that you want to set up automated data mapping in Amazon Connect Customer Profiles on.
On the Instance page, under Applications, choose Customer Profiles.
On the Data source integrations tab, choose Add data source integration to begin the process.
Select a Data source from any of the 70+ available no-code data connectors such as Adobe Analytics, Salesforce, or Amazon Simple Storage Service (S3). Each connector is designed to pull specific types of customer data. Configure the connector you have selected, then choose Next.
When you reach the Map data step, select the Auto-generate mapping option. This is where the generative AI capabilities are used to automatically map the attributes of your data to Amazon Connect Customer Profiles attributes. Then, choose Next to continue.
You will now be presented with a summary of all the automatically mapped Amazon Connect Customer Profiles attributes. This is your opportunity to review and make any necessary adjustments. When you are ready, choose Next to continue.
When you are satisfied with the configuration, choose Add data source integration to complete the connector configuration and begin ingesting your customer data into Amazon Connect Customer Profiles.
Your new connector will now be listed in your Amazon Connect Customer Profiles Data source integrations. When the connector is complete, it will show as Active.
RECALL & APPLY
Practice what you learned
Try answering in your own words, then compare with the suggested answer. These activities are for self-study and are not automatically graded.
01Why must a proposed field mapping be reviewed?
Topic: Data integration and generative mapping
Reveal suggested answer
Similar names can represent different business meanings. Incorrect mappings can corrupt identity, dates, or customer context.
02Why do identifiers and timestamps deserve special attention?
Topic: The data mapping process
Reveal suggested answer
Identifiers determine association and updates. Timestamps affect record ordering and which data appears current.
03What is the critical review point before completing the integration?
Topic: Data mapping demonstration
Reveal suggested answer
Check each mapping, required identifier, and timestamp interpretation. Verify that sample records create or update the intended profiles.
Capture a key idea, an example, or a question for your instructor.
There can be multiple profiles when customer records are captured across multiple channels and applications for the same customer. Without a common unique identifier linking these profiles, the same customer ends up with separate, disconnected profiles across different systems.
Identity Resolution in Amazon Connect Customer Profiles tackles this challenge by finding each profile and consolidating them. Identity Resolution uses both rules-based matching and ML matching to ensure that each customer has just one comprehensive profile.
Rule-based matching relies on predefined rules to determine if two profiles represent the same customer. This method compares specific attributes between profiles using a set of matching rules.
These rules examine attributes such as the following:
Account number
Address components (city, postal code, country)
Email addresses
Phone numbers
Name components (first name, last name, middle name)
Date of birth
You can configure up to 15 different matching rules to catch various scenarios where profiles might represent the same customer. The strength of rule-based matching is its precision and transparency. You know exactly why two profiles were matched.
You can choose how profiles are compared across attribute types and which attribute to use for matching from each type.
For example, if you want to match multiple email types, choose many-to-many to match across profiles and attribute types that your business uses.
You can choose from multiple attribute types.
Email type
Choose from the following:
EmailAddress
BusinessEmailAddress
PersonalEmailAddress
Phone number type
Choose from the following:
PhoneNumberNumber
HomePhoneNumber
MobilePhoneNumber
Address type
Choose from the following:
Address
BusinessAddress
MaillingAddress
ShippingAddress
matching (matches across sub-types) as follows:
ONE_TO_ONE – The system can only match if the sub-types are exact matches.
For example, when the EmailAddress fields of Profile A and B match, the two profiles are matched on the EmailAddress type.
MANY_TO_MANY – The system can match attributes across the sub-types of an attribute type.
For example, if EmailAddress for Profile A matches BusinessEmailAddress for Profile B, the profiles are matched based on EmailAddress type.
ML matching
For more sophisticated matching capabilities, Amazon Connect offers ML-based Identity Resolution. This approach uses AI to identify similarities that might not be caught by basic rules.
ML-based Identity Resolution reviews the following personal identifiable information (PII) attributes in each profile:
Names (first, middle, last)
Email addresses (personal, business)
Phone numbers (home, mobile, business)
Addresses (business, mailing, shipping, billing)
Date of birth
Unlike rule-based matching, ML-based matching can detect similarities even when information is not exactly the same.
For example, it might recognize that Ana Carolina Silva and Ana Silva with similar addresses are likely the same person.
Setting up Identity Resolution in Amazon Connect Customer Profiles
Transcript Setting up Identity Resolution in Amazon Connect Customer Profiles
Welcome to this demonstration on setting up Identity Resolution for your Amazon Connect Customer Profiles.
To start, sign in to the AWS Management Console.
In the search box at the top of the page, enter Amazon Connect. In the displayed list of services, choose Amazon Connect.
On the Amazon Connect page, the instances in your account are displayed. Choose the Instance alias that you want to set up Identity Resolution for Amazon Connect Customer Profiles.
On the Instance page, under Applications, choose Customer Profiles.
On the Amazon Connect Customer Profiles page, in the Identity Resolution section, choose Enable identity resolution to begin the process.
A message appears asking you to confirm that you understand that enabling Identity Resolution will not result in any profile merging, and that you will need to enable merging separately within Identity Resolution. After reading the message, choose Enable Identity Resolution to proceed.
You've now successfully enabled Identity Resolution with rule-based matching and machine learning-based matching for your Amazon Connect Customer Profiles domain.
After enabling Identity Resolution for a new domain with rule-based matching, the matching will start immediately if you have an integration running. For existing domains, the matching process will start within one hour. For machine learning-based matching, the Identity Resolution job will run for the first time within 24 hours of enabling the feature.
RECALL & APPLY
Practice what you learned
Try answering in your own words, then compare with the suggested answer. These activities are for self-study and are not automatically graded.
01Does enabling Identity Resolution immediately merge all matches?
Topic: Identity Resolution matching
Reveal suggested answer
No. The course distinguishes identifying matches from enabling merging. Review that separation during the demonstration.
02Why might a personal email need to match a business email field?
Topic: Attribute matching and setup
Reveal suggested answer
The same address may be stored under different subtypes in different source systems. Matching configuration controls whether those comparisons are allowed.
Capture a key idea, an example, or a question for your instructor.
After Identity Resolution identifies similar profiles (either through rules or ML), the next step is to consolidate them through the following auto-merging process:
Apply consolidation criteria that you define to determine which profiles should be combined.
Create a consolidated profile that combines information from all matching profiles.
Preserve information from the original profiles, and retain all values if there are conflicts.
Update references to the original profiles to point to the new consolidated profile.
Auto merging process
The Identity Resolution process in Amazon Connect Customer Profiles, automatically combines fragmented customer data from multiple sources into a unified customer profile.
In the following example, John Doe creates an account with his personal email. Later, he contacts support using his work email. Without Identity Resolution, the system would create two separate profiles. With Identity Resolution, if other attributes, such as the name and phone number match, the system recognizes these are the same person and merges the profiles.
The numbered markers in the following figure show the auto merging process.
Figure 22 Auto merging processSelect image to enlarge
1 - Fragmented first-party (1P) data collection
Customer data from various touchpoints:
Advertising: Mobile ID
Email: Name and email address
Site: Name, ID and email
Retail: Full name, ID, email, and address
App: Mobile ID and phone number
Contact center: Name and phone number
2 - Automatic matching and merging
The system processes fragments through a three-stage workflow:
1P deterministic match: Identifies exact matches like identical email addresses
1P probabilistic match: Identifies exact matches like identical email addresses
Auto-merging records: Combines matched information into a single profile using custom workflows
3 - Unified profile creation
The result is a comprehensive customer profile containing the following information:
Full name: John Doe
ID: CRM123
Email address: J.Doe@email.com
Phone number: 212-867-5309
Address: 123 4th St NY, NY
Mobile ID: M1234
This automated process helps businesses recognize the same customers across different channels, which can result in better customer service experiences and more effective marketing.
Setting up Auto Merging in Identity Resolution for Amazon Connect Customer Profiles
When similar profiles are detected by an Identity Resolution job, the process can automatically merge them into a unified profile based on auto-merging rules that you specify.
Transcript Setting up Auto Merging in Identity Resolution for Amazon Connect Customer Profiles
Welcome to this demonstration on setting up auto-merging in Identity Resolution for Amazon Connect Customer Profiles.
To start, sign in to the AWS Management Console.
In the search box at the top of the page, enter Amazon Connect. In the displayed list of services, choose Amazon Connect.
On the Amazon Connect page, the instances in your account are displayed. Choose the Instance alias of the instance that you want to set up Identity Resolution in Amazon Connect Customer Profiles on.
Now, on your Amazon Connect instance Overview page in the navigation pane, under Applications, choose Customer Profiles.
Next, in the Identity Resolution section, choose View Identity Resolution.
On the Identity Resolution page, in the Identity Resolution settings section, you will see the Rule-based resolution and Machine learning resolution. Choose the resolution that you want to setup auto-merging on. For this demonstration, choose the Rule-based resolution.
Now, on the Rule-based resolution settings page, in the Merge rule-based matches section, choose Edit.
The Missing timestamp pop-up appears and indicates whether you have custom object type mappings. Amazon Connect uses a timestamp attribute and timestamp format to determine when a profile was last updated. If the Missing timestamp pop-up appears, it means there is a timestamp missing from your custom objects. You can add it using the PutProfileObjectType API. If your object type does not have a proper timestamp attribute, you can acknowledge that a default timestamp will be applied for records ingested into Amazon Connect Customer Profiles. For this demonstration, select the acknowledgment. Then, choose Next.
On the Edit merge rule-based matches page, in the Merge matches section, you will define when to merge profiles based on the selected rule. Select the Merge matches found by rule-based matching checkbox.
Next, choose a rule level for merging from the list available. For this demonstration, choose Rule 1.
Then, choose Save.
In the Merge data pop-up, enter confirm in the textbox.
Then, choose Merge data.
Now, on the Rule-based resolution settings page, notice that the rule-based resolution settings have been updated.
Next, decide if you want to review matched profile IDs by having them written to an Amazon Simple Storage Service (Amazon S3) bucket. For this demonstration, assign an S3 bucket to the rule-based resolution. In the Match results location section, choose Edit.
Now, on the Edit match results location page, in the S3 location - optionalsection, select the Write profile ID matches to Amazon S3 checkbox.
To choose your bucket destination, either enter the S3 URI or choose Browse S3. For this demonstration, choose Browse S3.
In the Choose an archive in S3 pop-up, under Buckets, enter the name of the S3 bucket.
Now, choose the S3 bucket.
Then, select Choose.
Now, with the S3 bucket URI entered into the S3 URI textbox, choose Save.
On the Rules-based resolution settings page, an alert indicates you have successfully assigned the S3 bucket location.
Next, choose View Identity Resolution.
On the Identity Resolution page, notice that the rule-based resolution has an Active status for both Find matches and Merge matches. Also, S3 location displays the chosen S3 bucket.
Now, you will complete the machine learning resolution. In the Identity Resolution settings section, choose Machine learning resolution.
On the Machine learning resolution settings page, in the Merge matches section, choose Edit.
Now, on the Edit merge machine learning matches page, in the Merge matches section, define when to merge profiles based on attribute rules you create. Select the Merge matches found by machine learning matching checkbox.
Next, under Rule 1, choose the Attributes menu, and then choose one or more attributes from the list. For this demonstration, choose AccountNumber and EmailAddress.
To add more rules, choose Add merge rule. For the second rule, choose AccountNumber only, and for the third rule, choose EmailAddress only.
Then, choose Save.
In the Merge data pop-up, enter confirm in the textbox.
Then, choose Merge data.
Now, on the Machine learning resolution settings page, notice that the machine learning resolution settings have been updated.
Next, decide if you want to review matched profile IDs by having them written to an S3 bucket. For this demonstration, you will assign an S3 bucket to the machine learning resolution. In the Match results location section, choose Edit.
Now, on the Edit match results location page, in the S3 location - optional section, select the Write profile ID matches to Amazon S3 checkbox.
To choose your bucket destination, either enter the S3 URI or choose Browse S3. For this demonstration, choose Browse S3.
In the Choose an archive in S3 pop-up, under Buckets, enter the name of the S3 bucket.
Next, choose the S3 bucket.
Then, select Choose.
Now, with the S3 bucket URI entered into the S3 URI textbox, choose Save.
On the Machine learning resolutionsettings page, an alert indicates you have successfully assigned the S3 bucket location.
Next, choose View Identity Resolution.
On the Identity Resolution page, notice that the machine learning resolution has an Active status for both Find matches and Merge matches. Also, S3 location displays the chosen S3 bucket.
That's it. You successfully set up auto-merging in Identity Resolution and assigned S3 buckets to store matched profile IDs for Amazon Connect Customer Profiles.
RECALL & APPLY
Practice what you learned
Try answering in your own words, then compare with the suggested answer. These activities are for self-study and are not automatically graded.
01What risk does an overly broad merge rule create?
Topic: Auto merging customer profiles
Reveal suggested answer
It can consolidate different people into one profile, producing incorrect customer context and potential data exposure.
02What should you review before enabling a merge rule?
Topic: Auto merging demonstration
Reveal suggested answer
Review the attributes, rule level, timestamps, expected match examples, and likely false matches. Use an authorized training environment for the demonstration.
Capture a key idea, an example, or a question for your instructor.
In this section, you will learn to do the following:
Configure Amazon Connect outbound campaigns with AI-powered call classification to optimize agent productivity.
Create targeted customer segments using both natural language prompts and data-driven inspiration cards.
Implement segmentation strategies that improve campaign performance and customer engagement metrics.
Section introduction
Imagine you are tasked with reaching out to thousands of customers who haven't made a purchase in the past month. How would you identify these customers? How would you make sure that your outreach is personalized and effective? This is where outbound campaign intelligence comes into play.
Setting up Outbound Campaigns
Outbound campaigns are proactive communications initiated by businesses to their customers, and they are essential for everything from appointment reminders to marketing promotions.
Amazon Connect outbound campaigns can help with various proactive communications. Review the following examples of outbound campaigns:
Appointment reminders
Marketing promotions
Delivery notifications
Billing reminders
Service follow-ups
Customer satisfaction surveys
Preparing to set up Amazon Connect outbound campaigns
Before beginning your first campaign, you need to complete the following prerequisites:
In the following demonstration, you will learn about enabling outbound campaigns in the Amazon Connect. Amazon Connect outbound campaigns allow you to automate outbound campaign voice communications to your customers.
Transcript Configuring Outbound Campaigns in Amazon Connect
Welcome to this demonstration on how to set up Amazon Connect outbound campaigns, formerly known as high-volume outbound communications.
To start, sign in to the AWS Management Console.
In the search box at the top of the page, enter Amazon Connect. In the displayed list of services, choose Amazon Connect.
On the Amazon Connect page, the instances in your account are displayed. Choose the Instance alias of the instance that you want to enable outbound campaigns on.
Next, in the navigation pane, under Channels and communications, choose Outbound campaigns.
On the Outbound campaigns page, choose Enable. If you don't have this option, verify whether outbound campaigns is available in your AWS Region.
Next, under the Encryption section, you will need to either enter your own AWS KMS key or choose to create a new one. For this demonstration, you will create a new KMS key.
Choose Create an AWS KMS key. A new browser tab will open for the AWS Key Management Service (AWS KMS) console.
On the Create key page, in the Configure key step, review that the Key type is set to Symmetric and Key usage is set as Encrypt and decrypt. Then, choose Next.
On the Add labels step, enter a descriptive alias and description for your key. In this demonstration, for Alias, enter KMS-AmazonConnectOutboundCampaigns. For Description, enter KMS for Amazon Connect Outbound Campaigns. When you are done, choose Next.
On the Define key administrative permissions - optionalstep, choose Next.
On the Define key usage permissions - optionalstep, choose Next.
On the Edit key policy - optionalstep, choose Next.
Finally, on the Review step, choose Finish to complete the key creation process.
Now, return to the Amazon Connect console tab. Choose the AWS KMS key field, and review your list of available keys. Your newly created key should appear in the list. In this demonstration, you will choose the KMS-AmazonConnectOutboundCampaigns key.
With everything configured, choose Enable outbound campaigns. The process will take a few minutes to complete. After outbound campaigns is enabled, you can begin creating outbound campaigns for voice calls in Amazon Connect.
If the enabling process fails, you might need to verify that you have the required AWS Identity and Access Management (IAM) permissions for the key, including kms:DescribeKey, kms:CreateGrant, and kms:RetireGrant.
To enable your total instance limits for outbound campaigns, you need to make sure that your campaign permissions are up to date. Choose Upgrade permission.
This completes the demonstration on setting up Amazon Connect outbound campaigns.
RECALL & APPLY
Practice what you learned
Try answering in your own words, then compare with the suggested answer. These activities are for self-study and are not automatically graded.
01Why is the flow’s call progress branch important?
Topic: Outbound campaign setup
Reveal suggested answer
It determines how the campaign handles a person, an answering machine, or an uncertain outcome.
02Where would you investigate an enablement failure involving the key?
Topic: Outbound campaign demonstration
Reveal suggested answer
Review the key configuration and the required permissions identified in the course procedure. Confirm the intended instance and region as well.
Capture a key idea, an example, or a question for your instructor.
Customer segments are specific groups of people who share common characteristics or behaviors. Customer segments are dynamically evaluated based on attributes that you define and can change over time when the value of the attributes change. In Amazon Connect, segments serve as the foundation for targeted outbound campaigns.
Instead of sending the same promotional message to all customers, you might create different segments, such as the following:
Customers who abandoned their shopping carts recently
High-value customers who haven't purchased in 30 days
New customers who made their first purchase in the past week
Real world example
The following diagram illustrates how a customer can be segmented based on interaction type (inbound compared to outbound contacts) and priority levels. At the center is Amazon Connect, which serves as the hub for all customer interactions.
On the left side of the diagram, inbound contacts are categorized into the following three priority segments:
Cart abandoners (urgent priority) with an estimated 3,200 contacts who abandoned their carts within 24–72 hours
High-value customers (high priority) with 2,450 estimated contacts who haven't made purchases in more than 30 days and have historical spend exceeding $500 each month
New customers (medium priority) with approximately 1,850 contacts who made their first purchase within 7 days
The right side of the diagram displays the following outbound contact strategies:
Recovery and conversion efforts (urgent priority) targeting 200 contacts with limited-time offers and purchase hesitation solutions
Reengagement strategies (high priority) focusing on 350 contacts through personalized discounts and loyalty program enrollment
Welcome and onboarding (medium priority) reaching 1,500 contacts with product tutorials and support information
Figure 23 Understanding Customer SegmentsSelect image to enlarge
Segmentation benefits
The more targeted your segments, the more personalized and effective your campaigns can be. The following are some examples:
Higher engagement rates: When customers receive relevant messages, they're more likely to engage.
Improved customer experience: Personalized communications demonstrate to customers that you understand their needs.
Better resource allocation: By focusing on the right customers at the right time, you optimize your outreach efforts.
AI in Campaign Management
AI has revolutionized outbound campaign management, enabling businesses to work smarter, not harder. From detecting whether a human or machine answered a call to receiving recommendations based on trends in the customer data, AI-powered tools make campaigns more efficient.
In Amazon Connect, AI powers the following key aspects of outbound campaigns.
Call Classification
Call classification uses machine learning to determine whether a call was answered by a person or an automated system.
Amazon Connect call classification technology analyzes several factors:
Background noise analysis: It can detect background noise patterns typically associated with pre-recorded messages.
Speech patterns: The system recognizes long strings of words common in voicemail greetings.
Human response patterns: Call classification identifies typical human responses, such as "Hello," followed by pauses.
Amazon Connect call classification connects agents only with live customers.
Segment AI Assistant
Generative AI-powered segmentation helps non-technical business users to build audiences using natural language queries.
Natural language segment creation represents a significant breakthrough in making data more accessible to non-technical users. Instead of navigating complex filter interfaces or writing database queries, you can describe the customers you want to target in everyday language.
Amazon Connect segment AI assistant interprets your natural language description and translates it into a structured segment definition that identifies the exact customers matching your criteria.
Inspiration Cards
This generative AI-powered feature presents segment ideas tailored to specific customer data and trends, streamlining segment creation.
Inspiration cards are AI-powered recommendations that analyze your customer data to suggest potentially valuable segments for your campaigns. Unlike traditional segmentation that requires you to define criteria from scratch, inspiration cards proactively identify patterns and opportunities in your data.
Understanding call classification
Call classification uses machine learning to determine whether a call was answered by a person or an automated system.
Call progress analysis
When your outbound campaign makes a call, the Check call progress flow block branches based on the call classification outcome. The following are a few examples:
If a human answers, it branches to connect to an agent.
If an answering machine responds, it branches to leave a pre-recorded message.
If the ML model cannot determine the answer type, it branches to play a message before connecting to an agent.
This intelligence matters because studies show that many calls to consumers go to voicemail. Without call classification, agents would waste countless hours listening to voicemail greetings and leaving messages.
Figure 24 AI in Campaign ManagementSelect image to enlarge
Agent productivity benefits
Beyond call classification, AI improves agent productivity in several key ways:
Minimizing idle time: Predictive dialing helps agents spend more time speaking with customers.
Prioritizing live connections: Because answering machines, wrong numbers, and no-answers are filtered out, agents stay focused on live connections.
Providing context: When a call connects to an agent, AI can instantly display relevant customer information.
Figure 25 AI in Campaign ManagementSelect image to enlarge
Customer experience impact
AI does not just benefit agents and businesses, it also improves the customer experience in the following ways:
Reduced spam perception: When calls connect to live agents immediately, customers are more likely to engage.
Personalization: AI helps agents have more relevant conversations by providing customer context.
Figure 26 AI in Campaign ManagementSelect image to enlarge
RECALL & APPLY
Practice what you learned
Try answering in your own words, then compare with the suggested answer. These activities are for self-study and are not automatically graded.
01How would “high value customers” become a testable segment?
Topic: Customer segments
Reveal suggested answer
Define the spend metric, threshold, and time period, then inspect the resulting members and exclusions.
02What should happen when the classifier cannot determine the answer type?
Topic: AI in campaign management
Reveal suggested answer
Use the flow’s defined uncertain outcome handling. Demonstrate that the design must account for this result explicitly.
03Why is a high dialing rate insufficient as a success measure?
Topic: Campaign productivity and customer experience
Reveal suggested answer
It does not show whether customers received useful, timely service or whether the campaign achieved its intended outcome.
Capture a key idea, an example, or a question for your instructor.
Natural language segment creation represents a significant breakthrough in making data more accessible to non-technical users. Instead of navigating complex filter interfaces or writing database queries, you can describe the customers you want to target in conversational language.
Amazon Connect segment AI assistant interprets your natural language description and translates it into a structured segment definition that identifies the exact customers matching your criteria.
How natural language segment creation works
Behind the scenes, the segment AI assistant uses advanced natural language processing algorithms to do the following:
Analyze your description to identify key criteria and conditions.
Map these criteria to available customer attributes in your domain.
Construct a logical segment definition with the appropriate filters and relationships.
Apply this definition to your customer data to generate the segment.
For example, instead of manually configuring multiple filters and conditions, you could enter: Customers who have spent over $2,000 in the last 60 days.
Figure 27 Segment AI AssistantSelect image to enlarge
Best practices for your segment AI assistant prompts
Amazon Connect segment AI assistant relies heavily on the quality of prompts to generate effective responses for customer service agents. The difference between vague and better formed prompts can significantly impact the assistant's performance and usefulness. The quality of your segment depends significantly on how you phrase your prompt. Use the following best practices when composing prompts for the segment AI assistant:
Be specific: Include precise criteria instead of vague descriptions.
Reference existing attributes: When possible, use the names of attributes that exist in your data.
Include clear timeframes: Specify time periods clearly, such as in the last quarter.
Start simple: Begin with straightforward prompts and gradually add complexity.
Use business terminology: Frame your prompt in terms of business objectives.
Examples of vague and better formed prompts
Figure 28 Segment AI AssistantSelect image to enlargeFigure 29 Segment AI AssistantSelect image to enlarge
Refining AI generated segments
After the AI generates your segment, you can do the following:
Review the segment definition: Check that all conditions reflect your intent.
Adjust thresholds: Fine-tune values like purchase amounts or timeframes.
Add or remove conditions: Enhance the segment with additional criteria or simplify.
Preview results: See how many customers match your segment and review a sample.
Figure 30 Segment AI AssistantSelect image to enlarge
Inspiration Cards
Inspiration cards are Amazon Connect AI-powered recommendations that analyze your customer data to suggest potentially valuable segments for your campaigns. Unlike traditional segmentation that requires you to define criteria from scratch, inspiration cards proactively identify patterns and opportunities in your data.
These cards appear on the Customer segmentspage in Amazon Connect and present ready-to-use segment ideas based on actual trends in your customer profiles data.
Figure 31 Segment AI AssistantSelect image to enlarge
Inspiration cards generated recommendations
Inspiration cards use advanced analytics and AI to generate their suggestions. They do the following:
Analyze historical data: Examine customer behavior patterns over time.
Identify trends: Spot significant changes or patterns in customer activities.
Apply business logic: Categorize these patterns into meaningful business contexts.
Generate actionable segments: Create predefined segment criteria based on these insights.
The system organizes these suggestions into the following business-focused themes:
Promotion: Segments ideal for marketing promotions and special offers
Retention: Segments highlighting customers who might need attention to prevent churn
Support: Segments identifying customers who might benefit from proactive service
RECALL & APPLY
Practice what you learned
Try answering in your own words, then compare with the suggested answer. These activities are for self-study and are not automatically graded.
01What must you inspect after entering a segment prompt?
Topic: Natural language segment creation
Reveal suggested answer
Inspect the attributes, operators, thresholds, timeframes, AND or OR relationships, and a sample of matched customers.
02Improve the prompt “find customers who spend a lot.”
Topic: Segment prompts and refinement
Reveal suggested answer
Use a defined metric and time period, for example customers whose recorded spend exceeds 2000 in the last 60 days, using attributes available in the domain.
03Should a suggested retention audience immediately enter a campaign?
Topic: Inspiration cards
Reveal suggested answer
First review its definition, sample membership, suitability, and campaign requirements. A suggestion is a starting point for design.
Capture a key idea, an example, or a question for your instructor.
In this section, you will learn to do the following:
Identify common challenges that AI technology addresses in contact center agent workflows.
Compare key performance metrics before and after AI agent assistance implementation.
Apply Amazon Q in Connect features to resolve specific agent productivity challenges.
Evaluate the quality of AI-generated summaries against documentation best practices.
Section introduction
Have you ever called customer service and been put on hold while the agent searches for information? Or maybe you have been an agent yourself, struggling to find answers while a customer waits impatiently on the line. AI in contact centers aims to address these frustrating experiences.Amazon Connect delivers first-party AI across all channels, with unlimited AI that remains tied only to your underlying channel usage rather than AI consumption. Organizations of any size can use AI across all touchpoints, which empowers agents to reach their full potential with an intelligent digital partner. Think of it as giving each agent a super-smart assistant that works alongside them to handle routine tasks. Agents can focus on what humans do best: solving complex problems and building relationships with customers.
Current Agent Challenges and AI Solutions
Contact center agents face numerous challenges that affect their productivity, job satisfaction, and ability to deliver excellent customer service. Agents can use Amazon Connect AI solutions to help in the following ways:
Find solutions quickly
Resolve issues consistently
Manage customer emotions
Document after contact
Finding solutions quickly
Agents often need to search through multiple knowledge bases while customers wait and assemble a solution based on the information they find. They can then present this solution to the customer. Agents need to navigate disparate systems to complete the required actions.
Figure 32 Current Agent Challenges and AI SolutionsSelect image to enlarge
Amazon Connect AI solution:
Amazon Q in Connect, a personalized real-time agent assistant, analyzes customer conversations in real time. This provides agents with relevant information and actions to address issues. Amazon Q in Connect uses your knowledge bases to provide responses. It personalizes these responses based on customer-specific information, including order history, case details, loyalty status, and more.
For example, if a customer asks about a return policy for damaged items, the agent receives the appropriate policy details without needing to search. The agent is also provided with the steps to take to resolve the issue.
Consistent issue resolution
Some customer issues require consistent handling, regardless of the experience level of the agent. These issues have a well-defined process and steps that the agent needs to complete based on the conversation.
Amazon Connect AI solution:
Amazon Q in Connect guided agent workflows recommend step-by-step guidance that seamlessly walks agents through processes that require consistency.
Figure 33 Current Agent Challenges and AI SolutionsSelect image to enlarge
For example, when getting assistance to update information on a customer's account, a step-by-step guide can be launched to complete the process directly from the agent workspace. This intuitive support ensures consistent, high-quality service delivery for every customer scenario regardless of agent tenure.
Customer emotion management
It is difficult for agents to consistently identify customer emotions and conversation trends, especially in text-based channels such as chat.
Amazon Connect AI solution:
Contact Lens provides real-time analytics and sentiment analysis. AI can detect customer emotions and alert supervisors when interactions might need special attention.
Contact Lens detects negative sentiment in real time during customer interactions. For example, it recognizes frustration in customer statements such as, "This is the third time I've contacted you, and nobody seems to care." Supervisors receive alerts about deteriorating conversations, so they can choose to join or guide agents. This proactive approach helps prevent customer loss by addressing emotional signals in conversations.
Figure 34 Current Agent Challenges and AI SolutionsSelect image to enlarge
After contact documentation
Agents typically spend 30-45 seconds documenting each call, which adds up to hours of non-customer-facing time each week.
Amazon Connect AI solution:
Contact Lens automatically generates detailed summaries of customer conversations and captures key issues, actions taken, and next steps. This reduces the time agents spend on after contact documentation. These summaries can be copied for use in case notes or automatically added to systems of record without requiring agents to take action.
For example, Contact Lens created an automatic summary when a customer called about returning a watch. The summary captured key details like the return reason and exchange steps. The agent could immediately assist the next customer instead of spending time on documentation.
Figure 35 Current Agent Challenges and AI SolutionsSelect image to enlarge
Key Metrics Improved Through AI Agent Assistance
When contact centers implement AI agent assistance tools, they typically see improvements across multiple performance metrics. Understanding these metrics helps you set realistic goals and measure success after implementation.
AI typically reduces AHT through faster information access and reduced after contact work.
AI typically improves FCR by giving agents better information and process guidance.
AI-assisted agents typically achieve higher CSAT scores.
AI-enhanced customer service can lead to improved NPS scores.
Contact centers using AI assistance often see improved speed to competency and reductions in onboarding time, training time, and agent turnover.
Key metric examples
Learn about each metric and how AI enhances the experiences of customers in the following real-world examples.
AHT
Imagine an agent helping a customer with a complex return. Instead of putting customers on hold to search multiple systems, the AI assistant works instantly. It provides return policies, purchase history, and resolution steps all in one place. This reduces the total time it takes to complete the request.
FCR
A customer calls about internet connectivity issues. The AI system guides the agent through diagnostic steps and suggests solutions based on the customer's equipment. It can also help schedule technician visits when needed. The agent resolves the customer's issue on the first attempt. This is achieved without transferring the call to a more tenured resource or requiring a follow-up.
CSAT
A customer inquires about retirement planning options. The AI system provides the agent with personalized recommendations based on the customer's age, income, and risk tolerance. When the customer gets frustrated about the forecast gap, the agent can be empathetic and recommend the best options for the customer's unique circumstances. The customer is impressed by the tailored advice and quick service, which leads to a high satisfaction rating.
NPS
A frequent flyer calls an airline about a missed connection. AI instantly recognizes the customer's elite status, provides the agent with rebooking options, and suggests offering a lounge pass as compensation. The customer is delighted by the swift, personalized service and becomes more likely to recommend the airline.
Agent performance and retention
Consider a new agent who assists with a complex billing inquiry. Instead of feeling overwhelmed, they confidently navigate the interaction with AI-powered guidance to resolve the issue successfully. This positive experience contributes to their job satisfaction and likelihood of remaining with the company.
Introduction to Agent Assistance with Amazon Q in Connect
In this section, you will learn to do the following:
Identify specific ways Amazon Q in Connect reduces agent handling time and improves first contact resolution (FCR).
Identify how to configure Amazon Q in Connect knowledge bases to address common customer inquiries.
Recognize how to resolve common knowledge retrieval issues in Amazon Q.
Evaluate the effectiveness of Amazon Q responses for different types of customer scenarios.
Agent Assistance with Amazon Q in Connect
Amazon Q in Connect is a generative AI customer service assistant that helps contact center agents resolve customer issues quickly and accurately. Amazon Q in Connect listens to customer conversations in real time and identifies helpful information. It then presents this information to agents automatically, which eliminates the need for manual searching. Unlike traditional knowledge management systems that require manual keyword searches, Amazon Q in Connect works proactively without agent initiation. It uses conversation context to find the right information and generates natural language responses.
Amazon Q in Connect uses the three stages shown in the following figure.
Figure 36 Agent Assistance with Amazon Q in ConnectSelect image to enlarge
Conversation understanding
The system uses natural language understanding (NLU) to analyze what the customer is saying during calls or chats.
Information retrieval
Based on that understanding, Amazon Q in Connect searches through your connected knowledge sources to find relevant information.
Response generation
Using information from knowledge bases and customer data, the system provides agents with suggested responses and recommended actions. It also offers links to detailed documentation and step-by-step guides for efficient task completion.
Enable Amazon Q in Connect
A Domain, also known as an Assistant, is created when enabling Amazon Q in Connect. You can then associate knowledge bases to your domain, which will be used as the source of information during conversations.
Amazon Q in Connect comes preconfigured with AI agents for different tasks and are created within a domain. These agents can assist human agents in real time or support manual queries that agents might make. Amazon Q in Connect also provides AI agents for other use cases, such as self-service. You can use these default AI agents or create your own for more control over the behavior, associated knowledge bases, guardrails, chunking strategies, and more.
Key points to consider include the following:
Multiple domains can exist, but they operate independently.
One domain can link to multiple Amazon Connect instances.
Each Amazon Connect instance can only link to one domain at a time.
A default knowledge base can be used across the domain, or you can use separate knowledge bases by configuring AI agents.
Amazon Connect instances can be reassigned to different domains as needed.
Integration with Guided Workflows to Assist Agents
Guided workflows are step-by-step instructions that walk agents through specific processes. When integrated with Amazon Q in Connect, these workflows become even more powerful, automatically appearing based on the conversation.
Amazon Q in Connect Guided Workflows
When Amazon Q in Connect detects a specific customer intent, it can trigger the appropriate guided workflow to help the agent resolve that issue. Consider the following:
The customer explains their issue.
Amazon Q in Connect detects the intent.
The system recommends a relevant guided workflow.
The agent follows the workflow steps to resolve the issue.
For example, if a customer says, "I need to dispute a charge on my credit card," Amazon Q in Connect can do the following:
Recognize the intent as a dispute transaction.
Recommend a guided workflow that walks the agent through the dispute process.
Figure 37 Integration with Guided Workflows to Assist AgentsSelect image to enlarge
Flow Configuration
To enable Amazon Q in Connect in your contact flows, add an Amazon Q in Connect block to your contact flow. This block associates the Amazon Q domain with the current contact.
For voice calls only, do the following:
Add a Set recording and analytics behavior block.
Configure it for Contact Lens conversational analytics in real time.
Place this block anywhere in the flow.
Note: Contact Lens conversational analytics is required for Amazon Q to work with voice calls but is not required for chat interactions or self-service use cases.
Figure 38 Integration with Guided Workflows to Assist AgentsSelect image to enlarge
Providing Agent Access
Agents need access to Amazon Q in Connect to take advantage of its benefits during calls. The Admin security profile already includes all Amazon Q permissions by default. For more information about security profiles, see Security profile permissions for Amazon Q in Connect.
For agents to access and use Amazon Q in Connect, assign the following permissions in their security profile:
Amazon Q - Access – Agents can search for and view content. They can also receive automatic recommendations during calls if Contact Lens conversational analytics is enabled.
Custom views - Access – This is required if using the step-by-step guides integration.
Agent access to Amazon Q in Connect
Amazon Q in Connect offers two access methods for agents: directly through the Agent Workspace or through embedded integration in custom applications.
Amazon Connect Agent Workspace
If you are using the agent application provided with Amazon Connect, after you enable Amazon Q in Connect, share the following URL with your agents so they can access it:
Figure 40 Amazon Connect Agent WorkspaceSelect image to enlarge
Introduction to Enhancing Customer Service with AI Powered Tools
In this section, you will learn to do the following:
Identify key indicators in real-time transcription that require immediate agent action.
Differentiate between types of customer sentiment signals and select appropriate response strategies.
Apply Contact Lens AI summarization features to standardize post-call documentation.
Evaluate when to use different AI-powered tools based on specific customer interaction scenarios.
Real time Transcription Seeing What You Hear
Real-time transcription converts speech to text as the conversation happens. When a customer calls in, Amazon Connect Contact Lens analyzes the audio stream immediately and converts spoken words into text almost instantly. Real-time transcription with Amazon Connect Contact Lens offers significant advantages for customer service operations.
Let us explore how this capability enhances contact center effectiveness:
Capturing complex information
When a customer provides detailed information like reference numbers or addresses, you do not need to ask them to repeat information.
Agents can review the complete transcript during After Contact Work to verify details or use in back office operations. Reducing errors in documentation and follow-up actions. Supervisors can monitor calls more effectively without listening to entire conversations.
Providing accurate records of complex customer details, eliminating reliance on agent memory or incomplete notes.
Avoiding misunderstandings
If an agent is unable to recall what a customer has said, they can refer back to the transcript to check. Reducing the need to make an outbound contact asking them to repeat.
Helping clarify confusing statements or terminology that might otherwise lead to errors.
Handling accents and difficult audio
Sometimes phone connections are poor or accents make understanding difficult. Transcription helps bridge these communication gaps.
Improving comprehension for agents reviewing difficult conversations.
Focusing on the conversation
Instead of taking notes, you can focus on active listening while knowing that the transcript will capture the details. The transcript can then be summarized automatically and available at the end of the contact.
Creating more engaged customer interactions without worry about missing details.
Learning from past segments
While on a chat contact, if you need to reference something the customer mentioned earlier, you can scroll up in the transcript. For a voice contacts, agents can review the complete transcript during After Contact Work to reference information from earlier in the call. You can analyze conversation flow to identify patterns or missed information from completed transcripts.
Allowing more informed follow-up actions and improvements to customer journeys
Demystifying the conversation
Do your agents or customers commonly use acronyms, jargon, or other industry specific terms? It's common to have annual percentage rate said as APR, or explanation of benefits said as EOB. Custom vocabulary enables transcription of full words or phrases. This feature makes it easier for agents, supervisors, and quality analysts to better understand the complete context of conversations.
Specialized terminology appears accurately in documentation for improved training and analysis.
Supervisor assistance
Supervisors can access completed transcripts to provide targeted feedback and coaching to agents. This ensures guidance focuses on actual conversation details rather than recalled information.
Providing more effective training, when based on accurate records of customer interactions.
Best practices
The following are some best practices when using real-time transcription:
Do not rely on it exclusively – While transcription is remarkably accurate, it's not perfect. Use it as a tool to complement your listening skills, not replace them.
Verify critical information – For important details like credit card numbers or medical information, always verify directly with the customer.
Use it for follow-up – Reference the transcript to ensure you've addressed all the customer's concerns before ending the call.
Check for errors – Occasionally review the transcript for any obvious errors, especially with unusual names or technical terms.
Sentiment Analysis Understanding the Emotional Journey
Amazon Connect Contact Lens analyzes the sentiment of both the customer and the agent in a conversation as positive, negative, or neutral. It then considers the following two factors for each participant to assign a score that ranges from -5 to +5 for each period of the call:
Frequency – The number of times the sentiment is positive, negative, or neutral.
Sentiment streaks – The consecutive turns with same sentiment.
The overall sentiment score is the average of the scores assigned during each portion of the call.
Figure 41 Sentiment Analysis Understanding the Emotional JourneySelect image to enlarge
Agent Nikki handles a call from an initially frustrated customer. Amazon Connect Contact Lens detects negative sentiment at the start of the call. However, by midway through the call, the sentiment shifts to neutral and ends positive. This showcases Nikki's effectiveness at turning around negative situations.
Review the sentiment trend in the following figure and the explanations for its numbered points.
Figure 42 Sentiment Analysis Understanding the Emotional JourneySelect image to enlarge
Call start
Customer calls with an issue about their account.
Midpoint
Customer is frustrated at not being able to resolve her issue.
Resolution
The agent was able to resolve the customer's issue.
Completion
The customer is pleased with the outcome.
With sentiment analysis, you can see shifts happening in real time and adjust your approach accordingly. This can mean the difference between losing a customer or regaining a customer's trust. If this information was only available after the call completed, it could be too late to regain the customer's trust.
Post Contact Summarization with Amazon Connect Contact Lens
After contact work (ACW) significantly impacts agent productivity. Trying to remember all the details of a call while the next customer is waiting is stressful. Streamlining these post-interaction tasks can save your contact center substantial operational hours throughout the year, which leads to improved efficiency and reduced costs.
Common ACW Tasks
ACW typically involves the following:
Summarize what the customer needed.
Document what actions the agent took.
Record any follow-up tasks.
This documentation is crucial for several reasons. It helps agents who might handle the customer in the future, provides valuable data to identify trends, and creates an audit trail for compliance purposes.
The next section describes common after contact work challenges.
Common ACW Challenges
Agents may face the following common ACW challenges:
Forget important details from the beginning of the call.
Spend too much time writing comprehensive notes.
Create inconsistent documentation that varies from agent to agent.
Feel rushed and make documentation errors.
Now that you have reviewed common ACW challenges, move on to the remaining content.
Contact Lens summarization process
Amazon Connect Contact Lens uses generative-AI to automatically generate summaries of customer interactions. Think of it as having an assistant who listens to your calls and creates notes for you and then gives you an elevator pitch to what happened.
Explore how this works behind the scenes.
Recording and transcription
The system captures the audio from both sides (agent and customer) of the conversation and converts it to text.
Natural language processing
AI analyzes the transcript to identify key elements like main issues, actions taken, outcomes, and follow-up items.
Summary generation
Using these identified elements, AI creates a concise, structured summary.
For example, instead of reading through a 10-minute transcript, you might see a summary like: "Customer called about missing reimbursement for canceled flight. Agent verified eligibility and processed $250 refund. Customer will receive email confirmation within 48 hours."
Using Contact Lens summarization
The following figure and numbered explanations describe how Contact Lens summarization supports daily operations.
Figure 43 Post Contact Summarization with Amazon Connect Contact LensSelect image to enlarge
During the call
Focus entirely on helping your customer. You don't need to take extensive notes or worry about remembering every detail.
After the call ends
While you are in ACW mode, the system starts generating a summary.
Review
When the summary is ready (usually within seconds), it will appear ready to review.
Once you have reviewed the summary, you can close the contact and move on to your next customer. This is much faster than writing notes from the beginning. The summaries follow a consistent format that highlights issues, outcomes, and action items, which makes it easier for anyone reviewing the contact later.
Bringing It All Together The AI Enhanced Agent Experience
Call begins
As soon as the customer starts speaking, real-time transcription begins capturing their words, and sentiment analysis starts tracking their emotional state.
Issue identification
Amazon Q in Connect uses the transcript to provide real-time assistance, including solutions and next best actions. Contact Lens automatically identifies the issue for automated disposition and analytics.
Personalization
You are alerted to the negative customer sentiment and adjust your tone and approach to be more empathetic.
Resolution
As you work toward a solution, you can see sentiment improving in real time, which confirms you're on the right track
Documentation
After the call, the transcript and sentiment analysis contribute to an AI-generated post-contact summary. This streamlines your ACW.
Real-time sentiment analysis and AI suggestions transformed this standard billing dispute into an opportunity for exceptional service. When an agent receives automated guidance and early warning signals of customer dissatisfaction, they can prevent escalations and deliver faster resolutions.
This technology-driven approach reduces operational costs while boosting both customer satisfaction and agent confidence.
RECALL & APPLY
Practice what you learned
Try answering in your own words, then compare with the suggested answer. These activities are for self-study and are not automatically graded.
01Which agent activity should AI assistance improve first in your environment?
Topic: AI assistance for contact center agents
Reveal suggested answer
Accept a defined workflow problem with an observable baseline, such as time spent searching or incomplete case notes.
02How do these capabilities support an inexperienced agent?
Topic: Agent challenges and AI support
Reveal suggested answer
They reduce information gathering and procedural uncertainty while the agent remains responsible for appropriate customer handling.
03Which measures would you compare before and after deployment?
Topic: Measuring agent assistance
Reveal suggested answer
Use a consistent baseline and comparable contact types. Review efficiency, resolution, customer experience, and agent outcomes together.
04What could cause an assistant to give an irrelevant recommendation?
Topic: Amazon Q in Connect assistance
Reveal suggested answer
The detected issue, retrieved source, knowledge quality, or configuration may be wrong. Investigate those stages separately.
05Why does the domain association matter during troubleshooting?
Topic: Domains, knowledge bases, and AI agents
Reveal suggested answer
The contact must use the intended assistant configuration and knowledge sources. The wrong association can produce missing or irrelevant assistance.
06What should a voice demonstration verify before testing recommendations?
Topic: Guided workflows and flow configuration
Reveal suggested answer
Verify the relevant recording and analytics behavior, Amazon Q block, domain association, and the agent’s access.
07The feature is enabled but an agent cannot see it. What do you check?
Topic: Agent access and workspace
Reveal suggested answer
Check the agent’s security profile, workspace or embedded application configuration, and the assistant association.
08Should an agent accept every transcribed account number without checking?
Topic: Real time transcription
Reveal suggested answer
No. Speech recognition can make errors. Confirm critical identifiers through the organization’s established process.
09Why should a supervisor inspect the conversation behind a sentiment score?
Topic: Sentiment across a conversation
Reveal suggested answer
The score is an estimate. The surrounding language and issue context are needed to decide whether and how to intervene.
10What should an agent verify before using a generated summary?
Topic: Post contact summaries
Reveal suggested answer
Verify the customer issue, actions, outcome, commitments, and any follow up. Correct omissions or unsupported statements.
11Describe an effective handoff from AI assistance to agent judgment.
Topic: The assisted agent journey
Reveal suggested answer
The assistant supplies relevant facts and guidance. The agent checks applicability, responds with empathy, and decides how to handle exceptions.
Capture a key idea, an example, or a question for your instructor.
Introduction to AI Powered Supervisor Capabilities
Supervisors Role and AI Technologies
AI's role for supervisors
AI in Amazon Connect Customer combines advanced technologies that analyze customer interactions in real-time and post-contact. Unlike traditional tools that simply record calls or track metrics, AI-powered solutions can actually understand conversations and identify important patterns. For example, imagine a customer calling about a billing issue. Traditional tools might just record the call and track how long it lasted. Amazon Connect has built-in AI that can detect payment disputes and rising customer frustration, alerting supervisors before issues escalate.
Amazon Connect Customer uses the following key AI capabilities to support supervisors.
Core AI technologies in Amazon Connect Customer
Review the following core AI technologies in Amazon Connect Customer.
Generative AI
Technology that can create new content based on patterns learned from training data.
Generative AI can produce human-like text summaries, categorizations, and other content derived from customer interactions.
For example, a customer has a 15-minute call with an insurance company discussing a complex claim involving water damage, contractor estimates, and coverage questions. After the call, generative AI analyzes the conversation and creates a concise summary with key issues and actions. The customer care agent reviews the summary, and makes necessary edits. The AI system files it while drafting follow-ups that may be needed later.
Natural language processing
Natural language processing (NLP) allows computers to understand and meaningfully respond to human language through artificial intelligence.
NLP bridges the gap between human communication and computer understanding by analyzing text or speech.
For example, a customer calls a telecom company's support line and says, "My internet keeps dropping every evening." The contact center's NLP system identifies service disruption as the intent, identifies evening as the time, routes to tech support, and retrieves account data for the support agent.
Natural language understanding
Natural language understanding (NLU) is a specialized subset of NLP. NLU technology is focused specifically on comprehending the meaning and intent behind language.
NLP handles the overall processing of language, whereas NLU drills deeper into extracting specific intents, entities, and meaning.
For example, a customer calls a telecom company's support line and says, "My internet isn't working right when I need it most." The NLU system interprets vague complaints as service issues, infers evening from context, detects unstated frustration, and prioritizes routing accordingly.
Automatic speech recognition
Automatic speech recognition (ASR) converts spoken language into text. ASR uses acoustic modelling to transcribe speech accurately. It handles various accents and filters background noise.
ASR enables real-time call transcription, making voice conversations searchable and analyzable. Supervisors can quickly scan conversation content without listening to recordings. This technology supports sentiment analysis on voice channels.
For example, a customer calls an airline's support line and says, "I need to change my flight to Dallas on Thursday." The contact center's ASR system accurately transcribes the spoken words into text in real-time, despite the customer's accent and background noise from an airport. This transcription instantly reaches the agent and booking system, which searches for Dallas Thursday flights before the agent responds.
Sentiment analysis
This technology analyzes transcribed text using NLP to detect positive, negative, or neutral sentiment in customer and agent conversations, helping supervisors identify interactions that might need attention.
For voice calls, the system first transcribes speech to text, then NLP classifies each speaker turn as positive, negative, or neutral based on the words used. For chat, text is analyzed directly. A score between -5 (most negative) to +5 (most positive) is generated for each portion, with the overall score calculated as the average across all portions.
For example, a customer calls a bank's support line saying, "I've been trying to access my account all day and nobody has helped me." The sentiment analysis system analyzes the transcribed text using NLP and classifies the turn as negative based on the language patterns. The system flags the interaction as having negative sentiment, alerting supervisors.
Speech analytics
Speech analytics analyzes voice data for insights beyond basic transcription. Speech analytics is broader than sentiment analysis, evaluating aspects like tone, pace, non-talk time, interruptions, and conversation dynamics.
Sentiment analysis and speech analytics are different but related technologies. Sentiment analysis uses NLP on transcribed text to classify each speaker turn as positive, negative, or neutral based on language patterns.
For example, a customer calls a healthcare provider about a billing issue. During the call, speech analytics detects extended silences, faster customer speaking, and cross talk. The system flags this call for review, revealing communication issues and suggesting agent training needs.
Text analytics
Text analytics derives meaningful patterns and insights from written text. Text analytics identifies topics, themes, and anomalies across text-based interactions.
For example, the text analytics system scans thousands of weekend customer chats and detects a 400 percent rise in checkout error messages. The system alerts IT to investigate payment processing issues before social media complaints begin. The system also notifies managers to prepare response strategies for this emerging problem.
Knowledge retrieval and question answering
Knowledge retrieval systems locate information from knowledge bases. Question answering tools extract specific answers from content. Both technologies respond intelligently to natural language queries.
For example, a customer calls a travel agency asking, "What's your cancellation policy for Mediterranean cruises during hurricane season?" The knowledge retrieval system quickly analyzes the question and displays policy information on the agent's screen. The agent provides immediate, accurate details without putting the customer on hold.
PII detection and redaction
This technology identifies and can remove sensitive personally identifiable information (PII) from text and recordings.
These algorithms recognize patterns associated with sensitive information like credit card numbers or addresses.
For example, a banking customer sends a chat message: "I can't access my account. My username is jstiles1980, my password is Spring2023!, and my account number is 1111-2222-3333.
ML for pattern recognition
This technology uses ML to identify patterns in customer interactions and contact center data.
These algorithms learn from historical data to recognize trends, themes, and clusters across conversations.
For example, a credit card company's ML system analyzes millions of service interactions over the most recent six-month period. The system discovers customers interactions that discuss international travel without mentioning transaction fees often result with customers calling back with disputes. The contact center creates a new protocol for agents to explain these fees during travel conversations. This change reduces follow-up calls and boosts customer satisfaction scores.
AI tools in Amazon Connect Customer do not replace effective supervisors, they make them more efficient. These technologies handle routine tasks and spot important patterns automatically. Supervisors can now focus on what matters most for your business.
The most effective contact centers combine human skills with AI capabilities. This partnership helps teams respond faster and understand customers better. As tech evolves, one constant remains: when skilled people use intelligent tools, everyone wins.
Shifting from Reactive to Proactive Monitoring
Quality Monitoring
Traditional quality management and monitoring are largely reactive. Supervisors typically review a small percentage of calls after they occur or respond when an agent asks for help. AI transforms this model by enabling a proactive approach with real-time monitoring, alerts, and data driven insights.
For example, without AI, you might discover a compliance issue days after it occurred. With Amazon Connect Customer AI capabilities such as Conversational Analytics, you can receive an alert the moment an agent forgets to read a required disclosure. This facilitates immediate issue resolution. Another is the ability to gain insights across the entire contact center while not requiring the manual evaluation of contacts to see trends.
With traditional supervision, interventions happen only after issues are identified through review, potentially too late to change the outcome.
AI-powered supervision allows real-time monitoring. Supervisors can detect issues during calls before they escalate. This prevents discovering problems through post-call reviews when intervention is too late. The key difference is the sequence. AI detects and alerts sooner than manual review therefore allowing earlier action.
Figure 44 Quality MonitoringSelect image to enlarge
Amazon Connect Customer AI tools eliminate traditional limitations by doing the following:
Simultaneously analyzing every customer interaction
Automatically identifying patterns across thousands of conversations
Surfacing the specific interactions that need human attention
Benefits of AI to Supervisors
Amazon Connect Customer AI capabilities streamline supervisor tasks by automating contact monitoring and analysis. This transforms traditional quality management processes that relied on manual call reviews and report creation.
These efficiencies translate directly to cost savings because of the following:
Reduced time spent on routine monitoring (from hours to minutes)
More focused effort on coaching and agent development
Faster identification and resolution of emerging customer trends and issues
Enhanced quality management
AI transforms quality management by automating routine evaluations, freeing supervisors to invest their expertise where it matters most.
Complete interaction coverage: Instead of reviewing a random 2–3 percent sample of calls, AI can analyze 100 percent of interactions, making sure that no critical issues are missed.
Consistent measurement: AI can provide consistent analysis across all interactions, eliminating human bias in evaluations.
Automated scoring: Many aspects of quality evaluation can be automated, such as compliance statement detection or greeting verification.
Real time intervention possibilities
Perhaps the most transformative benefit is the shift from after-the-fact review to real-time intervention.
Live alerts: Supervisors can receive notifications when AI detects issues like customer frustration, compliance risks, or agents struggling.
In-call assistance: When alerted to an issue, supervisors can provide real-time guidance through chat or call barging.
Proactive intervention: AI can detect which calls might become problematic before issues escalate by detecting changes in customer sentiment.
Voice of the customer: Amazon Connect Customer AI can transform routine customer calls into valuable voice-of-customer data. It automatically identifies common frustration points across thousands of conversations. Amazon Connect Customer can categorize these points by key features such as product or service type and provide sentiment assessment based on those categories.
Marketing insights: AI analytics reveal unmet customer needs before they appear in formal research. Amazon Connect Customer can highlight customer confusion around specific product terms or benefits. These insights help refine campaign targeting based on real customer pain points.
Strategic advantage: The AI analytics in Amazon Connect Customer mines valuable business intelligence from every customer conversation. This tool uncovers competitor mentions that agents might miss. It reveals genuine customer sentiment about pricing strategies. The system exposes emerging market opportunities hidden within daily interactions.
Without AI, these insights would remain buried in conversation data.
For example, a supervisor might spend 15 hours weekly reviewing a limited sample of calls. With automation analyzing 100 percent of interactions, the supervisor can now focus on strategic coaching to enhance customer experience metrics. This helps them make data-driven decisions based on comprehensive insights rather than limited snapshots.
Introduction to Conversational Analytics
Amazon Connect Customer Conversational Analytics
Conversational Analytics overview
Amazon Connect Customer Conversational Analytics is an AI-powered analytics solution built into Amazon Connect Customer. It uses machine learning to analyze conversations between customers and agents across voice and chat channels, providing insights that would be impossible to gather manually at scale. At its core, Conversational Analytics acts like an extremely attentive listener that can process thousands of conversations simultaneously. Conversational Analytics transcribes speech to text and analyzes the language to understand meaning and identify important patterns.
Figure 45 Conversational Analytics overviewSelect image to enlarge
AI generated contact summary
Contact Lens uses generative AI to create concise summaries of customer interactions.
Conversational analytics
Customer sentiment trend: Shows how customer sentiment changes as the contact progresses
Customer sentiment: Shows the distribution of customer sentiment for the entire call
Talk time: Shows the distribution of talk time and non-talk time during the entire call—talk time is further split into agent and customer talk time
Audio analysis
Interactions can be recorded and are split into Customer, Agent, or System/Bot. Customer and Agent sentiment are marked in the timeline of the recording.
Transcript Key highlights and categories
Transcript automatically identifies and labels key parts of customer conversations and displays highlights of the conversations. Managers can view those highlights on the Contact details page.
Contact Lens rules let you automatically categorize contacts, receive alerts, or generate tasks based on keywords.
How Conversational Analytics works with Amazon Connect Customer
Conversational analytics works in two primary modes:
Post-contact analysis: After a conversation ends, Conversational Analytics processes the full interaction to provide comprehensive insights, transcripts, and analytics.
Real-time analysis: During live conversations, Conversational Analytics can analyze the interaction as it happens, alerting supervisors to potential issues that might require immediate intervention.
Conversational Analytics features
The following capabilities work across multiple channels (voice calls, chat conversations, and more), giving you a comprehensive view of all customer interactions:
Conversational Analytics with AI-powered summaries, transcription, and sentiment analysis
Performance evaluation with automated assessments and calibration tools
Real-time monitoring capabilities with supervisor alerts and screen recording
Search, recording, and supervisor intervention tools
Enabling Contact Lens in an Amazon Connect Customer instance
The following demonstration steps help you understand how to enable Contact Lens in a Connect Customer instance.
Welcome to this demonstration on enabling Contact Lens in an Amazon Connect instance.
To start, sign in to the AWS Management Console.
In the search box at the top of the page, enter Amazon Connect. In the displayed list of Services, choose Amazon Connect.
On the Amazon Connect page, the instances in your account are displayed. Choose the Instance alias of the instance that you want to enable Contact Lens on.
Next, in the navigation pane on the left of the screen. Under Applications, choose Analytics tools.
On the Analytics tools page, notice the state of Contact Lens. In this example, Contact Lens is not enabled. To access the settings for Contact Lens, choose Edit.
To enable Contact Lens, select the checkbox. Then choose Save.
On the Analytics tools page, a success alert indicates your instance features have been updated. Notice that the status for Contact Lens has changed to Enabled.
This concludes the demonstration on enabling Contact Lens in an Amazon Connect Customer instance.
Thank you for your participation.
Conversational Analytic Security Features
Conversational Analytics incorporates comprehensive security protocols designed specifically for sensitive contact center environments. Conversational Analytics safeguards customer conversations through the following protective layers that work seamlessly together:
Sensitive data redaction: This feature identifies and removes sensitive data like credit card numbers, names, and other PII information.
Data encryption: All data is encrypted both in transit and at rest.
Retention controls: Organizations can set data storage policies.
Insider tip: Ensure that you configure access permissions through security profiles to determine which users can access Contact Lens insights. To learn more about security profiles, see Security Profiles for Connect Customer and Contact Control Panel (CCP) Access.
These security features help ensure that organizations in regulated industries can confidently deploy Amazon Connect Customer while maintaining their compliance obligations and protecting sensitive customer information.
Sentiment Analysis, Issue Detection, and Summarization
Understanding sentiment analysis
Sentiment analysis is like having an emotional thermometer for every conversation. It automatically detects the sentiment of interactions, helping you respond appropriately before issues escalate. Conversational Analytics analyzes transcribed text using NLP to determine whether sentiment is positive, neutral, or negative based on the language patterns in each speaker turn. The system tracks sentiment throughout the conversation, showing how it changes over time.
Figure 46 Sentiment Analysis, Issue Detection, and SummarizationSelect image to enlarge
Sentiment analysis example
Sentiment analysis estimates the sentiment of what the customer and agent are saying throughout the interaction. This metric is represented as both a quantitative value (with a range from -5 to +5) and a qualitative value (positive, neutral, mixed, or negative). Scores are based on language patterns in the transcribed text, weighted by both the frequency of positive/negative turns and sentiment streaks (consecutive turns with the same sentiment).
Real time sentiment alerts
Conversational Analytics sends immediate notifications when a conversation's sentiment turns negative, empowering managers to join in-progress contacts and help resolve issues before they escalate.
For example, when a customer chat shows rapidly deteriorating sentiment, a supervisor can receive an alert and review the transcript. The supervisor can then either coach the agent privately or join the conversation directly.
Figure 47 Sentiment Analysis, Issue Detection, and SummarizationSelect image to enlarge
Using sentiment data for coaching
Sentiment analysis provides valuable data for agent coaching and training. By setting up agent alerts, you can use real-time sentiment data for coaching in the following ways:
Identify which agents consistently maintain positive customer sentiment.
Recognize conversation patterns that tend to improve sentiment.
Spot common triggers that lead to negative sentiment.
Create training scenarios based on real interactions that show sentiment shifts.
Issue detection
Issue detection works like an early warning system, automatically flagging potential problems before they become serious.
Conversational Analytics analyzes conversations in real-time, looking for patterns that indicate potential problems including the following:
Technical difficulties described by customers
Requests for supervisor escalation
Mentions of cancellations or refunds
Specific product problems or recurring complaints
Expressions of significant frustration
Real time issue alerts
When Conversational Analytics detects a potential issue, it can do the following:
Send alerts to supervisors.
Categorize the issue type automatically.
Provide relevant context for whoever steps in to help through Tasks.
Trend analysis for systemic issues
Beyond individual contacts, issue detection also consolidates data to help identify systemic problems and trends. Issue groupings include the following:
Product defects affecting multiple customers
Website or app functionality problems
Confusing policies or processes
Training gaps among agents
Seasonal or campaign-related issues
Summarization
Conversational Analytics uses generative AI to create concise summaries of customer interactions. These summaries highlight the following:
Main issues discussed
Actions taken by the agent
Outcomes of the interaction
Follow-up items or commitments
Key customer information or requests
Figure 48 Sentiment Analysis, Issue Detection, and SummarizationSelect image to enlarge
Benefits of automated summarization
Time efficiency: Understand contact outcomes without reviewing entire transcripts or recordings
Structured insights: Highlight customer issues, agent actions, and interaction outcomes
Objective representation: Focus on facts instead of subjective interpretations
Performance tracking: Identify common customer issues and agent coaching opportunities
Compliance Monitoring
Compliance monitoring overview
Conversational Analytics uses advanced speech and text analysis that you can use to identify contacts for compliance with various requirements. Compliance monitoring provides peace of mind in a complex regulatory landscape.
By automatically checking every interaction against compliance requirements, Conversational Analytics helps organizations reduce risk, protect customer information, and demonstrate their commitment to regulatory standards. This proactive approach not only helps avoid penalties but also builds customer trust through consistent, compliant interactions.
Common compliance examples are as follows:
Required disclosures and statements
Prohibited language or topics
Sensitive data collection processes
Authentication procedures
Required information collection
For example, in a financial services contact center, agents might be required to verify identity using specific questions and inform customers about call recording. conversational analytics can automatically check whether actions were completed in each interaction.
Sensitive data protection
Conversational Analytics also helps protect sensitive customer information by automatic redaction of sensitive personal information from call recordings and transcripts.
Beyond checking for required actions, compliance monitoring also helps protect sensitive customer information with the following features:
Automatic redaction: Conversational Analytics can automatically redact sensitive personal information from transcripts and recordings.
PII detection: The system identifies when personally identifiable information is shared.
Compliance reporting: Aggregate data shows compliance rates across agents and teams.
This protection works for both recorded calls and chat transcripts, helping to ensure sensitive information is handled appropriately across all channels.
For example, Conversational Analytics can automatically redact credit card numbers spoken by customers during calls from recordings and transcripts. This feature helps you maintain payment card industry (PCI) compliance.
Compliance consideration
Although the Conversational Analytics redaction feature uses sophisticated technology to protect sensitive information, human verification remains essential to validate complete compliance. The following are examples of situations they might require verification:
Review redacted output
Because of the predictive nature of machine learning, Conversational Analytics might not identify and remove all instances of sensitive data in a generated transcript. You should review any redacted output to ensure the transcript meets your compliance requirements. For more information, see Use sensitive data redaction to protect customer privacy using Contact Lens.
Scrubbing PII or PCI
If PII or PCI data is captured in call recordings, the sensitive data must be scrubbed from the recording and obfuscated from any logs or transcriptions. For more information, see Best practices for PCI compliance in Amazon Connect.
Conversational Analytics Expanded Channels and AI Interactions
Conversational Analytics for Email
The Conversational Analytics capabilities you learned about now extend to additional areas. For the email channel, categorization, PII redaction, and summarization are available. For AI agent (bot) interactions, sentiment analysis, summarization, and compliance monitoring apply.
Conversational Analytics automatically processes email interactions alongside voice and chat. This unifies analytics across all customer communication channels.
What it does
For every email interaction, the system automatically:
Categorizes the email by topic and intent
Redacts PII (personally identifiable information) from stored records
Generates contact summaries so supervisors can quickly understand the interaction without reading entire email threads
This means your analytics workflows for categorization, summarization, and PII redaction now apply to email alongside voice and chat.
Why it matters for supervisors
Email interactions are often longer and harder to review than a quick chat or call. A single email thread might span multiple days and dozens of exchanges. AI-generated summaries make these reviewable at a glance, while automatic categorization helps identify trends across the email channel.
PII redaction ensures that stored analytics data remains compliant with privacy requirements, even when customers include sensitive information in their emails.
Conversational Analytics for AI Agents
Conversational Analytics also covers self-service interactions including conversations handled by AI agents across voice, chat, and messaging channels.
This is a fundamental shift. Previously, analytics focused exclusively on human agent performance. Now supervisors can apply the same analytical lens to AI agent interactions:
Sentiment analysis for bot conversations — are customers getting frustrated with the AI?
Redaction — PII is protected in bot interactions just as it is in human ones
Issue identification — where are AI agents struggling or generating poor outcomes?
This gives supervisors a unified view of the entire customer experience, regardless of whether a human or AI agent handled the interaction. Patterns become visible: if customers consistently express negative sentiment during a specific type of AI agent interaction, that signals a tuning opportunity.
AI powered case summaries
Supervisors can also generate concise summaries that span multiple interactions for a single case or customer issue.
A customer might contact your organization three times about the same billing dispute:
Figure 49 Conversational Analytics for EmailSelect image to enlarge
Previously, a supervisor reviewing this case would need to read transcripts from all three interactions individually.
With AI-powered case summaries, supervisors generate a single concise summary covering all related interactions with one click. The summary captures the key facts, actions taken, and current status across the full history of the case.
This streamlines review workflows and helps supervisors quickly identify cases that need attention without spending time reading through multiple full transcripts.
Introduction to Real Time Analytics in Contact Centers
Understanding Real Time Analytics
Real time analytics overview
Real-time analytics refers to the collection, processing, and analysis of data as it is being generated. In contact centers, this means having access to up-to-the-minute information about customer interactions, agent performance, queue status, and overall operational metrics.
Unlike historical reporting that looks at what happened yesterday or last week, real-time analytics focuses on what is happening now. It is like the difference between watching a live sports game compared to reading about it in tomorrow's newspaper.
Key components of real time analytics
Review the following key components of real-time analytics:
Real-time dashboards
Dashboards transform complex data into understandable charts, graphs, and status indicators that update continuously.
A good dashboard is like the cockpit of an airplane. The dashboard shows you all the critical information you need to manage your contact center safely. Managers can quickly see if service levels are on target, if queues are growing too large, or if too many agents are unavailable.
Figure 50 Key components of real time analyticsSelect image to enlarge
Data collection systems
Real-time analytics begins with systems that capture information as it happens. These systems track every customer interaction across channels like phone, chat, email, and task. They monitor how long customers wait, what agents are doing, and the status of every interaction.
Amazon Connect Customer automatically collects all these data points without requiring any additional configuration. From the moment your instance is set up, the built-in telemetry of Amazon Connect Customer continuously gathers metrics on contacts, queues, and agent activities. These metrics are immediately available for real-time monitoring and analysis.
Think of it like sensors placed throughout your contact center, constantly feeding information to Amazon Connect Customer. These sensors might track when calls come in, how quickly they are answered, and what is happening during the conversation.
Alert systems
Real-time analytics become truly powerful when combined with alert systems that notify the right people when metrics fall outside acceptable ranges.
For example, if customer wait times suddenly spike beyond 5 minutes, the system could automatically send text messages to team leads. The system can also display visual alerts on the dashboard. You do not need to constantly monitor it, but you will know immediately when there is a problem.
Integration capabilities
Amazon Connect Customer real-time analytics seamlessly integrate with both AWS services and external systems to create a comprehensive supervisor system.
Real-time analytics directly connect with the Amazon Connect Customer Forecasting, Capacity Planning, and Scheduling capabilities.
Supervisors can use live metrics like occupancy rates and service levels to make immediate staffing adjustments through the built-in scheduling tool. This helps to ensure optimal coverage during unexpected contact volume fluctuations.
Amazon Connect Customer real-time metrics can be accessed through APIs and streamed through Amazon Kinesis. This allows integration with third-party workforce management tools, CRM platforms, and business intelligence systems.
For example, when real-time analytics detect an unusual spike in contacts about a specific issue, this data can trigger automated workflows in external systems.
Importance of real time analytics
With real-time analytics, supervisors can quickly respond to issues and make positive changes, such as the following:
Immediate problem resolution
When issues arise, every minute counts. Real-time analytics helps supervisors to identify problems instantly, such as system errors, call quality issues, or service disruptions. Then, supervisors can implement immediate corrective actions.
Dynamic resource allocation
Contact centers often experience unexpected surges in volume. With Real-time analytics, supervisors anticipate needs and shift resources where they are needed most, preventing potential bottlenecks before they impact customer experience.
Improved customer experience
When supervisors have visibility into what is happening right now, they can take steps to improve the customer experience immediately.
Agent support and coaching
Common mistakes to avoid with real time metrics
Even with powerful tools, organizations often make mistakes when implementing real-time analytics. Examples of these mistakes include the following:
Information overload
Trying to monitor too many metrics at once can lead to analysis paralysis. Focus on the vital few metrics that drive immediate action, not every possible data point.
Failing to act on insights
Real-time data is only valuable if it drives real-time action. Some organizations invest in sophisticated analytics tools but do not empower their supervisors to make quick adjustments based on what they see.
Neglecting training
Managers and supervisors need proper training to interpret real-time data and understand which actions are most appropriate for different situations.
Real Time Analytics Tools and Features
Real time dashboards
Have you ever watched a busy restaurant kitchen during the dinner rush? The head chef constantly monitors multiple cooking stations, checks food quality, manages timing, and coordinates the staff, all simultaneously. Contact center managers face a similar challenge, and like that chef, they need specialized tools to help them succeed.
The dashboard brings real-time analytics to life, serving as your contact center's command center for monitoring current activity.
Figure 51 Real time dashboardsSelect image to enlarge
Channel filter
To view metrics with a specific channel or channels, add a filter to limit the data shown in the dashboard.
Combine data
View combined graphs of data to give a more complete picture of what is happening in your contact center.
Customize view
Modify metrics displayed in each widget for a more customized view of your contact center.
Real-time dashboards include the following:
Channel specific metrics
Effective dashboards show metrics for channels, such as phone, chat, or email. These channels might highlight to indicate when one channel is performing significantly worse than others.
Queue status displays
These displays show how many customers are waiting, how long they have been waiting, and which service they are waiting for. This information helps managers make immediate staffing adjustments.
Customizable views
With advanced dashboards, users create personalized views based on their role and responsibilities.
Queue and agent performance dashboard
The Queue and agent performance dashboard display the real-time state for each agent. This is shown as Activity on the dashboard. Timers track duration in each state against performance thresholds. The dashboard has the ability to set color-coded indicators to show green, yellow or red custom threshold.
Service level report
Service level appears as a percentage of contacts answered within target time frames. For example, answering 80 percent of calls within 30 seconds would show as 80 percent for SL30. Time frames, queues and the performance indicator colors for Service Level (SL) metrics can all be configured on the report. In the report below you can see that only 1 percent of contacts have been answered with 15 seconds (SL15) for the AnyCompany Finance queue.
Real time alerts
Although dashboards provide visibility, alert systems help ensure that you do not miss critical changes even when you are not actively monitoring the dashboard.
Threshold based alerts
Trend based alerts
Anomaly detection
Alert delivery methods
Real time monitoring and intervention
Real-time monitoring and intervention is the ability to monitor individual interactions as they happen and intervene when necessary.
Live contact monitoring
Monitoring options
Speech and sentiment analysis
Supervisors can listen to live calls or read live chats to assess quality, identify training needs, or step in to assist with difficult situations.
Integrating Real Time Analytics with Other Systems
The effectiveness of real-time analytics multiplies when integrated with the following systems:
Workforce management systems
When real-time analytics shows high call volumes, workforce management integration can automatically identify available staff to call in or agents who could work longer shifts.
CRM integration
Real-time analytics can pull customer information from CRM systems to provide context for current interactions. For example, a simple dashboard might only show that call volumes are high. With a CRM integration, you might find that 80 percent of those calls come from high-value customers.
Quality management integration
When real-time analytics identifies potential quality issues in customer interactions, it can automatically flag these interactions for later review by the quality team.
Automated response systems
Real-time analytics can trigger automated system responses without requiring human intervention. For example, when wait times exceed thresholds, systems can automatically offer callbacks and update website messaging to set realistic response time expectations.
Practical applications of real time analytics
Managing unexpected volume fluctuations
Improving first contact resolution
Supporting agents in the moment
Making data-driven staffing adjustments
Real-time analytics help managers respond quickly by indicating the following:
Identifying volume surges as they begin
Showing which channels and contact types are most affected
Predicting the impact on service levels if no action is taken
Tracking the effectiveness of mitigation efforts
Practical example: When a retail company launches a new product, the contact center sees a 35 percent increase in call volume within the first hour. The manager takes the following immediate action:
Checks agent adherence to identify any scheduled agents who have not logged in.
Reassigns cross-trained agents from email to phone support.
Updates the IVR message to acknowledge higher-than-normal call volume.
Adds a banner to the website with answers to common questions about the new product.
Best practices for implementing real time analytics
Now you will review best practices for real-time analytics with Amazon Connect.
Start with clear objectives
Focus on actionable metrics
Design intuitive visual displays
Create tiered thresholds
Develop clear response protocols
Integrate with workforce management
Continuously refine your approach
Tactics that undermine success
Use the following strategies to overcome various factors that can undermine success in your contact center:
Information overload: Limit main dashboards to 5–7 key metrics with drill-down capabilities for details.
Alert fatigue: Carefully balance early warning with the amount of notifications and regularly adjust thresholds.
Focusing on technology instead of people: Allocate at least as much time to training and change management as to technical implementation.
Neglecting feedback loops: Create regular processes to review real-time data patterns and use them to drive systemic changes.
The Queue and agent performance dashboard display the real-time state for each agent. This is shown as Activity on the dashboard. Timers track duration in each state against performance thresholds. The dashboard has the ability to set color-coded indicators to show green, yellow or red custom threshold.
Figure 52 Queue and agent performance dashboardSelect image to enlarge
Service level report
Service level appears as a percentage of contacts answered within target time frames. For example, answering 80 percent of calls within 30 seconds would show as 80 percent for SL30. Time frames, queues and the performance indicator colors for Service Level (SL) metrics can all be configured on the report. In the report below you can see that only 1 percent of contacts have been answered with 15 seconds (SL15) for the AnyCompany Finance queue.
Figure 53 Queue and agent performance dashboardSelect image to enlarge
Real time alerts
Although dashboards provide visibility, alert systems help ensure that you do not miss critical changes even when you are not actively monitoring the dashboard.
Threshold based alerts
These alerts trigger when a metric crosses a predefined threshold, such as oldest contact age exceeding 10 minutes.
Trend based alerts
These notify you when metrics are moving in a concerning direction, like call volume increasing by 30 percent compared to the previous hour.
Anomaly detection
You could use anomaly detection in Amazon CloudWatch logs to detect unusual patterns from your contact center metrics like queue size and waiting times that might not trigger standard threshold alerts. To learn more about CloudWatch, see Amazon Monitoring and Observability.
Alert delivery methods
Alerts can be delivered through visual indicators on dashboards, tasks, emails, or events to Amazon EventBridge. To learn more about EventBridge, see Amazon EventBridge.
Real time monitoring and intervention
Real-time monitoring and intervention is the ability to monitor individual interactions as they happen and intervene when necessary.
Live contact monitoring
Monitoring options
Speech and sentiment analysis
Supervisors can listen to live calls or read live chats to assess quality, identify training needs, or step in to assist with difficult situations.
Integrating Real Time Analytics with Other Systems
The effectiveness of real-time analytics multiplies when integrated with the following systems:
Workforce management systems
When real-time analytics shows high call volumes, workforce management integration can automatically identify available staff to call in or agents who could work longer shifts.
CRM integration
Real-time analytics can pull customer information from CRM systems to provide context for current interactions. For example, a simple dashboard might only show that call volumes are high. With a CRM integration, you might find that 80 percent of those calls come from high-value customers.
Quality management integration
When real-time analytics identifies potential quality issues in customer interactions, it can automatically flag these interactions for later review by the quality team.
Automated response systems
Real-time analytics can trigger automated system responses without requiring human intervention. For example, when wait times exceed thresholds, systems can automatically offer callbacks and update website messaging to set realistic response time expectations.
Practical applications of real time analytics
Managing unexpected volume fluctuations
Improving first contact resolution
Supporting agents in the moment
Making data-driven staffing adjustments
Real-time analytics help managers respond quickly by indicating the following:
Identifying volume surges as they begin
Showing which channels and contact types are most affected
Predicting the impact on service levels if no action is taken
Tracking the effectiveness of mitigation efforts
Practical example: When a retail company launches a new product, the contact center sees a 35 percent increase in call volume within the first hour. The manager takes the following immediate action:
Checks agent adherence to identify any scheduled agents who have not logged in.
Reassigns cross-trained agents from email to phone support.
Updates the IVR message to acknowledge higher-than-normal call volume.
Adds a banner to the website with answers to common questions about the new product.
Best practices for implementing real time analytics
Now you will review best practices for real-time analytics with Amazon Connect.
Start with clear objectives
Focus on actionable metrics
Design intuitive visual displays
Create tiered thresholds
Develop clear response protocols
Integrate with workforce management
Continuously refine your approach
Tactics that undermine success
Use the following strategies to overcome various factors that can undermine success in your contact center:
Information overload: Limit main dashboards to 5–7 key metrics with drill-down capabilities for details.
Alert fatigue: Carefully balance early warning with the amount of notifications and regularly adjust thresholds.
Focusing on technology instead of people: Allocate at least as much time to training and change management as to technical implementation.
Neglecting feedback loops: Create regular processes to review real-time data patterns and use them to drive systemic changes.
Real-time monitoring and intervention is the ability to monitor individual interactions as they happen and intervene when necessary.
Live contact monitoring
Supervisors can listen to live calls or read live chats to assess quality, identify training needs, or step in to assist with difficult situations.
Monitoring options
Amazon Connect Customer offers the following monitoring options:
Silent monitoring: Supervisors can listen to live calls or read live chats without the agent or customer knowing.
Barging: Supervisors can join a live voice or chat conversation. For voice, the supervisor speaks to both agent and customer. For chat, the supervisor sends messages visible to both parties.
Speech and sentiment analysis
Amazon Connect Customer can analyze conversations in real-time, detecting customer sentiment, identifying key phrases, or flagging compliance issues.
Integrating Real Time Analytics with Other Systems
The effectiveness of real-time analytics multiplies when integrated with the following systems:
Workforce management systems
When real-time analytics shows high call volumes, workforce management integration can automatically identify available staff to call in or agents who could work longer shifts.
CRM integration
Real-time analytics can pull customer information from CRM systems to provide context for current interactions. For example, a simple dashboard might only show that call volumes are high. With a CRM integration, you might find that 80 percent of those calls come from high-value customers.
Quality management integration
When real-time analytics identifies potential quality issues in customer interactions, it can automatically flag these interactions for later review by the quality team.
Automated response systems
Real-time analytics can trigger automated system responses without requiring human intervention. For example, when wait times exceed thresholds, systems can automatically offer callbacks and update website messaging to set realistic response time expectations.
Practical applications of real time analytics
Managing unexpected volume fluctuations
Real-time analytics help managers respond quickly by indicating the following:
Identifying volume surges as they begin
Showing which channels and contact types are most affected
Predicting the impact on service levels if no action is taken
Tracking the effectiveness of mitigation efforts
Practical example: When a retail company launches a new product, the contact center sees a 35 percent increase in call volume within the first hour. The manager takes the following immediate action:
Checks agent adherence to identify any scheduled agents who have not logged in.
Reassigns cross-trained agents from email to phone support.
Updates the IVR message to acknowledge higher-than-normal call volume.
Adds a banner to the website with answers to common questions about the new product.
Improving first contact resolution
Real-time analytics helps improve first contact resolution (FCR) in the following ways:
Identifying when customers are repeatedly calling about the same issue
Highlighting knowledge gaps that prevent agents from resolving issues
Detecting when transfers between departments are increasing
Supporting agents in the moment
One of the most powerful applications is providing immediate support to agents when they need it most.
Practical example: A new agent has been on a call for 12 minutes, significantly longer than the average 5 minutes. The supervisor acts by doing the following:
Using silent monitoring to listen to the call
Recognizing the customer is asking about a complex product feature the agent has not been trained on
Using barge-in mode to guide the customer and agent through the process
After the call, briefly meets with the agent to reinforce the learning
Making data-driven staffing adjustments
Data can be useful for staffing adjustments in the following ways:
Allowing immediate response to unexpected staffing issues, such as call-outs or traffic surges
Allowing intra-day adjustments instead of waiting for the next scheduling cycle
Providing visibility into emerging patterns before they become established trends
Supporting dynamic reallocation of staff across departments as needs shift
Historical data analysis remains valuable for long-term planning and identifying seasonal patterns. The responsiveness that real-time solutions offer for day-to-day workforce optimization cannot be matched by historical analysis.
Best practices for implementing real time analytics
Now you will review best practices for real-time analytics with Amazon Connect.
Start with clear objectives
Define exactly what you want to achieve before implementing real-time analytics.
Focus on actionable metrics
Prioritize metrics that drive immediate action when they change, connect to customer experience, and can be influenced by real-time decisions.
Design intuitive visual displays
Use color coding consistently, group related metrics, use appropriate visualization types, and limit information on each screen.
Create tiered thresholds
Not all issues require the same response or urgency. Create a tiered system, such as the following:
Green: Metrics are not exceeding thresholds.
Yellow: Metrics crossed thresholds but are not critical.
Red: A significant service impact requires immediate action.
Develop clear response protocols
For each type of threshold, develop standard response protocols so managers know exactly what actions to take.
Integrate with workforce management
Make real-time analytics even more powerful by integrating with workforce management systems like Amazon Connect Customer forecasting, capacity planning, and scheduling.
Continuously refine your approach
Regularly review and adjust dashboard layouts, alert thresholds, and response protocols.
Tactics that undermine success
Use the following strategies to overcome various factors that can undermine success in your contact center:
Information overload: Limit main dashboards to 5–7 key metrics with drill-down capabilities for details.
Alert fatigue: Carefully balance early warning with the amount of notifications and regularly adjust thresholds.
Focusing on technology instead of people: Allocate at least as much time to training and change management as to technical implementation.
Neglecting feedback loops: Create regular processes to review real-time data patterns and use them to drive systemic changes.
Note - AI-powered manager assistance is currently available as a preview feature. To request access, contact your AWS account team or an AWS representative.
The real-time analytics tools include AI-powered manager assistance (preview), which lets supervisors ask natural language questions about their metrics instead of navigating dashboards manually.
This feature allows supervisors to ask questions like, "Why is average handle time increasing?" and receive instant diagnostic answers drawn from over 150 metrics.
How it works
Instead of clicking through dashboard filters and visually scanning for anomalies, supervisors type or speak questions in natural language. The AI assistant:
Interprets the question
Queries the relevant metrics across queues, agents, and time periods
Identifies the likely cause
Presents a concise answer with supporting data
For example: "Which queue has the longest wait time right now?" or "Are there any agents who haven't taken a call in the last 30 minutes?"
What kinds of questions it answers
The assistant can address:
Diagnostic — "Why is service level dropping in the billing queue?"
Status — "How many agents are currently available?"
Comparison — "How does today's call volume compare to last Tuesday?"
Alerting — "Is anything unusual happening right now?"
It draws from 150+ metrics to form answers, so supervisors do not need to know which specific metric to look at; they describe what they want to understand.
Impact on supervisor workflow
This shifts supervisors from reactive dashboard monitoring to proactive, question-driven management. Rather than waiting for an alert or noticing a number change, they can investigate hunches, confirm suspicions, and diagnose issues conversationally.
Additional enhancements
Several additional enhancements make real-time analytics more powerful and customizable.
Custom metrics no code
Supervisors can now create custom metrics using mathematical operations without technical skills. For example, combining existing metrics to calculate a "first-contact resolution rate" specific to your organization's definition. No developer involvement needed.
Custom business dimensions
Metrics can now be filtered by business divisions, product lines, or customer segments. Instead of seeing aggregate data across the entire contact center, a supervisor can focus on "enterprise customers in the returns queue" or "small business accounts handled by Team A."
Enhanced real time alerts
Alerts now include the specific context that triggered them, such as which agents, queues, flows, or routing profiles are involved. Previously, an alert might say "service level dropped below 80%." Now it says "service level dropped below 80% in the Billing queue, affecting agents Smith, Park, and Chen."
Flow designer analytics
Supervisors can view aggregate traffic through each step in contact flows, identifying where customers drop off, which branches handle the most volume, and where errors occur. This helps optimize flow design based on real usage patterns rather than assumptions.
Introduction AI Agent Monitoring and Enhanced Evaluations
AI Agent Monitoring and Analytics
AI agent dashboards
AI agent monitoring dashboards provide supervisors with key performance indicators specific to autonomous AI interactions.
Key dashboard metrics
The dashboards surface the following metrics for AI agent interactions:
Hand-off rate — percentage of interactions transferred to a human agent
Conversation turns — average number of exchanges per interaction
Goal success rate — percentage of interactions resolved successfully
Faithfulness score — whether responses are grounded in source material
These metrics update in real time, giving supervisors an always-current view of AI agent performance.
Version comparison
When you update an AI agent's configuration, such as changing its knowledge bases, adjusting guardrails, or modifying its tools, the dashboard lets you compare performance between versions side by side.
This answers the critical question: "Did our change make things better or worse?" You can see whether the new version improved goal success rate, reduced hand-offs, or inadvertently increased conversation turns.
Difference from human agent monitoring
Human agent dashboards focus on availability, adherence, and individual performance coaching. AI agent dashboards focus on:
System-level quality (is the agent working correctly?)
Configuration effectiveness (are the knowledge bases and tools sufficient?)
Boundary detection (where does the agent fail and need human takeover?)
The supervisory action differs too. For human agents, you coach. For AI agents, you tune
When to intervene
Dashboards show you what is happening. The next step is knowing what to do about it. The following patterns indicate an AI agent needs attention.
Signal
What it means
Action to take
Rising hand-off rate
Agent cannot resolve a growing number of requests
Check for new customer query types not covered by knowledge bases
Dropping faithfulness score
Agent is generating unsupported responses
Review and expand knowledge base content; tighten guardrails
Increasing conversation turns
Agent is taking longer to resolve issues
Analyze transcripts for confusion loops; improve tool selection
Low goal success rate
Agent is failing to resolve requests
Investigate specific failure categories; consider adding tools
Negative sentiment in bot analytics
Customers are frustrated with the AI
Review interaction patterns; adjust tone and escalation thresholds
Understanding the faithfulness score
The faithfulness score measures how closely an AI agent's responses align with its configured knowledge sources. Scores range from 0 to 1, where higher values indicate stronger grounding in approved content. When faithfulness scores drop below acceptable thresholds, supervisors should investigate the agent's traces to identify whether the agent is generating unsourced responses or hallucinating information.
AI agent traces
For deeper investigation beyond what dashboards provide, supervisors and developers can access AI agent traces through APIs.
Traces provide a step-by-step record of an AI agent's reasoning and actions during an interaction. They show:
What the agent "thought" at each step (its reasoning)
Which tools it considered and selected
What data it retrieved from knowledge bases
Why it made the decisions it made
This level of detail is essential for diagnosing specific failures. When the dashboard shows a problem, traces help you understand the root cause so you can fix it precisely rather than guessing.
Enhanced Evaluations
Multilingual evaluation support
Automated evaluations previously supported English only. They now operate in five additional languages, with cross-language evaluation support.
Portuguese
French
Italian
German
Spanish
Cross-language evaluation means a supervisor working in English can evaluate interactions that occurred in any of the supported languages. The system handles the translation and analysis, so language barriers do not prevent quality oversight.
New evaluation question types
Evaluation forms now support additional question types beyond the original format, enabling more nuanced assessment.
Multiple choice questions
Evaluators can now include multiple choice questions in evaluation forms. This enables:
Categorizing the type of issue handled
Selecting from predefined quality levels
Classifying interaction outcomes into specific buckets
Multiple choice provides structured data that is easy to aggregate and trend over time.
Date questions
Date-type questions allow evaluators to record specific dates within evaluations. This is useful for:
Tracking when issues were first reported
Recording promised follow-up dates
Noting deadlines communicated to customers
Date fields enable time-based analysis of evaluation data that was previously captured inconsistently in free-text fields.
Automated follow up evaluations
A powerful new workflow allows evaluation results to trigger additional evaluations automatically.
When an initial evaluation reveals a concern — for example, a low score on compliance or accuracy — the system can automatically trigger a follow-up evaluation on subsequent interactions by the same agent. This creates a continuous improvement loop:
Initial evaluation identifies a problem
Follow-up evaluations are automatically scheduled
Subsequent interactions are assessed for improvement
The cycle continues until performance meets the threshold
This removes the manual step of remembering to re-check agents who received poor scores. The system handles the follow-through.
Evaluating AI agent interactions
Evaluations are no longer limited to human agent interactions. Automated evaluations can now assess AI agent self-service interactions, with aggregated insights across all bot conversations.
How it works
The system automatically evaluates AI agent interactions against quality criteria you define:
Did the agent resolve the customer's request?
Did it stay within compliance guidelines?
Was the tone appropriate?
Did it escalate appropriately when needed?
Results are aggregated into dashboards showing AI agent quality trends over time.
Difference from human agent evaluations
Human agent evaluations often focus on soft skills, adherence to scripts, and coaching opportunities. AI agent evaluations focus on:
Configuration effectiveness (is the agent set up correctly?)
Knowledge base adequacy (does it have the information it needs?)
Guardrail performance (is it staying within boundaries?)
Escalation accuracy (is it handing off at the right moments?)
The corrective actions differ: for humans, you coach. For AI agents, you adjust configuration, expand knowledge bases, or refine guardrails.
RECALL & APPLY
Practice what you learned
Try answering in your own words, then compare with the suggested answer. These activities are for self-study and are not automatically graded.
01What changes when review expands beyond a small contact sample?
Topic: AI support for supervisors
Reveal suggested answer
Supervisors can examine wider patterns, but still need validated criteria, appropriate interpretation, and targeted human review.
02How should a supervisor respond to a recurring complaint pattern?
Topic: Proactive quality monitoring
Reveal suggested answer
Review representative contacts, verify the pattern, and determine whether the cause is a product issue, policy problem, or coaching need.
03Which view would help investigate why a customer became frustrated?
Topic: Conversational analytics views
Reveal suggested answer
Use the sentiment timeline to locate the change, then inspect the corresponding transcript and recording where available.
04Is enabling the instance capability the only configuration step?
Topic: Analytics modes and enablement
Reveal suggested answer
No. Review the applicable flow and analytics settings and the permissions needed for users to access the results.
05How does an issue category differ from a sentiment value?
Topic: Sentiment, issue detection, and summaries
Reveal suggested answer
A category describes what the contact concerns. Sentiment estimates emotional tone. They answer different questions and can be used together.
06Does automated redaction prove that every sensitive value was removed?
Topic: Compliance monitoring and redaction
Reveal suggested answer
No. The source explicitly notes that predictive systems can miss information. Review output against the organization’s requirements.
07Why is a case summary useful for a repeated billing complaint?
Topic: Analytics across channels and AI interactions
Reveal suggested answer
It combines the important facts and actions from related contacts so the reviewer can see what remains unresolved.
08What makes an operational metric actionable?
Topic: Real time operational analytics
Reveal suggested answer
A change in the metric has an understood meaning, an accountable owner, and a practical response the team can take.
09What must be known before interpreting a service level percentage?
Topic: Dashboard views and service levels
Reveal suggested answer
The target time, included contacts, time period, and queue or channel scope. Percentages with different definitions are not directly comparable.
10How can an alert become a useful response instead of another notification?
Topic: Alerts and live intervention
Reveal suggested answer
Include the relevant context, route it to the right owner, and define what action is expected at that level of urgency.
11A product launch produces a sudden call surge. What should the supervisor inspect first?
Topic: Responding to contact volume changes
Reveal suggested answer
Confirm the affected queue, current staffing and adherence, and the reasons customers are contacting support before selecting a response.
12Why might a long agent activity duration need context?
Topic: Queue and agent performance examples
Reveal suggested answer
A complex contact may legitimately take longer. Review the interaction and workload before treating duration alone as a performance problem.
13An agent is handling an unusually long call. What is an appropriate first step?
Topic: Monitoring and operational integrations
Reveal suggested answer
Review the contact context through the permitted monitoring workflow. Determine whether the agent needs assistance before intervening.
14How would you reduce alert fatigue?
Topic: Effective real time analytics practices
Reveal suggested answer
Review false or low value alerts, adjust thresholds, clarify priorities, and keep notifications tied to actions that teams can take.
15What should an instructor verify before demonstrating a preview capability?
Topic: AI assisted management and custom metrics
Reveal suggested answer
Confirm that the training environment has access and that its behavior and interface match the planned demonstration. Preserve the source’s preview qualification.
16Why can a lower hand off rate be a poor result?
Topic: AI agent monitoring
Reveal suggested answer
The agent may be retaining contacts it cannot safely or correctly resolve. Review success, quality, and appropriate escalation together.
17Faithfulness declines after a knowledge update. Where should the investigation begin?
Topic: AI agent investigation and traces
Reveal suggested answer
Compare affected interactions, retrieved content, and configuration versions. Look for missing, outdated, or irrelevant knowledge and unsupported answers.
18Why should an evaluation form use observable criteria?
Topic: Evaluation capabilities
Reveal suggested answer
Observable criteria support consistent scoring and make disagreements easier to investigate using contact evidence.
19How does corrective action differ for a human agent and an AI agent?
Topic: Evaluating AI agent interactions
Reveal suggested answer
Human follow up often involves coaching. AI follow up involves knowledge, instructions, tools, guardrails, and configuration testing.
Capture a key idea, an example, or a question for your instructor.
Imagine running a contact center where you have the right number of agents available at the right times. In this contact center, agent performance is evaluated automatically, and both customers and agents are happier. Welcome to the world of AI-powered workforce optimization in Amazon Connect Customer. Contact center workforce optimization involves strategically scheduling the right number of skilled agents to match staffing needs. It is about balancing excellent customer service with operational efficiency. In a typical contact center, workforce optimization involves several key components.
Forecasting
A forecast attempts to predict future contact volume and average handle time by predicting how many customer contacts you will receive and when they will come in.
Amazon Connect Customer uses historical metrics to generate the forecast.
Capacity planning
Capacity planning determines how many agents you will need to handle the forecasted contact volume while meeting your service level goals.
A capacity plan in Amazon Connect Customer helps you estimate the long-term full-time equivalent (FTE) requirements for your contact center up to 64 weeks in the future. It specifies how many FTE agents are required to meet your service level target for a certain period of time.
Scheduling
Contact center schedulers or managers need to create agent schedules for day-to-day workloads that are flexible and meet business and compliance requirements.
Amazon Connect Customer helps you create efficient schedules that are optimized for per-channel service level or average speed of answer (ASA) targets.
You can generate and manage agent schedules based on the following:
A published forecast
Shift profiles (templates for weekly shifts)
Staffing groups (agents that can handle specific types of contacts from a specific forecast group)
Human resources and business rules
Time off management
Contact center leaders coordinate team members' time off requests and manage schedule activities due to operational changes.
With Amazon Connect Customer time off management capabilities, you can manage agent time off requests that comply with preconfigured regional labor and business rules.
Administrators or managers with the appropriate security profile permissions can configure the time off settings. Amazon Connect Customer automatically approves or rejects requests depending on how you've configured both the time off rules and the daily maximum allowed time off hours.
Supervisors or managers with permissions can view agent time off requests and override automatically approved or rejected time off.
Overtime and voluntary time off management
Effectively balancing agent supply with customer demand is essential for meeting service goals while controlling costs. When you properly match the number of agents with incoming contacts, you create an environment where service levels and response times meet targets without unnecessary expenses. This forms the foundation of efficient workforce management.
Overtime (OT) and voluntary time off (VTO) provide the flexibility needed to adapt to changing conditions. When contact volumes spike or agent availability drops, you can use overtime to extend coverage without hiring new staff. In contrast, during slower periods, willing agents can use VTO to take unpaid time off, which reduces costs while maintaining appropriate staffing levels. Together, these capabilities create a responsive workforce model that can quickly adjust to business needs while keeping both customers and financial objectives in focus.
Schedule adherence
Contact center supervisors or managers track schedule adherence to understand when agents are following the schedule that you have created.
This helps ensure you achieve your service level targets, while improving agent productivity and customer satisfaction.
Intraday performance monitoring
Managers and team leaders can observe daily patterns and workload levels to adjust staffing schedules as needed. They can make predictions for rest-of-day contact volumes, average queue answer time, average handle time, and effective staffing.
The Amazon Connect Customer Intraday Forecast Performance Dashboard provides forecasts for the following:
Contact volume and average handle time for queues that have a minimum of 5,000 unique contacts per week per queue-channel for the last 4 weeks. This threshold is evaluated on a rolling basis, with intraday forecasts refreshed every 15 minutes throughout the day covering future intervals of the current day.
Average queue answer time follows the same threshold of 5,000 unique contacts per week per queue-channel over the trailing 4-week period.
Quality management
AI-powered quality management ensures consistent service delivery through intelligent agent evaluation and automated insights. Amazon Connect Customer agent evaluations use machine learning (ML) and generative AI to enhance quality assessments and provide data-driven coaching recommendations.
Key Amazon Connect Customer agent evaluations features include the following:
Generative AI-powered evaluations that automatically answer evaluation questions with rich context and detailed reasoning
Flexible automation options including manual assessment, rule-based automation, and AI-powered automation based on business needs
Multi-evaluator calibration capabilities that enable simultaneous evaluation of contacts to reduce bias and improve accuracy
Custom evaluation forms with ML-powered Conversational Analytics and intelligent contact selection
Coaching insights that provide targeted recommendations for agent development
Native integration built directly into Amazon Connect Customer without requiring external systems
AI-powered quality evaluations help maintain service consistency, ensure compliance through automated monitoring, and support agent development through intelligent feedback and generative AI-driven coaching recommendations.
Workforce optimization cycle
Amazon Connect workforce management components work together as services that consume shared forecast data, with forecasting as the required foundation. These six core components help optimize your contact center experience. The components include forecasting, capacity planning, scheduling, schedule adherence, agent flexibility, and intraday operations. Forecasting publishes predictions that the other five components consume independently—capacity planning and scheduling, for example, each use the published forecast without depending on each other, creating a comprehensive optimization system.
Forecasting
Analyze and predict contact volume and handle time based on historical data. Forecasts provide the foundational data that other components consume. Each downstream component uses published forecasts without depending on the others.
Capacity planning
Determine how many agents you need to handle forecasted contact volumes while meeting service level goals, up to 64 weeks in the future.
Scheduling
Generate agent schedules for day-to-day workloads that are flexible and meet business and compliance requirements.
Schedule adherence
Track whether agents are following published schedules. This helps ensure you achieve your service level targets while improving agent productivity
Agent Flexibility
Manage overtime, voluntary time off, shift exchanges, and time off requests to balance agent supply with customer demand.
Intraday operations
Monitor daily patterns and workload levels to adjust staffing schedules as needed, with forecasts refreshed every 15 minutes throughout the day.
Traditional Workforce Management Challenges
Common challenges
With traditional workforce management, you are relying on various manual processes, communication from various teams, and inaccurate or out of date forecasting and data.
Common challenges include the following:
Inaccurate forecasts leading to overstaffing or understaffing.
Time-consuming manual scheduling processes that can't quickly adapt.
Inconsistent quality evaluation dependent on individual reviewer preferences.
Limited visibility into true agent performance and productivity.
Difficulty balancing business needs with agent preferences.
How AI Enhances Workforce Management
ML powered forecasting
ML transforms workforce forecasting with powerful capabilities and insights that were previously difficult and time consuming. ML-powered forecasting in Amazon Connect Customer uses advanced machine learning to identify complex patterns in your contact data to predict contact volumes and their arrival rates.
By predicting contact volumes and arrival rates, Amazon Connect Customer improves the accuracy and efficiency of forecasts and schedules. Amazon Connect Customer generates these forecasts using an ML model specifically tailored for contact center operations.
It accounts for the following two key factors:
Seasonal trends across multiple years
Recent arrival patterns
Figure 54 ML powered forecastingSelect image to enlarge
The result? Short-term forecasts are computed daily and long-term forecasts weekly. Once you publish a forecast, it drives your scheduling and capacity planning with greater accuracy than traditional methods.
Modern workforce optimization is not about separate tools and processes. Instead, it focuses on creating a seamless system where each component supports the others.
Integrated approach to workforce optimization
Imagine you are the manager at a banking contact center. Your forecast shows an expected increase in calls about a new mortgage product. That information automatically flows into your capacity planning tool, which calculates the needed staffing. The scheduling system then creates efficient schedules, and the quality management system knows to focus evaluations on mortgage-related calls during that period.
Figure 55 Integrated approach to workforce optimizationSelect image to enlarge
Forecasts created
Historical contact data is used to create forecasts that flow into capacity planning.
Staffing requirements
The scheduling system then creates optimal schedules.
Agent schedules
This creates efficient agent schedules based on the configured staff rules and shift patterns.
Targeted evaluations
Contact evaluations can be targeted toward mortgage-related calls, ensuring agents are adhering to company guidelines and policies.
Introduction to ML Driven Forecasting and Capacity Planning
How ML Forecasting Works and Benefits
How ML forecasting works
ML-driven forecasting uses machine learning models tailored for contact center operations to analyze historical contact center data and predict future contact volumes and handle times. Unlike traditional forecasting methods, ML models can identify complex patterns, seasonal trends, and subtle relationships between variables that human forecasters might miss. For example, imagine you are running a retail customer service center. Your ML forecasting system analyzes two years of data and notices patterns of call volume increases outside of obvious holiday periods.
Figure 56 How ML forecasting worksSelect image to enlarge
Data collection
The system gathers historical information including:
Contact volumes
Handling times
Abandon rates
Special events
2: Data preprocessing
AI cleans the data, removing anomalies that could skew predictions.
Pattern recognition
ML algorithms identify recurring patterns and correlations.
Model training
The system creates and refines mathematical models based on the historical data.
Forecasting
The trained models generate predictions for future time periods.
Continuous learning
As new data comes in, the system updates its models to improve accuracy.
Amazon Connect automatically updates short-term forecasts daily and long-term forecasts weekly to provide fresh forecasts based on current information.
Forecasting Types
Contact centers typically use three types of forecasts.
Short term forecasts daily/weekly
Short-term forecasts (daily/weekly) Short-term forecasts predict contact volumes for the immediate future, typically up to 18 weeks ahead. These forecasts are crucial for day-to-day staffing decisions and scheduling. Think of this like checking the weather app before getting dressed in the morning. You need accurate information to make immediate decisions.
Figure 57 Forecasting TypesSelect image to enlarge
Long term forecasts monthly/quarterly/yearly
Long-term forecasts (monthly/quarterly/yearly) Long-term forecasts look further ahead (up to 64 weeks) to help with strategic decisions around hiring, training, and infrastructure planning. This is more like studying climate patterns before deciding where to build a house. You are making bigger decisions that cannot be easily changed.
Figure 58 Forecasting TypesSelect image to enlarge
Intraday forecasts real time
Intraday forecasts (real time) Intraday forecasts are ultra-short-term forecasts that help managers make immediate staffing adjustments throughout the day and update every 15 minutes. Imagine having a weather app that updates every 15 minutes during a storm. That is how intraday forecasting helps contact center supervisors navigate unexpected surges in call volume.
Figure 59 Forecasting TypesSelect image to enlarge
Benefits of ML forecasting
Using ML for contact center forecasting delivers several key advantages as follows:
Improved accuracy – ML models typically reduce forecast errors compared to traditional methods. This translates directly to better staffing decisions.
Automated updates – The system automatically refreshes forecasts (daily for short-term, weekly for long-term), which eliminates manual work and ensures forecasts stay current.
Pattern detection – ML identifies complex relationships that humans might miss, such as how weather conditions in specific regions affect certain types of support calls.
Multi-channel integration – Modern ML forecasting handles forecasts across all customer contact channels (phone, chat, email, social media) allowing for unified planning.
Common forecasting mistakes to avoid
Even with ML assistance, contact centers still make the following common forecasting errors:
Ignoring outliers
While days with unusual spikes or drops in activity should be cleaned from regular training data, they should still be analyzed for what caused the spike or drop.
Over reliance on technology
ML provides recommendations, but human judgment remains valuable for interpreting results and applying business context.
Using insufficient historical data
For best results, provide at least 12 months of historical data to capture seasonal patterns. Amazon Connect Customer models can use a maximum of 156 weeks (approximately 3 years) of data. The minimum requirement is 1,000 contacts per month within the last 6 months.
Not adjusting forknown future events
Always supplement ML forecasts with information about upcoming promotions, product launches, or other events that may affect contact volumes.
Forecasting is where all workforce planning begins. Without accurate forecasts, your scheduling and capacity planning will always be flawed, no matter how sophisticated your other processes are.
Capacity Planning Components and Process
Capacity planning overview
Capacity planning is the process of determining how many agents your contact center needs to handle forecasted contact volumes while achieving your service level targets. It bridges the gap between your forecasts (what you expect to happen) and your scheduling (who works when). Think of capacity planning like determining how many checkout lanes to open at a grocery store. Open too few, and customers wait in long lines. Open too many, and you are paying cashiers to stand around. The goal is to find the right balance between customer wait times and resource utilization.
Successful capacity planning requires the following three key components:
Accurate forecasts
Before you can create a capacity plan, you need reliable forecasts of contact volume and handle time. You must publish long-term forecasts before generating capacity plans.
Planning scenarios
Scenarios let you model different business conditions to see how they affect staffing needs. A scenario typically includes the following:
Maximum occupancy rate – Percentage of time agents spend actively handling contacts
Shrinkage factors – Breaks, training, absences, and after-contact work
Daily attrition rate – Model staff turnover rates
FTE hours per week – Working hours per week for full-time employees
Outsourced contacts – Model outsourcing percentages to third parties
Business operation days – Days per week or month that the contact center operates (for example, 5 days per week, 7 days per week, or specific calendar days)
Maximum OT – Amount of OT allowed
Maximum VTO – Amount of VTO allowed
Service level goals
You must define what success looks like using one of the following two metrics:
Service level – Percentage of contacts answered within a target time (for example, 80 percent in 20 seconds)
Average speed of answer (ASA) – The average time customers wait before speaking with an agent (for example, 30 seconds)
Figure 60 Service level goalsSelect image to enlarge
Capacity planning process
The capacity planning process includes the following steps.
Prepare your inputs
Before starting, you need:
Published long-term forecasts for the relevant time period
At least one planning scenario defining your operational parameters
Clear service level or ASA targets
Generate the capacity plan
Using the forecasts and scenarios, the capacity planning system calculates:
Required FTE employees with and without shrinkage
Forecasted occupancy rates
Gaps between available and required FTEs
Maximum OT and VTO allowances
Analyze the results
Review your capacity plan to identify:
Periods of understaffing where you will need to hire or use OT
Periods of overstaffing where you might offer VTO
Patterns that suggest you need to adjust your resourcing strategy
Create action plans
Based on your analysis, develop plans to address any gaps:
Recruiting and hiring timelines
Training schedules for new hires
OT policies for covering short-term gaps
Cross-training opportunities to increase flexibility
Historical Data Utilization
Using Historical Data and ML Forecasts
Did you know that your contact center is a repository of valuable data? Every call, chat, and email contains clues about future customer behavior. You can use the power of your historical contact data combined with ML forecasts to make smarter business decisions. This is about extracting practical insights that help you deliver better customer experiences while controlling costs.
Historical data is the foundation of all contact center planning. ML forecasts of historical data provide the following:
Baseline understanding of typical patterns
Trend identification for growing or declining contact types
Seasonal pattern recognition
Think of historical data as your contact center's memory. Just as you remember that your coffee shop is busiest on Saturday mornings, your historical data remembers when your contact center is busy or slow.
The following image shows an example of an ML forecast based on historical data in Amazon Connect Customer.
Figure 61 Using Historical Data and ML ForecastsSelect image to enlarge
Key historical metrics for forecasting
When using historical data for ML-powered forecasting, focus on the following essential metrics:
Volume metrics
Contact counts by channel
Interval distribution (daily, 15-minute and 30-minute breakdowns)
Day-of-week patterns
Month-of-year patterns
Handling metrics
Average handle time
Amazon Connect Customer automatically generates forecasts, with the ability for manual adjustments when needed through override capabilities.
While ML excels at finding patterns, human judgment remains crucial for contextual understanding, interpretation of unusual results, strategic decision making, and stakeholder communication.
Introduction to Intelligent Scheduling
AI Optimized Scheduling Process
AI scheduling overview
To run a contact center, you need the right number of agents working at the right times to achieve your operational goals.
Scheduling in Amazon Connect Customer helps with the following:
Generate agent schedules for day-to-day workloads that are flexible and meet business and compliance requirements.
Ensure you have the right number of agents to avoid both overstaffing (which leads to overspending) and understaffing (which impacts service levels).
AI scheduling starts with forecasting to predict how many contacts will arrive at different times. Contact centers used to rely on educated guesses to create schedules. Today, AI-optimized scheduling transforms this process. You can use Amazon Connect Customer to create AI optimized schedules that are per-channel service level or average speed of answer (ASA) targets.
The following image shows an example of an Amazon Connect Customer forecast graph with predicted call volumes by hour and day with peak periods highlighted.
Figure 62 AI scheduling overviewSelect image to enlarge
AI trend detection
Unlike humans who might miss subtle patterns, AI can detect trends across factors like day-of-week patterns, time-of-day variations, seasonal trends, and historical anomalies.
Calculating staffing needs
After the AI system has predicted contact volume, it calculates how many agents are needed for each 15-minute or 30-minute period considering the following:
How long it typically takes to handle each contact
The target service level
Required break times
Staff rules
Figure 63 AI scheduling overviewSelect image to enlarge
Scheduling optimization
The system then creates shift activities that put the right number of agents in the right places, trying several combinations to find the optimal solution.
Figure 64 AI scheduling overviewSelect image to enlarge
Improved accuracy
AI scheduling addresses the following common problems:
Overstaffing – AI optimizes staffing levels to predicted demand, ensuring you're not paying for idle time.
Understaffing – By accurately predicting busy periods, AI ensures adequate coverage during peak times.
Inefficient shift patterns – AI creates shifts that precisely match your contact patterns to maximize efficiency.
Example of the AI optimized schedule
Real-world example: For an online retailer, the AI system might notice that Monday mornings have high call volumes about weekend orders. Additionally, Thursday evenings see increased chat requests about weekend deliveries. The system would schedule more phone agents on Monday mornings and more chat agents on Thursday evenings.
Balancing Agent Preferences and Contact Center Needs
Business and agent needs
Contact center scheduling has traditionally been employer-centric and often leads to high turnover, lower productivity, and increased absenteeism.
Contact center agents and businesses face common needs.
Business needs
The following core operational metrics drive sustainable contact center performance:
Sufficient coverage for all expected contacts
Meeting service levels
Cost efficiency
Skill coverage
Agent needs
The following critical workplace flexibility factors support a stable and engaged workforce:
Consistent schedules
Work-life balance
Preferred shift times
Time off for personal events
Consecutive days off
Balancing solutions
Amazon Connect Customer improves your contact center's approach with scheduling optimization as follows:
Staff rules for agent-level customization – Before creating schedules, contact centers set agent preferences for shift times, preferred days off, maximum consecutive workdays, and specific time-off needs.
Shift exchange – Agents can trade shifts with each other, provided the exchange doesn't create coverage issues.
Flexible time off options – Progressive contact centers offer voluntary time off (VTO) during slow periods, advanced time-off requests, and optional overtime during busy periods.
Staff Rules Configuration
Staff contract rules
You can use Amazon Connect Customer to set customized rules at the agent level. Before creating schedules, contact centers set agent preferences for shift times, preferred days off, maximum consecutive workdays, and specific time-off needs.
Min working time
Minimum working hours/minutes per day/week
Max working time
Maximum working hours/minutes per day/week
Min consecutive working days
Minimum consecutive working days
Max consecutive working days
Maximum consecutive working days
Min time gap
Minimum time gap between shifts (hours)
Min consecutive days off
Minimum consecutive days off
Flexibility with scheduling adjustments
There is flexibility in scheduling adjustments so that contact centers can take a progressive approach when balancing business and agent needs.
Shift exchange options
Flexible time off options
Agents can trade shifts with each other, provided the exchange doesn't create coverage issues.
Real world examples of staff rules configuration
For the holiday season, a retail contact center might do the following:
Ask for volunteers to work key dates with overtime incentives.
Create shorter shifts on major holidays.
Plan far in advance so agents know their holiday schedule.
For agents who are parents, a contact center might do the following:
Create dedicated parent shifts during school hours.
Allow part-time options during school terms.
Create flexible scheduling during school breaks.
Flexibility with scheduling adjustments
Shift exchange options
Agents can trade shifts with each other, provided the exchange doesn't create coverage issues.
Flexible time off options
Progressive contact centers offer the following:
Voluntary time off (VTO) during slow periods
Advanced time-off requests
Optional overtime during busy periods
Real world examples of staff rules configuration
For the holiday season, a retail contact center might do the following:
Ask for volunteers to work key dates with overtime incentives.
Create shorter shifts on major holidays.
Plan far in advance so agents know their holiday schedule.
For agents who are parents, a contact center might do the following:
Create dedicated parent shifts during school hours.
Allow part-time options during school terms.
Create flexible scheduling during school breaks.
Intraday Forecast Performance Dashboard
Dashboard overview
The Intraday forecast performance dashboard is a real-time monitoring tool in Amazon Connect Customer that helps contact center managers optimize daily operations. It tracks and forecasts key metrics including contact volume, handle time, answer speed, and staffing levels over 24-hour periods in 15-minute intervals.
The dashboard combines current data with historical patterns and short-term forecasts so managers can anticipate operational needs. This helps them adjust staffing levels and respond to emerging trends. Its color-coded indicators and comparison features help you quickly assess performance against benchmarks, which leads to immediate, data-driven operational decisions.
The Amazon Connect Customer Intraday Forecast Performance Dashboard provides forecasts for the following:
Contact volume and average handle time for queues that have a minimum of 5,000 unique contacts per week per queue-channel for the last 4 weeks. This threshold is evaluated on a rolling basis, with intraday forecasts refreshed every 15 minutes throughout the day covering future intervals of the current day.
Average queue answer time follows the same threshold of 5,000 unique contacts per week per queue-channel over the trailing 4-week period.
The following image shows the Intraday dashboard in Amazon Connect Customer.
Figure 65 Intraday Forecast Performance DashboardSelect image to enlarge
Performance overview chart
The Intraday trailing performance overview chart provides aggregated metrics based on your filters. Each metric in the chart is compared to your compare to benchmark time range filter.
The following image shows an example Intraday trailing performance overview chart.
Figure 66 Intraday Forecast Performance DashboardSelect image to enlarge
Image features
This image shows the following:
Contact volume during your time range selection was 1,213.
This is down ~13 percent compared to your benchmark number of contacts handled.
Avg. handle time has increased by 3.92 percent.
Avg. speed of answer has increased by 41 seconds.
As average speed of answer increases, service quality decreases. This metric inversely relates to contact center performance, with longer wait times indicating less efficient customer experience.
Abandonment rate shows the current day performance and is not compared to the previous performance.
The colors that appear for the metrics indicate negative (red) compared to the benchmark.
Comparison trend graphs
The Intraday performance dashboard displays the following trend graphs, which cover different metrics:
Contact volume
Average handle time
Average speed of answer
Effective staffing
These graphs include the Intraday forecast that projects up to 24 hours on a 15-minute interval based on the following:
The value of the respective metric
The historical actuals from the current day
The historical actuals from the same time in the past week
These trend graphs provide data only for the next 24 hours and the past 24 hours. There is no option to change the time range.
The following image shows an example of a Contact volume trend graph from the Intraday dashboard.
Figure 67 Intraday Forecast Performance DashboardSelect image to enlarge
Comparison against short term forecasts
You can compare Average handle time and Contact volume against published short-term forecasts.
To select this option, choose the Compare to button and select Short-term published forecast. This automatically picks up the published short-term forecast for the time range selected. You can't select an unpublished forecast or a specific published forecast.
For historical widgets, it compares against the same time range as the widget. For the daily projection widget, it compares against the entire day.
Figure 68 Intraday Forecast Performance DashboardSelect image to enlarge
Daily projection chart
The daily projection chart provides a projection of how the day will end by combining historical metrics for the day so far with Intraday forecasts for the remainder of the day. This is available for the following metrics:
Average handle time
Average queue answer time
Contact volume
Effective staffing
This widget only supports comparing against short-term forecasts for contact volume and average handle time.
Figure 69 Intraday Forecast Performance DashboardSelect image to enlarge
Real Time Schedule Adjustments
Why Adjustments Are Needed
Even with the best AI and forecasting, schedules need adjustment because of the following:
Agents may be absent (illness, family emergencies, transportation problems)
Contact volumes might differ from forecasts (marketing campaigns perform differently than expected)
Handle times might change (new issues might take longer to resolve)
Business priorities might shift suddenly
Real-time schedule adjustment is the process of identifying these mismatches between planned and actual conditions, then making quick, intelligent changes to minimize their impact.
Identifying the mismatch
A supervisor can identify mismatches between planned and actual conditions in Amazon Connect Customer using real-time adherence monitoring and intraday forecasting tools. The real-time metrics dashboard refreshes every 15 seconds, giving supervisors a live view of agent states compared to their published schedules. When an agent deviates from their planned activity, Schedule Adherence Notification Rules automatically alert the supervisor via email, task, or third-party integrations. Configurable adherence thresholds (1–10 minutes per activity) help distinguish genuine non-adherence from minor operational variances, reducing alert fatigue. On the demand side, intraday forecasts operate at 15- or 30-minute intervals, allowing supervisors to compare forecasted contact volume against actual arrival patterns and spot emerging staffing gaps before they impact service levels.
Mitigating the mismatch
Once a mismatch is identified, supervisors have several tools to respond quickly. Schedules can be edited in real time with minute-level precision. Agents with overlapping shifts can swap schedules, and when agents need to be reassigned, supervisors can move them to the correct staff group and regenerate the schedule for that individual without disrupting the rest of the team's optimized schedule. For non-phone activities such as coaching, training, or meetings, the Optimize Activity Placement feature automatically places them at algorithmically optimal times within shifts to minimize service level impact. When forecasts change, supervisors must manually trigger a schedule refresh because published schedules do not auto-update.
Types of adjustments
Understanding what happened and why can help to refine staffing strategies for similar situations. Review the following types of adjustments.
Schedule extensions
When contact volumes exceed forecasts or agent availability falls below required levels, implementing schedule extensions provides essential coverage to maintain service levels.
Asking agents to stay later (overtime)
Bringing agents in earlier
Converting scheduled training to work time
Schedule reductions
During periods of unexpected low contact volume, strategically reducing scheduled agent hours helps align staffing with actual business needs while managing labor costs.
Offering voluntary time off (VTO) during slow periods
Sending agents to training during slow times
Allowing longer breaks during slow periods
Skill adjustments
Flexible modification of agent skill assignments allows for dynamic reallocation of workforce resources to match changing contact patterns and business priorities.
Temporarily reassigning agents to different channels
Enabling backup skills
Moving agents between departments
When deciding how to adjust schedules, consider the following:
Impact on service levels
Impact on agent well-being
Cost implications
Fairness and process
Duration of the need
Insider tip: Create an opportunity to improve future scheduling by documenting each change and the outcome.
Intelligent Scheduling Enhancements
Multi Skill Agent Scheduling
The AI-optimized scheduling process you learned about has received several enhancements that give supervisors more control and produce better results. Previously, scheduling optimization treated agents as interchangeable. Now it accounts for the fact that agents have different specialized skills, and customer needs vary by skill type.
Introduced in November 2025, multi-skill scheduling optimizes based on each agent's specific skill set. If your contact center handles billing, technical support, and sales, the scheduler now ensures the right mix of skills is available at every time slot, not just the right headcount.
How it works
The scheduling algorithm considers:
Each agent's certified skills (billing, tech support, sales, Spanish-speaking, etc.)
Forecasted demand per skill
Service level targets per queue or skill group
The result is a schedule where you have enough billing-skilled agents during billing peaks and enough tech support agents during the hours technical issues spike — even if total headcount remains the same.
Why it matters
Without multi-skill scheduling, a contact center might be fully staffed but still miss service levels because the wrong skills are available. You could have 20 agents on shift but only 2 who can handle Spanish-language billing inquiries during a Spanish-language peak.
Multi-skill scheduling eliminates this mismatch by treating skills as a dimension of the optimization, not an afterthought.
Schedule adherence thresholds and notifications
Supervisors can now define how early or late agents can start or end shifts and activities, with automated notifications when thresholds are exceeded.
Introduced in October 2025, this feature adds proactive compliance monitoring to scheduling:
Define thresholds — set acceptable ranges for how early or late agents begin and end shifts and scheduled activities (breaks, training, lunch)
Automated notifications — when agents exceed these thresholds, supervisors receive email or text alerts automatically
No manual checking required — the system monitors adherence continuously and flags exceptions
This shifts schedule adherence from a reactive review (checking yesterday's data) to a real-time compliance tool. Supervisors know immediately when patterns emerge, allowing coaching conversations while the behavior is fresh.
Individual scheduling and bulk operations
Two operational improvements simplify day-to-day scheduling management for larger teams.
Individual agent scheduling
Managers can now schedule individual agents and merge their schedule with the existing team schedule. This is useful for:
New hires who join mid-cycle and need schedules created outside the normal generation run
Agents returning from leave who need reintegration into the rotation
Special assignments or temporary schedule adjustments for specific individuals
Bulk operations
For larger changes, managers can now:
Copy existing schedule configurations as templates for new schedules
Bulk edit scheduling settings across multiple agents simultaneously
Apply rule changes to groups rather than updating agents one by one
These efficiency improvements reduce the administrative time supervisors spend on scheduling mechanics, freeing them to focus on optimization and coaching.
Introduction to AI Powered Agent Performance Evaluation
Automated Quality Management
AI powered quality management
Traditional quality management approaches limit contact centers to evaluating a small sample of interactions, typically 1–3 percent, leading to inconsistent evaluations and delayed coaching opportunities.
Amazon Connect Customer Conversational Analytics (formerly known as Contact Lens) transforms quality management through generative AI-powered automation as follows:
Use AI to automatically evaluate 100 percent of customer interactions.
Generate performance insights from conversations as they happen.
Consistently apply standardized and objective evaluation criteria.
Flag to supervisors for immediate attention.
Extract performance trends and insights across agent population.
Statistical analysis provides valid insights into trends.
AI-powered analytics in Amazon Connect Customer go deeper by analyzing what actually happens during interactions.
Amazon Connect Customer Conversational Analytics with generative AI can do the following:
Evaluate agent performance based on natural language criteria automatically.
Detect compliance issues or script adherence problems.
Identify successful techniques used by top performers.
Provide objective, consistent evaluations.
Deliver feedback much faster than manual processes.
Instead of a supervisor reviewing 3–5 calls per agent per month, the AI system can evaluate every single interaction. This ensures no problematic interactions slip through while identifying top performers consistently.
Effective Evaluation Forms and Best Practices
Designing effective evaluation forms
Evaluation forms are important for measuring the quality and effectiveness of your contact center. You can use AI to help with this in many ways, but it's important to implement best practices when creating evaluation forms.
To design evaluation forms that work effectively with AI automation, use the following best practices:
Be specific
Use specific, observable behaviors rather than vague criteria. For example, use: "Did the agent address the customer by name at least once during the call?", rather than: "Was the agent professional?"
Use binary criteria when possible
Questions with clear yes/no answers are easier for AI to evaluate accurately.
Include verification points
For compliance items, specify exactly what the agent should say.
Balance process and outcome
Include both process measures (did the agent follow steps?) and outcome measures (was the issue resolved?)
Test and refine
Start with a small set of criteria, validate AI evaluations against human evaluations, and refine as needed.
The following image shows the evaluation form builder in Amazon Connect Customer.
Figure 70 Automated Quality ManagementSelect image to enlarge
Evaluation Forms in Action
Implementing effective evaluation forms and processes creates a foundation for quality improvement across your contact center. The most effective forms combine AI automation with human insight to measure what matters most. Scores on evaluations should do more than grade agent performance. They should guide coaching conversations and highlight development opportunities.
Overview
Review all the key information about the evaluation form such as:
Evaluation score
Evaluation ID
Status
Timestamps
Timezone
History
Evaluation scores
See how this contact evaluation compared to the agent's trailing four-week average and the average of all contacts.
Weighting
See how each section and question is weighted within the evaluation from.
AI generated answer
Receive a generative AI-powered recommendation for the answer, along with context and justification (reference points from the transcript that were used to provide answers).
Did you notice how AI assists with objective scoring in the completed evaluation form example? This technology saves supervisors time while providing consistent feedback to agents. Through smart form design and regular evaluations, agents learn exactly what success looks like. Organizations benefit from better customer experiences and customers receive more consistent service. Your evaluation strategy becomes a powerful tool that benefits everyone involved in the customer service journey.
Performance Evaluation Enhancements
Generative AI email overviews
Beyond evaluating 100 percent of interactions with standardized criteria, automated quality management includes features that improve both the evaluation process and the actions taken based on results.
Agents handling email interactions now receive AI-generated overviews that help them respond faster and more accurately.
This feature provides agents with:
Email overview — a concise summary of the customer's email, highlighting the key issue and any relevant history
Suggested actions — recommended next steps based on the email content and organizational policies
Draft responses — AI-generated response suggestions the agent can review, edit, and send
This reduces the time agents spend reading lengthy email threads and deciding how to respond. The AI handles the analysis; the agent reviews and acts.
Introduction to Audio Enhancement
Two modes of audio enhancement
Audio Enhancement offers two specialized modes. Each targets a different type of audio interference on the agent's side of the call.
Voice isolation
Voice Isolation suppresses all noises and background speech, leaving only the agent's voice audible to the customer.
Best for environments where:
Multiple agents sit close together and cross-talk bleeds into calls
The agent works from home with family or housemates nearby
Confidential conversations from adjacent desks could be overheard by the customer
Voice Isolation is the more aggressive mode. It filters out everything that is not the primary speaker's voice.
Noise suppression
Noise Suppression suppresses only background noises — things like keyboard clicking, air conditioning, traffic, dog barking — while preserving human speech.
Best for environments where:
The primary issue is environmental noise rather than other voices
Agents occasionally need nearby colleagues to contribute to a call
The workspace has mechanical noise (printers, HVAC) but limited cross-talk
Noise Suppression is the lighter-touch mode. It cleans up the audio environment without silencing nearby speech.
Selecting the right mode
Choosing the right mode depends on the agent's physical environment. The following comparison helps.
Environment
Recommended mode
Why
Open-plan office, desks close together
Voice Isolation
Prevents cross-talk from adjacent agents
Work-from-home, shared living space
Voice Isolation
Filters household speech and activity
Private office with street noise
Noise Suppression
No cross-talk risk; just environmental noise
Contact center floor, good desk spacing
Noise Suppression
Background hum without significant voice bleed
Quiet home office with occasional pet noise
Noise Suppression
Intermittent noise, no ongoing speech interference
Configuration and control
Audio Enhancement is managed at two levels: administrators enable and configure it, and agents with proper permissions can adjust their own settings.
Administrator setup
Administrators enable Audio Enhancement through User Management settings in the Amazon Connect Customer console. They can:
Enable or disable Audio Enhancement for specific security profiles
Set the default mode (Voice Isolation or Noise Suppression)
Control whether agents can change their own settings
Agent self service
Agents with proper permissions can adjust Audio Enhancement settings from the CCP. Changes take effect on the agent's next call. They can:
Toggle Audio Enhancement on or off
Switch between Voice Isolation and Noise Suppression
This gives agents flexibility to adapt to changing conditions — for example, switching to Voice Isolation when a noisy meeting starts in a nearby conference room.
Impact on call quality metrics
Audio Enhancement improves several operational metrics:
Reduced "I can't hear you" complaints from customers
Lower repeat-request rates (customers asking agents to repeat themselves)
Improved transcription accuracy for Conversational Analytics
Better sentiment scores when audio quality is no longer a frustration factor
RECALL & APPLY
Practice what you learned
Try answering in your own words, then compare with the suggested answer. These activities are for self-study and are not automatically graded.
01How do forecasting and scheduling answer different questions?
Topic: Workforce optimization components
Reveal suggested answer
Forecasting estimates the workload. Scheduling assigns agents and activities to meet that workload within the applicable rules.
02What would you inspect before offering voluntary time off?
Topic: Agent flexibility and intraday operations
Reveal suggested answer
Check the remaining demand, available skills, staffing position, service targets, and applicable rules for the rest of the day.
03Must scheduling wait for a completed capacity plan?
Topic: The workforce planning relationship
Reveal suggested answer
The source explains that capacity planning and scheduling can each consume published forecasts independently. Distinguish the conceptual planning cycle from a required technical dependency.
04Why might a technically valid forecast still need adjustment?
Topic: ML forecasting and workforce challenges
Reveal suggested answer
It may not reflect a future event absent from historical data, such as a product launch or a planned service change.
05A new mortgage product is expected to increase calls. Which teams need the information?
Topic: An integrated workforce example
Reveal suggested answer
Forecasting, staffing and scheduling teams need the expected workload. Training and quality teams need the new contact types and required handling practices.
06Why should an unusual spike be investigated before it is removed?
Topic: How ML forecasting works
Reveal suggested answer
The spike may represent a data error or a real event likely to recur. Those cases require different treatment.
07Which forecast would support a hiring plan rather than today’s queue response?
Topic: Forecast horizons
Reveal suggested answer
A long term forecast supports hiring and capacity decisions. Intraday information supports immediate operational adjustment.
08What would you do if forecasts repeatedly miss a promotion period?
Topic: Forecast quality and limitations
Reveal suggested answer
Review the historical data and the promotion’s effect, then incorporate the known business event into the planning process and assess the revised result.
09Why would two plans based on the same demand forecast require different staffing?
Topic: Capacity planning inputs
Reveal suggested answer
Different service targets, occupancy assumptions, available hours, attrition, or other scenario inputs can change the staffing requirement.
10What makes a capacity plan useful to a manager?
Topic: Capacity planning process
Reveal suggested answer
It identifies the size and timing of staffing gaps and links them to concrete hiring, training, outsourcing, or scheduling decisions.
11Why can stable contact volume still create a staffing shortfall?
Topic: Historical data for forecasting
Reveal suggested answer
Handling time may rise, changing the workload even if the number of contacts is unchanged.
12Why is daily headcount alone insufficient for scheduling?
Topic: Intelligent scheduling
Reveal suggested answer
Demand changes across intervals and contact types. Coverage must match the required times and skills, including planned non-contact activities.
13How should an organization handle competing requests for the same time off?
Topic: Business needs and agent preferences
Reveal suggested answer
Apply the configured rules and the organization’s process consistently, considering coverage, fairness, and any applicable constraints.
14Why can a schedule fail to meet demand despite enough total agents?
Topic: Staff rules and scheduling constraints
Reveal suggested answer
The combination of skills, availability, shift patterns, and staff rules may prevent enough coverage in particular intervals.
15What should the instructor establish before interpreting the chart?
Topic: Intraday forecast performance
Reveal suggested answer
Identify the selected metric, queue or channel, timeframe, comparison, and which points are actuals versus forecasts.
16How would you investigate higher answer times?
Topic: Trend graphs and comparisons
Reveal suggested answer
Review volume, handle time, and effective staffing together. A change in any of them can affect waiting, so avoid assuming one cause from one metric.
17How does a daily projection differ from actual performance so far?
Topic: Published forecasts and daily projections
Reveal suggested answer
It combines observed results with forecast values for the remaining intervals. It is not a completed-day measurement.
18Does a changed forecast automatically update every published schedule?
Topic: Real time schedule adjustments
Reveal suggested answer
The source says supervisors must trigger the relevant refresh. Verify the published schedule and communicate any approved change.
19A team is fully staffed but lacks the required language skill. What is missing?
Topic: Multiskill scheduling and adherence
Reveal suggested answer
The staffing view needs a skill dimension. Total headcount does not show whether the right agents can handle the waiting work.
20Why compare automated and human evaluations during rollout?
Topic: Automated quality management
Reveal suggested answer
The comparison reveals ambiguous criteria, missing context, and scoring errors that need refinement before the results drive decisions.
21Improve the question “Was the agent good?”
Topic: Evaluation form design
Reveal suggested answer
Use a specific observable criterion, such as whether the agent confirmed the customer’s requested next step, with a clear scoring rule.
22What should an agent verify in a draft email response?
Topic: Email support and evaluation follow up
Reveal suggested answer
Check the actual issue, policy applicability, factual details, tone, and any promised action before sending the response.
23Which mode fits nearby conversations that should not reach the customer?
Topic: Audio enhancement modes
Reveal suggested answer
Voice isolation is the mode described for suppressing background speech. Test it with the actual headset and working environment.
24Why should a configuration change be tested on the next call?
Topic: Audio enhancement configuration
Reveal suggested answer
The source describes changes taking effect on the next call. Confirm the active mode and listen for both noise reduction and clear primary speech.
Capture a key idea, an example, or a question for your instructor.
Transform your customer experience with the AI-powered self-service capabilities of Amazon Connect Customer. This course explores how intelligent automation can reduce call volumes, enhance customer satisfaction, and lower operational costs.
You will learn about natural sounding voice interactions using our text-to-speech service, Amazon Polly, and how to design impactful conversational interfaces using Amazon Lex. You will also discover how to deploy generative AI solutions with Amazon Q in Connect to handle customer requests without agent intervention.
Through practical examples and use cases, you will learn to create seamless self-service experiences that satisfy customer needs while optimizing your contact center resources.
Introduction to AI Powered Self Service
Self Service in Modern Contact Centers
Modern AI self service
Imagine you run a retail contact center. During holiday seasons, your call volume triples, mostly with customers asking about order status. Without strong self-service, you would need to hire seasonal staff, which could increase costs and create training challenges. With the self-service capabilities of Amazon Connect Customer, you could automatically handle these common requests, leaving your agents available to tackle more complex issues.
Review the following to learn how modern AI self-service can help:
Traditional limitations – Self-service capabilities have evolved far beyond basic touch-tone menus (for example, "Press 1 for sales, 2 for support").
Modern capabilities – Today's AI-powered solutions can understand natural language, personalize responses, and complete actions on behalf of customers.
Business benefits – For businesses, this means reduced call volumes, lower operational costs, and more satisfied customers.
Self service or human agents
Self-service can handle the standard queries that your contact center receives daily, which leaves human agents free to handle specialized queries. This provides the following typical benefits:
Continuous 24-hour availability of information and assistance
Immediate access to answers without waiting in queues
Consistent experience regardless of time or day
Reduced call volume for routine inquiries and simple transactions
Lower operational costs per customer interaction
Improved agent satisfaction by focusing on complex, value-added tasks
Figure 71 Amazon Connect AI Self Service CapabilitiesSelect image to enlarge
Adding self-service options to call centers helps everyone. Customers get quick answers whenever they need them without waiting in queues for an agent. The contact center saves money by having self-service capabilities handle simple questions while skilled agents focus on specialized problems. The information collected from self-service also helps companies understand what customers want. As more people prefer solving problems quickly on their own, having good self-service options is necessary. Companies that use technology effectively while still providing human help when needed will benefit.
Evolution from Traditional IVR to AI Powered Self Service
Four Types of IVR
Self-service capabilities have evolved beyond basic touch-tone menus and directed dialog options. Today's AI-powered solutions can understand natural language, personalize responses, and complete actions on behalf of customers. For businesses, this can mean reduced call volumes, lower operational costs, and more satisfied customers.
The following sections explore the four types of Interactive Voice Response (IVR).
Traditional IVR
The following are features of the first generation of automated phone systems:
Limited to touch-tone keypads or directed dialog
Required customers to navigate menus by pressing buttons
Included examples such as "Press 1 for sales, 2 for support, 3 for billing"
Customer experience:
System: "Please listen carefully as our menu options have changed." Customer: Sighs and waits to press the right number
Automatic speech recognition ASR
These systems recognize spoken words but not context as described in the following:
Understands natural speech instead of requiring button presses
Matches spoken words to predefined commands
Capabilities limited to specific phrases and commands
Customer experience:
Customer: "I need to update my billing address." System: Recognizes "billing address" and routes accordingly
Natural language understanding NLU
These systems interpret meaning behind words as described in the following:
Identifies customer intent from conversational language
Understands context and variations in expression
Handles multiple intents in a single statement
Customer experience:
Customer: "My internet is really slow today." System: Identifies intent as "technical support – service issue" and routes to the appropriate department automatically
Intelligent action systems
These advanced systems understand and process decisions such as the following:
Processes complete transactions without human intervention
Handles complex security and verification processes
Performs multi-step processes while maintaining context
Customer experience:
Customer: "I'd like to make a payment." System: Verifies identity, checks account balance, processes payment, and sends confirmation
Customer: Completes entire transaction without agent interaction
Compare technologies
Feature
Traditional IVR
ASR
NLU
Intelligent actions
Input Method
Keypad buttons
Voice commands
Conversational language
Natural conversation
Understanding
None
Word recognition
Intent recognition
Context-aware understanding
Customer Effort
High
Medium
Low
Minimal
Resolution Speed
Slow
Medium
Fast
Fastest
Response Latency
Instant (pre-recorded)
Low
Low
Higher (LLM inference)
Business Value
Basic automation
Improved routing
Better experience
Complete automation
By using the intelligent actions modern IVR approach, Amazon Connect Customer empowers customers to speak naturally and get help quickly, instead of navigating through multiple menu layers.
When implementing self-service, companies may worry that customers will feel they are getting a reduced experience compared to talking with an agent. However, customers may prefer the speed and availability of self-service to efficient inquiry resolution.
Think about how you interact with modern voice assistants like Amazon Alexa. You speak naturally, and they understand your intent. Amazon Connect Customer brings this same technology to contact centers, which creates experiences that are natural and helpful rather than robotic and limiting.
Remember that good self-service should not be more complicated. It should be faster and easier than talking to an agent.
How Amazon Connect Customer AI Enhances Self Service Experience
Three Core Services
Amazon Connect Customer self-service capabilities are built on three underlying AWS services: Amazon Polly, Amazon Lex, and Amazon Q in Connect Customer. These three AWS services are pre-integrated with Amazon Connect Customer, so you can access and utilize them from within the Amazon Connect Customer admin workspace.
Amazon Polly
This service converts text to natural-sounding speech, which gives your self-service system a human-like voice. Gone are the days of robotic, monotone IVR systems.
These three services, along with Amazon Connect Customer, work together to create natural, powerful customer experiences.
Amazon Lex
Amazon Lex, a conversational AI service, powers NLU in your self-service solutions. It uses conversational AI technology similar to that powering Amazon Alexa, which makes it possible for customers to speak or type naturally instead of navigating rigid menus.
Traditional IVR systems followed rigid scripts with limited options. Discover how AI transforms this experience in the following comparison:
Traditional IVR: "For account balance, press 1. For payment information, press 2."
AI-powered self-service: "Hi there! How can I help you today?"
Customer: "I want to check my account balance."
AI-powered self-service: "I'd be happy to help you check your balance. For security, can you please verify the last four digits of your social security number?"
The difference is dramatic. Amazon Lex understands the customer's intent, which makes interactions feel more like talking to a helpful person than navigating a machine.
Amazon Q in Connect Customer
This generative AI assistant, the same assistant that helps agents, can be used for self-service. Providing dynamic responses, step-by-step guidance, and can complete actions on behalf of customers.
Self-service capabilities have evolved beyond basic touch-tone and directed dialog menus. Examples include "Press 1 for sales" and "Press or say 1 for sales, 2 for support."
Imagine a customer is contacting their bank in the following example:
Customer: "I noticed a strange charge on my account from yesterday."
Amazon Q in Connect Customer: "I understand you're concerned about an unfamiliar transaction. I can help you review recent charges and place a temporary hold on your card if needed. Would you like me to show you the most recent transactions on your account?"
This level of understanding and assistance was not possible with traditional self-service systems. Amazon Q in Connect Customer doesn't just follow scripts—it comprehends situations and responds appropriately.
Amazon Q in Connect Customer can do the following:
Answer open-ended customer questions using knowledge from your business.
Recommend personalized next steps based on:
What is understood of the customer's current situation
What is known about the customer in Amazon Connect Customer Profiles (this could include products owned, recent orders or contact history) or other data sources
Hand off to deterministic flows, or human agents seamlessly when needed, with full conversation context.
Introduction to Amazon Polly for Text to Speech
Understanding Amazon Polly
Amazon Polly overview
Amazon Polly is an AWS service that converts text into lifelike speech. Your contact center can use Amazon Polly, integrated with Amazon Connect, to deliver natural-sounding voice prompts to callers. Instead of recording every prompt, you can type what you want your IVR to say, and Amazon Polly says it to your customers.
Imagine you manage a bank's contact center. Traditionally, contact centers hired voice actors to record prompts for every account balance, transaction amount, and address. This made customer experiences rigid and changes costly. With Amazon Polly, your system can dynamically generate speech based on real-time data: "Hello Jane, your current balance is $523.45." This creates a personalized experience without extensive recording sessions.
Figure 72 Introduction to Amazon Polly for Text to SpeechSelect image to enlarge
Using Amazon Polly
Amazon Connect Customer uses the built-in, managed integration with Amazon Polly to provide you with the ability to dynamically create speech-based prompts through text. Amazon Polly Neural and Standard voices are included at no additional charge within Amazon Connect Customer.
To learn more about using Amazon Polly in Amazon Connect Customer, review the following steps.
Step 1
In the Amazon Connect Flow designer, select the Play Prompt block from the Interact section of the Block Library.
Add this block to your canvas by dragging it into position or using keyboard navigation.
This block enables text-to-speech capabilities through Amazon Polly.
Step 2
Click on the play prompt block to open the block settings.
Choose the 'Text-to-speech or chat text' option. A text field will appear.
Select the Interpret as option and select Speech Synthesis Markup Language (SSML) to allow you to add SSML tags to modify the speech output.
Add your message such as 'Welcome to customer service. How may I help you today?'
Then click Save.
Text to speech best practices
Text-to-speech (TTS) allows you to instantly update IVR messages, offer personalized self-service, and support multiple languages without the cost of re-recording audio files. However, poorly implemented TTS can sound robotic and damage customer trust. To ensure your prompts sound natural and professional, apply the following best practices.
Script for the ear
When writing TTS prompts, remember that your customer is listening, not reading:
Keep sentences short – Break down complex information into short, conversational sentences to avoid cognitive overload.
Put actions first – Start instructions with the required action. For example, use "Say billing or press 2" instead of "To hear your billing information, please say 2."
Avoid jargon – Use simple, everyday language that the average caller easily understands.
Use punctuation for natural pauses – Insert commas and periods into the text to force the TTS engine to pause appropriately.
Control pronunciation and pacing with SSML
TTS engines can struggle with acronyms, phone numbers, or custom product names. Use SSML to refine the output:
Control pronunciation – Use SSML tags to specify phonetic pronunciations for specialized terms or to spell out characters individually (for example, case numbers).
Emphasize key actions – When listing menu options, present the descriptive phrase first, then speak the digit. For example: "Sales, press 1."
Adjust speaking rate – Slow the rate for critical information such as account numbers or confirmation codes.
Choose and maintain a consistent voice persona
Your TTS voice represents your brand in every customer interaction:
Align with your brand – Choose a TTS voice that matches your brand identity. A casual brand might use a friendly, energetic voice, while a financial institution should use a reassuring, professional tone.
Maintain consistency – Do not mix pre-recorded studio prompts with default TTS voices. Keep the voice and tone consistent across all menus and sub-menus.
Disclose AI usage – State upfront that the caller is interacting with an automated system to preserve trust and set appropriate expectations.
Handle dynamic content and allow interruptions
Personalization and flexibility improve the customer experience:
Personalize intelligently – Use TTS for dynamic, real-time content (for example, "Your account balance is $50.12" or "Your appointment is confirmed for Tuesday"). Do not ask for information the system already has on file.
Allow barge-in – Configure your system to accept customer input while TTS is playing. This lets returning customers skip prompts they have heard before.
Test and provide fallbacks
Thorough testing prevents poor experiences in production:
Test your prompts – Call an assigned phone number in Amazon Connect and listen as a customer would. Test in simulated noisy environments to ensure the generated voice remains intelligible. You can also test directly through Amazon Polly in the AWS console.
Provide a live agent option – Make sure the option to reach a human agent is available on the main menu and that the system routes there if a customer gets stuck.
Consider the full flow – Listen to the entire conversation path from the caller's perspective. Engage a conversational UI designer for expertise in creating pleasant customer experiences.
Voice Selection and Customization
Voice selection
Amazon Polly offers a variety of voices across many languages. Each voice has its own character and tone, so you can select one that best represents your brand. For example, you might choose a warm, friendly voice for a travel agent, or a more professional, authoritative voice for financial services.
Use the Set voice flow block in Amazon Connect to set the text-to-speech (TTS) language and voice for the flow. You can use this flow block to choose from dozens of voices in a variety of languages and accents.
To learn more about the Set voice flow block, review the following key configuration areas:
Language
Set the Language for your flow.
Voice selection
Select one of the voices available for the selected language.
For customers that want to provide an even more unique and brand focused experience to customers, creating your own Polly voice is possible as well.
Customize options
Amazon Connect offers specialized speaking styles for select voices. Examples include the following:
Conversational Style – Sounds like a friendly conversation rather than a formal announcement. This is typically the style recommended for customers for contact center workloads.
Newscaster Style – Mimics the clear, professional delivery of a news broadcaster.
Using SSML for enhanced speech control
Speech Synthesis Markup Language (SSML) gives you additional control over how Amazon Polly pronounces text.
Using SSML in Amazon Connect
Amazon Connect supported SSML tags
Practical customer support example
Using SSML for enhanced speech control
Speech Synthesis Markup Language (SSML) gives you additional control over how Amazon Polly pronounces text.
Using SSML in Amazon Connect
To use SSML in Amazon Connect do the following:
In any text-to-speech flow block, set the Interpret as field to SSML
Wrap your text in tags.
Add SSML tags to customize pronunciation, pauses, and more.
Amazon Connect supported SSML tags
The following is a selection of SSML tags supported in Amazon Connect:
break: Adds a pause (e.g., adds a one-second pause)
prosody: Controls volume, rate, or pitch (e.g., speaking slowly)
say-as: Specifies how to pronounce characters, words, and numbers
phoneme: Makes a specific phonetic pronunciation
For a full list of tags, see SSML tags supported by Amazon Connect in the Amazon Connect Administrator Guide.
Practical customer support example
Review the practical customer support example:
"I'm sorry you are experiencing an issue with your service. Your case number is CX25791."
In this example Amazon Polly is instructed to speak at 90 percent of normal speed. It then adds a short pause before reading the case as individual characters rather than, "CX twenty-five thousand, seven hundred and ninety-one."
Introduction to Amazon Lex for Intelligent Bots and IVR
Amazon Lex Core Concepts and Terminology
Amazon Lex core concepts
Amazon Lex is a fully managed AI service you can use to build conversational interfaces or chatbots for applications. It uses AI technology to recognize customer speech and text input, understand the meaning behind it, and respond appropriately.
When integrated with Amazon Connect Customer, Amazon Lex transforms traditional automated phone menus from rigid, button-pressing experiences into natural conversations.
For example, customers experience, "How can I help you today?" and "I'd like to check my account balance," instead of, "Press 1 for sales. Press 2 for support."
At the heart of Amazon Lex are the following two capabilities:
Natural language processing (NLP) refers to the overall ability of computers to work with human language.
Natural language understanding (NLU) is a subset of NLP that focuses specifically on comprehending the meaning and intent behind what a person says.
Conversational customer experience using AI is not just about recognizing the words but understanding what the customer wants to accomplish. Amazon Lex goes beyond recognizing words in customer statements. When a customer says, "I need to update my shipping address," it understands their intent to change their shipping information.
Examples of Amazon Lex
Banking example
The transition from button-pressing experiences into natural conversation is illustrated through a banking example.
The example starts with pressing 1 on a keypad, progresses to basic voice commands, and culminates in natural conversation. In the natural conversation stage, customers can say something like, "I'd like to transfer money from my checking to savings account." Each stage shows how customer interaction becomes more natural and conversational.
Figure 73 Introduction to Amazon Lex for Intelligent Bots and IVRSelect image to enlarge
The image shows how Amazon Lex processes customer speech input. The diagram uses an example: "I need to update my shipping address to AnyStreet 100." First, automatic speech recognition (ASR) processes the speech. The text moves through NLP and NLU. These systems identify the intent (UpdateShippingAddress) and extract slot information (AddressSlot: AnyStreet 100). Finally, the system fulfills the intent and sends a confirmation to the customer.
Figure 74 Recognizing a change in customer intentSelect image to enlarge
Amazon Lex terminology
You can use Amazon Lex to build applications (bots) to elicit information from users to accomplish a task. For example, you can create a bot to order flowers, book a hotel room, or order a pizza.
There are some key terms to understand when working with Amazon Lex.
Bot
An Amazon Lex bot is powered by ASR and NLU capabilities.
Amazon Lex bots can understand user input provided with text or speech and converse in natural language.
Language
An Amazon Lex V2 bot can converse in one or more languages.
Each language is independent of the others. You can configure Amazon Lex V2 to converse with a user using native words and phrases.
For more information about slots, see Languages and locales supported by Amazon Lex V2 in the Amazon Lex V2 Developer Guide.
Intent
An intent represents an action that the user wants to perform. You create a bot to support one or more related intents. For example, you might create an intent that orders pizzas and drinks. For each intent, you provide the following required information:
Intent name – A descriptive name for the intent. For example, OrderPizza.
Sample utterances – How a user might convey the intent. For example, a user might say, "Can I order a pizza," or "I want to order a pizza."
How to fulfill the intent – How you want to fulfill the intent after the user provides the necessary information. We recommend that you create an AWS Lambda function to fulfill the intent. You can optionally configure the intent so Amazon Lex V2 returns the information back to the client application for the necessary fulfillment.
In addition to custom intents, Amazon Lex V2 provides built-in intents to quickly set up your bot.
For more information about slots, see Built-in intents in the Amazon Lex V2 Developer Guide.
Amazon Lex V2 always includes a fallback intent for each bot. The fallback intent is used whenever Amazon Lex can't deduce the user's intent. For more information about the fallback intent, see AMAZON.FallbackIntent in the Amazon Lex V2 Developer Guide.
Slot
An intent can require zero or more slots or parameters. You add slots as part of the intent configuration. At runtime, Amazon Lex prompts the user for specific slot values. The user must provide values for all required slots before Amazon Lex can fulfill the intent.
For example, the OrderPizza intent requires slots such as pizza size, crust type, and number of pizzas. In the intent configuration, you add these slots. For each slot, you provide slot type and a prompt for Amazon Lex to send to the client to elicit data from the user. A user can reply with a slot value that includes additional words, such as, "large pizza please," or "let's stick with small." Amazon Lex can still understand the intended slot value.
Slot type
Each slot has a type. You can create your custom slot types or use built-in slot types. Each slot type must have a unique name within your account. For example, you might create and use the following slot types for the OrderPizza intent:
Size – With enumeration values Small, Medium, and Large.
Crust – With enumeration values Thick and Thin.
Amazon Lex also provides built-in slot types. For example, AMAZON.NUMBER is a built-in slot type that you can use for the number of pizzas ordered.
For more information about slot types, see Built-in slot types in the Amazon Lex V2 Developer Guide.
Version
A version is a numbered snapshot of your work publishable for various workflow stages. These stages include development, beta deployment, and production.
Once you create a version, you can use a bot as it existed when the version was made. After you create a version, it stays the same while you continue to work on your application.
Alias
An alias is a pointer to a specific version of a bot. With an alias, you can update the version your client applications are using.
For example, you can point an alias to version 1 of your bot. When you are ready to update the bot, you publish version 2 and change the alias to point to the new version.
Because your applications use the alias instead of a specific version, all of your clients get the new functionality without needing to be updated.
Putting Amazon Lex into Action
Now that you have explored the key terminology of Amazon Lex, review how intents and slots interact with each other in Amazon Connect.
Customer "I want to transfer money."
Amazon Lex determines that the customer is wanting to transfer money by the sample utterances provided in the TransferFunds intent. The sample utterances are used to train the bot to allow it to classify the customer's utterance.
Bot "Sure, I can help with that. Which account would you like to transfer from"
Now that Amazon Lex knows what intent the customer needs it can provide clarifying questions to help gather data needed to fulfill the intent.
Slots provide the ability to gather data to fulfill the recognized intent. The next step in the conversation is determined by prompting the user to provide the required data for intent fulfillment. In this scenario, the slot prompts for the user to provide the account details of where they would like funds to be transferred from.
Customer "My checking account." Bot "And which account would you like to transfer to"
After the customer provides a response, Amazon Lex determines whether it needs to fill any further slots. If needed, it will ask the customer to provide the necessary information.
The second slot asks for the destination account.
Customer "My savings account." Bot "How much would you like to transfer"
Amazon Lex continues to ask the customer to provide the necessary information until all slots are filled.
The third slot is for a dollar amount.
Customer: "Five hundred dollars." Bot: "Just to confirm, you want to transfer 500 dollars from your checking account to your savings account. Is that correct?"
Now that all slot data has been gathered, the intent can be fulfilled. In this scenario, a confirmation prompt has been configured to confirm the provided details before completing.
Seeing it in action in the Amazon Lex test window
You can test your bot in the Amazon Lex test window before using it within a live customer journey.
This gives you the ability to ensure that your slots are capturing the correct information.
Insider tip: Confirmation is not required for all intents and should be used selectively. For transactional intents or actions with significant consequences, implement confirmation to validate the bot's intended action before execution.
When using Amazon Lex with Amazon Connect for phone interactions, it benefits from special optimizations:
Telephony audio sampling – Amazon Lex is trained on telephony audio at an 8 kHz sampling rate, which provides increased speech recognition accuracy specifically for phone-based use cases. Because Amazon Lex is optimized to process 8 kHz audio directly, it avoids the quality loss that occurs when upsampling or downsampling between different rates, providing reliable speech recognition accuracy for phone calls.
Dual tone multi-frequency (DTMF) input – In addition to voice recognition, Amazon Lex can accept numeric input from keypad presses, which gives customers flexibility in how they respond.
Understanding 8 kHz telephony audio
The 8 kHz sampling rate (narrowband audio) is the standard for traditional telephone networks (PSTN). This limits the frequency range to a maximum of 4 kHz, which is sufficient for speech intelligibility but removes higher-frequency consonant detail.
Why this matters for your bot design:
Consonant clarity – Frequencies above 4 kHz are filtered out, which can make it harder to distinguish between similar-sounding consonants (for example, "s" versus "z" or "f" versus "th").
Background noise tolerance – In noisy environments, narrowband audio provides less separation between the caller's voice and ambient sounds.
Speech recognition accuracy – Amazon Lex is specifically trained on 8 kHz audio, which mitigates these limitations for telephony use cases. You can further improve accuracy by configuring a custom vocabulary for specialized terminology.
The evolution - Modern carriers increasingly support wideband audio (16 kHz) through VoLTE and HD Voice, which doubles the available frequency range. Amazon Connect supports these higher-fidelity connections when the caller's network provides them.
Best practice - Use an Amazon Lex custom vocabulary to help the transcription layer recognize specialized terminology (industry jargon, product names, or acronyms) that may be harder to distinguish at 8 kHz.
Building Conversational Interfaces
Key Principles
Building Conversational Interfaces
Building effective conversational interfaces is about creating a natural dialogue that feels helpful and intuitive. Here are some key principles to keep in mind:
Be human-centered – Design with the customer's needs and expectations in mind.
Keep it simple – Use clear, straightforward language.
Provide context – Help customers understand where they are in the conversation and what's happening next.
Handle errors gracefully – Plan for misunderstandings and create friendly ways to get the conversation back on track.
Confirm important information – Confirm details for critical actions before proceeding.
Well-designed conversations anticipate and adapt to customer needs. The image shows this through a typical banking exchange. A customer asks to check their account balance. The bot responds by asking which account type. When the customer changes their mind, the bot seamlessly transitions to help with a money transfer instead.
Figure 75 Building Conversational InterfacesSelect image to enlarge
Example conversational interaction between a customer and bot.
Map customer journeys – Identify the most common reasons customers contact you.
Draft sample dialogues – Write out example conversations for each scenario.
Plan for detours – Customers often change their minds or ask tangential questions.
Conversational interface designers
A conversational designer creates intuitive, engaging interactions between users and AI-powered interfaces. They design natural interactions for chatbots, voice assistants, and other conversational tools. Their work combines elements of user experience (UX) design, linguistics, psychology, and copywriting to craft dialogue flows that align with user needs and brand goals.
Designing conversational interfaces requires shifting from traditional visual UI design to natural human interaction patterns. The goal is to make the interaction feel organic, efficient, and low-effort for the user. The following best practices apply whether you are designing for voice (IVR and voicebots) or text (chatbots and AI assistants).
Voice design
Voice interfaces process information sequentially. Callers cannot scan back through what they heard, so every design decision must account for linear delivery and limited short-term memory.
Structure for cognitive load. Use the "one breath" rule: prompts should take fewer than four seconds to speak aloud. Front-load the critical action or keyword at the beginning of each sentence. For example, use "To book a flight, say flight" rather than "Say flight if you would like to book a flight."
Limit options per decision point. Offer a maximum of three to four choices at any single prompt. Long lists cause callers to forget the first options by the time they hear the last ones. If you need more options, organize them into categories and let callers navigate progressively.
Support barge-in. Allow callers to interrupt a prompt the moment they hear the option they want. Forcing a caller to listen to a 20-second recording before responding wastes their time and increases abandonment.
Handle over-answering. If the system asks "What date would you like to travel?" and the caller says "Next Tuesday with my wife," the system should extract the date and store the extra context (two tickets) rather than returning an error.
Text design
Text interfaces give users control over pacing. They can read at their own speed, scroll back, and see visual anchors. This creates different design opportunities and constraints.
Use visual structure. Break responses into short paragraphs. Use bullet points for lists and bold text for key terms. Callers read on small screens, so concise formatting improves comprehension.
Provide interactive elements. Buttons, quick-reply chips, and carousels reduce typing effort and guide users toward valid responses. These visual aids have no equivalent in voice channels.
Allow longer responses when appropriate. Unlike voice, text interfaces can present a full paragraph or a comparison table without overwhelming the user. However, keep each message focused on a single topic to maintain clarity.
Support rich media. Link to images, documents, or calendar widgets when they help the user complete a task faster than text alone.
Amazon Lex and Amazon Connect Integration
Integration Overview
Amazon Lex and Amazon Connect Integration
Amazon Connect and Amazon Lex work seamlessly together to create a powerful platform for conversational customer service.
With this integration, you can do the following:
Replace rigid IVR menus with natural conversations.
Automate routine customer inquiries and transactions.
Collect information before transferring to agents.
Provide 24/7 self-service options.
Deliver consistent experiences across voice and chat.
Building and Testing an Amazon Lex Bot in Amazon Connect
Transcript Building and Testing an Amazon Lex Bot in Amazon Connect
Create and configure the bot
AWS has streamlined the Amazon Lex bot creation process, so you can build complete conversational experiences without leaving Amazon Connect. First, you will need to log into your Amazon Connect admin workspace by navigating to your instance URL, such as https://yourinstancename.my.connect.aws. You will either need the Admin security profile or make sure your security profile includes the "Channels and Flows - Bots - Create" permission.
Once logged in, navigate to the bot creation area. In the navigation menu, choose on Routing, then select Flows.
Now on the Flows page, you can see a Bots tab. Choose Bots, then select Create bot.
Excellent! The Details dialog box has opened. Now configure our bot with the following information. For bot Name, enter "BankerBot". Remember, this needs to be a unique name within your AWS account. Bot Description is optional, but adding context is recommended. Enter, "Bot to help with banking actions". For compliance with COPPA, or the Child Online Privacy Protection Act, for this banking bot, select No. Since this application will not target children under 13. Then, choose Create to proceed.
Perfect! Amazon Connect has successfully created the bot and directed us to the bot configuration page. As you can see, the BankerBot is ready for configuration. The next step is adding language support. To choose the initial language to use for the bot, choose Add language.
You can add multiple supported languages to your bot, but for this demonstration, choose English (US) as the initial language.
Define the transfer intent
Now that language has been added, you can add your intents, which represent the goals your users want to accomplish. Remember, there are two types of intents. Custom intents represent specific actions your bot should handle. Built-in intents are for common actions. Every bot automatically includes a fallback intent for unrecognized requests.
To create your first custom intent, choose Add intent, then choose Add empty intent.
In the Add intent dialog, enter "TransferMoney" for the Intent Name and "Providing the ability for customers to transfer funds from one account to another" for the Description. Then, choose Add.
Great! Your intent has been created. Now you can configure the utterances. These are phrases users might say to trigger this intent. Choose Add and enter phrases like "Please transfer money from { FromAccount } to { ToAccount } for { Amount }", then choose Add. You can add phrases like "I want to transfer some money" and so on. Any word wrapped in parentheses represents a slot parameter. You can add those next. Once you have entered all of your phrases, choose Save.
Next, you can configure the slots. Slots are parameters required to fulfill the intent. To start defining the slots, choose Add. Remember the slot parameters are the words you added in the utterance phrases that were wrapped in parentheses. Start with Amount, setting the slot Name, then select the slot Type as AMAZON.Number. Also, add the slot Prompt to "How much would you like to transfer?". Finally, choose Add.
Then, repeat the process for the FromAccount slot, and the ToAccount slot. Set the slot types to AMAZON.Number and add the Slot Prompts of, "Which account would you like to transfer funds from?" and "Which account would you like to send funds to?". Then, choose Add for each slot, then Save when finished.
Now you can configure the prompts. To configure the messages your bot will use, choose Edit. For the Confirmation prompt message, enter "Would you like me to proceed in transferring { Amount } from { FromAccount } to { ToAccount }?" For the Decline response, enter "Okay, I won't transfer the money.". When you are finished, choose Save.
The last step in the process of building a bot is to build the bot's language. By building the bot's language, you can use the bot within our Amazon Connect Flow that we will configure. To start the build process, select the Build language button. Shortly after that the build will be complete. That's how you build a complete Amazon Lex bot directly within Amazon Connect. You've successfully created a functional financial services bot with intents, utterances, slots, and conversational prompts - all without leaving the Amazon Connect interface.
Build the contact flow
Now navigate to the Flows management area. In the navigation menu, choose Routing, then choose Flows.
On the Flows page, choose Create flow.
In the flow designer, you can enter the Flow name "Banking bot flow".
In the flow designer, drag the Play prompt block from the Interact section of the Block Library onto the canvas.
Drag the arrow from the Entry point block and connect it to the Play prompt block.
Next, choose the Play prompt block to open the block configuration pane on the side of the screen. Then choose the Text-to-speech or chat text option.
Enter your welcome message into the textbox. In this demonstration you can welcome the customer with, "Welcome to the example Banking Bot," and then choose Save.
Now, from the Block Library, drag a Get customer input block from the Interact section, onto the canvas.
Connect the Success branch of the Play prompt block to the input of the Get customer input block. Then choose the Get customer input block to open the configuration pane. Now choose the Amazon Lex tab under the Config tab.
Next, choose the Select a Lex bot list, and choose your Amazon Lex bot. Select Banker Bot.
Next choose the Alias list and select your bot alias. In this demonstration, you can use the TestBotAlias. Remember that the TestBotAlias should not be used for production traffic.
Next, under the Customer prompt or bot initialization section, choose the Text-to-speech or chat text option and provide the text you want to play to the customer when they meet your Amazon Lex bot. For this demonstration, enter "How can I help you today?".
Under the Intents section, choose the Add an intent link twice to add two empty intent branches. Next you need to type the names of the intents from the BankerBot Amazon Lex bot. You can enter TransferMoney and FallbackIntent into the empty intent branch boxes. Then when complete, choose Save.
You can now see that the Get customer input block has updated with the two intent branches that you added.
From the Block Library, drag a Play prompt block from the Interact section, onto the canvas. Connect the TransferMoney branch of the Get customer input block to the new Play prompt block. Next, choose the Play prompt block to open the block configuration pane on the side of the screen. Select the Text-to-speech or chat text option. In the textbox, enter the message to give the customer once they have completed the TransferMoney intent. In this demonstration, you can say, "Thank you for contacting us." and then choose Save.
Drag another Play prompt block from the Interact section, onto the canvas. Connect the FallbackIntent branch of the Get customer input block to the new Play prompt block. Then choose the Play prompt block to open the block configuration pane on the side of the screen. Select the Text-to-speech or chat text option. In the textbox, enter the message to give the customer if they hit the Fallback intent as you will transfer them through to an agent. In this demonstration, you can say, "I will now transfer you to an agent." and then choose Save.
From the Block Library, under the Set section, drag a Set working queue block onto the canvas. Connect the FallbackIntent linked Play prompt block Success branch to the Set working queue block. Then, choose the Set working queue block to open the block configuration pane on the side of the screen. Select the Search for queue list. Then choose the BasicQueue and choose Save.
From the Block Library, under the Terminate section, drag a Transfer to queue block onto the canvas. Connect the Success branch from the Set working queue block to the Transfer to queue block. Then drag a Disconnect block onto the canvas. Connect the At capacity and Error branches from the Transfer to queue block to the Disconnect block.
Now, connect the Error branch from the Set working queue block to the Disconnect block.
Then, connect the Error branch from the FallbackIntent linked Play prompt block to the Disconnect block. Next, connect the Success branch and the Error branch from the TransferMoney linked Play prompt block to the Disconnect block. Now, connect the Default branch of the Get customer input block to the FallbackIntent linked Play prompt block. And the Get customer input block Error branch to the Disconnect block. Then, connect the Error branch from the welcome message Play prompt block to the Disconnect block.
Publish and test
Finally, you need to publish the flow so that it can be used in a customer journey. Only published flows can be assigned to a phone number or configured in a communications widget. In the corner of the flow designer, choose the Publish button.
At the Publish confirmation dialog, choose Publish to activate the flow immediately.
Now the flow is active, so you can test the complete customer experience. From the navigation menu, choose Home to go to the Configuration guide.
In the Configuration guide, under Step 1. Explore your channels of communication, choose the Test chat link to go to the Test chat page.
On the Test chat page, choose the Test Settings link, which will open the Test Chat Settings dialog. Under System Settings, select the Contact Flow list and find the contact flow that was just published. In this demonstration, you can choose the Banking bot flow that was just built and published. Then, choose Apply.
The test chat widget will load and enter the flow that was selected. In this demonstration, you're presented with the configured welcome message and the message that was configured in the Get customer input block. The flow is now waiting for an input, so you can enter, "I want to transfer some money", one of the sample utterances that was configured in the BankerBot.
The BankerBot has understood the request and recognized that you want to transfer some money by selecting the TransferMoney intent. Now the Slots parameters need to be filled to be able to fulfil the intent. You can provide the account number "123456789" as a source account and send the message back to the bot.
Now the ToAccount slot needs to be filled. You can provide the account number "987654321" as the destination account and send the message back to the bot.
The third slot, Amount, provides the remaining customer prompt. You can enter the amount "12345".
The last prompt back from the bot is the Confirmation prompt, asking to confirm the transfer of the funds. You can send, "yes", back to the bot.
The bot has collected and confirmed the transfer details, and the sample flow has completed. Executing a real transfer requires an authorized backend fulfillment integration.
After implementing your Amazon Lex bot in Amazon Connect, test thoroughly and optimize over time.
Important language considerations
Figure 76 Building Conversational InterfacesSelect image to enlarge
When using an Amazon Lex V2 bot, the language attribute in Amazon Connect must match the language locale used to build your Amazon Lex bot. For example, if your Lex bot uses Australian English (en_AU), you must configure Amazon Connect to use the same language locale using either of the following:
A Set voice block to indicate the Amazon Connect language model
A Set contact attributes block to specify the language
This ensures that speech recognition works correctly for your customers.
Optimizing Self Service with Generative AI
Generative AI Capabilities for Amazon Lex
Applying Generative AI with Lex Bot Creation
Optimize your Amazon Lex bot creation and self-service performance by using generative AI. You can take advantage of the Amazon Bedrock generative AI capabilities to automate and speed up your Amazon Lex bot building process. Amazon Bedrock generative AI capabilities for Amazon Lex can generate intents, slots, and sample utterances automatically.
Amazon Lex currently provides the following generative AI capabilities to optimize Amazon Lex V2 bots:
Create new bots and populate them with relevant intents and slot types efficiently using natural language description.
Generate sample utterances for your bot intents automatically.
Improve your bots' slot resolution performance.
Create an intent to help answer your customer questions.
Use Amazon Bedrock Agents and Amazon Bedrock knowledge bases to help answer your customer's questions.
Improve intent classification and slot resolution.
Note: These features use generative AI. As you use Amazon Lex, remember that it may give inaccurate or inappropriate responses. For more information, see AWS Responsible AI Policy.
Powered by Amazon Bedrock: AWS implements automated abuse detection. Amazon Lex V2 generative AI features are built on Amazon Bedrock. As a result, users inherit Amazon Bedrock controls for enforcing safety, security, and responsible AI use.
You can activate generative AI capabilities for Amazon Lex V2 either through the console or API.
Using the console
Sign in to the AWS Management Console and open the Amazon Lex V2 console at https://console.aws.amazon.com/lexv2/home.
Select the bot and the locale in the bot for which you want to turn on generative AI capabilities.
In the Generative AI configurations section, select Configure.
Toggle the Enabled button for each feature that you want to activate. Select the model and version that you want to use for that feature. Enabling a feature may incur additional charges.
Select Save after you turn on the features that you want to activate. A green success banner appears to confirm that the capabilities are turned on.
Using the API
To enable generative AI capabilities for a new bot, use the CreateBot operation to create a new bot.
Send a CreateBotLocale request, modifying the generativeAISettings object as necessary. If you are enabling the capabilities for an existing bot, send an UpdateBotLocale request instead.
To enable usage of the descriptive bot builder, modify the descriptiveBotBuilder object. Specify the foundation model to use in the modelArn field and set the enabled value to True.
To enable slot resolution improvement, modify the slotResolutionImprovement object. Specify the foundation model to use in the modelArn field and set the enabled value to True.
To enable sample utterance generation, modify the sampleUtteranceGeneration object. Specify the foundation model to use in the modelArn field and set the enabled value to True.
Generative AI Capabilities in Detail
Review each of the capabilities to learn how generative AI can help you build effective customer experience journeys using Amazon Lex V2.
Descriptive bot builder
Create a bot by using natural language to describe what the bot should be able to do. Amazon Lex V2 invokes Amazon Bedrock models to generate intents and slot types that fit your bot's use case.
Here are some helpful example bot descriptions you can use with descriptive bot builder in Amazon Lex V2:
Industry
Example prompt
Financial services
"Our financial card service assists users with essential tasks for new cards. These tasks include card activation, PIN delivery through email or mail, and card verification using a zip code. We help customers with tasks associated with their existing credit cards. These tasks include inquiring about benefits, reporting lost cards, requesting new cards, resetting PINs, and paying bills."
Food services
"I want a bot to help customers order food (using item ID, quantity, size), check order status, and cancel an order. Use Order ID for indexing orders."
Airline
"We are an airline domain that helps users book flight tickets and manage their reservations. Our services include checking reservation details, obtaining receipts, inquiring about flight status, rescheduling, eliciting flight details, and canceling booked flights. You can also generate additional intents if they help support functions in the domain description."
Insurance
"We are an insurance company that sells car, home, and annuity insurance policies. I want a bot that can check claim status, file a claim, make policy payments and cancel a policy. We use policy_id and last four digits of the Social Security Number (SSN) for account identification and validation."
Vehicle management
"We are building a Towed Cars Lookup bot that helps drivers whose car has been towed to find where the car is located. This bot should ask for the address or location where the automobile was towed from. It should also gather details about the vehicle, including its license plate, make, model, and year."
Travel
"I am a travel agent, and I want a bot to help my customers book a trip to a AnyCompany theme park. AnyCompany has several parks all over the world to choose from and also has hotels, dining, and special entertainment that can be reserved."
For more information, see Use a description to build a bot in Lex V2 with the descriptive bot builder in the Amazon Lex V2 Developer Guide.
Utterance generation
Use utterance generation to automate the creation of sample utterances for your intent.
Amazon Lex V2 generates sample utterances for you based on the intent name, description, and existing examples. This feature reduces the time and effort you spend in discovering and writing your own sample utterances.
For more information, see Use utterance generation to generate sample utterances for intent recognition in the Amazon Lex V2 Developer Guide.
Assisted slot resolution
You can improve the accuracy of some built-in slots in your bot's conversation flow by using assisted slot resolution.
Assisted slot resolution uses Amazon Bedrock large language models (LLMs) to improve recognition of some built-in slots, which results in an improved interpretation of customer responses during slot elicitation. For utterances that could not be resolved normally, you can attempt to resolve them a second time using Amazon Bedrock.
With assisted slot resolution, you can use the power of Amazon Bedrock foundation models to improve the accuracy of the following built-in slots:
AMAZON.Alphanumeric without regex support
AMAZON.City
AMAZON.Country
AMAZON.Date
AMAZON.Number
AMAZON.PhoneNumber
AMAZON.Confirmation
For more information, see Using assisted slot resolution to clarify slot values in Amazon Lex V2 in the Amazon Lex V2 Developer Guide.
QnAIntent
Amazon Lex V2 offers a built-in AMAZON.QnAIntent that you can add to your bot. This intent harnesses generative AI capabilities from Amazon Bedrock by recognizing customer questions and searching for an answer in a knowledge store you have set up.
For example, "Can you provide me details on the baggage limits for my international flight?"
This feature reduces the need to configure questions and answers using task-oriented dialogue within Amazon Lex V2 intents. This intent also recognizes follow-up questions such as, "What about domestic flights?" based on the conversation history and provides the answer accordingly.
For more information, see AMAZON.QnAIntent in the Amazon Lex V2 Developer Guide.
Amazon Bedrock Agents
Use Amazon Bedrock Agents to handle complex workloads requested by customers without having to go through a comprehensive task definition process. Amazon Lex V2 offers a built-in AMAZON.BedrockAgentIntent that you can add to your bot.
This intent harnesses generative AI capabilities from Amazon Bedrock by recognizing customer requests, analyzing them, reasoning them, and responding. It also has the capability to ask any follow-up questions to achieve the task needed.
For example, imagine you defined a retail agent that can check customer's order status. When the customer asks for order status, agent first requests customerId or associated emailId to retrieve the details and responds with correct order status. You can also decide to integrate your AMAZON.BedrockAgentIntent with an Amazon Bedrock knowledge base to directly answer any customer queries.
For more information, see Using BedrockAgentIntent to use a Amazon Bedrock Agent in Amazon Lex V2 in the Amazon Lex V2 Developer Guide.
Improve intent classification and slot resolution
Assisted NLU is a feature that uses LLMs to improve Amazon Lex V2 intent classification and slot resolution capabilities. It enhances accuracy while staying within your bot's configured intents and slots.
The feature does not generate or modify any bot content. This feature helps to improve the overall accuracy of the NLU system, resulting in a more seamless and effective conversational experience for users.
For more information, see Improve intent classification and slot resolution in Lex V2 with assisted NLU in the Amazon Lex V2 Developer Guide.
By using the generative AI capabilities for Amazon Lex, you can significantly transform and enhance your self-service customer experience. This streamlines bot development, automatically generates comprehensive dialogue components, and continuously improves accuracy through advanced NLU. This powerful combination reduces operational costs and accelerates time-to-market for new self-service solutions. Most importantly, it delivers more intuitive and effective customer interactions. As you implement these generative AI features, you can create responsive bots that better understand customer intent and provide more accurate, contextual responses. This leads to higher customer satisfaction and reduced support workload for your organization.
Amazon Lex Next Generation NLU and Voice Options
LLMs as Primary NLU
Three major advances in Amazon Lex change how you design and build conversational self-service experiences: LLMs now serve as the primary engine for understanding customer intent, third-party voice AI models can be integrated alongside native AWS services, and language support has expanded considerably.
Earlier in this course, you learned how Amazon Lex uses natural language understanding (NLU) to identify customer intent and extract slot values. Traditionally, this required defining intents with sample utterances and configuring slots manually. That approach still works, but a newer and more powerful option is now available.
Amazon Lex now offers large language models as the primary option for understanding customer intent. This changes the development experience in a few important ways:
What changes for developers
With LLM-powered NLU, you no longer need to provide exhaustive lists of sample utterances for each intent. The model generalizes from fewer examples and handles unexpected phrasing more gracefully.
A customer might say, "Hey, my internet has been acting weird since Tuesday and I'm pretty frustrated." Traditional slot-based NLU might struggle to map this to the correct intent. LLM-powered NLU interprets the meaning naturally and routes it appropriately.
What changes for customers
Customers can speak naturally without adapting their language to match what the system expects. Multi-part requests work without requiring separate conversation turns. The system maintains context better across longer exchanges.
When to use each approach
LLM-powered NLU is the recommended default for new bots. Traditional slot-based NLU remains appropriate when:
You need highly predictable, deterministic routing
Your use case has a small number of well-defined intents
Amazon Lex integrates with Amazon Bedrock through built-in native features and custom AWS Lambda hooks. This integration allows you to enhance Lex with large language models (LLMs) to handle complex, unscripted customer conversations.
Runtime integration answering users
When a user speaks or types to Amazon Lex, Bedrock handles the open-ended conversation layers through two native built-in intents:
AMAZON.QnAIntent (Retrieval-Augmented Generation) – You can link a Lex intent directly to an Amazon Bedrock Knowledge Base (which typically points to company documents vectorized in Amazon OpenSearch). When a user asks a question, Lex intercepts it, queries Bedrock, retrieves the document context, runs it through an LLM, and returns the summarized answer to the customer without requiring custom backend code.
AMAZON.BedrockAgentIntent (complex orchestration) – If you require your bot to execute multi-step tasks (for example, "Book a flight and email me the confirmation"), Lex can delegate the conversation to an Amazon Bedrock Agent. The agent breaks down the request, determines what APIs to call, and prompts the user for missing details dynamically.
Guardrails and fallback using AWS Lambda
If you require strict control over when the LLM is allowed to respond to your customer, you can route the integration through an AWS Lambda function:
Intent shield – Lex handles deterministic tasks first (for example, collecting account numbers, capturing alphanumeric codes).
Bedrock fallback – If the user says something unexpected, Lex triggers its FallbackIntent. The Lambda function forwards the transcript to Amazon Bedrock, validates the response, and passes it back to Lex for delivery.
Automated bot building developer layer
Amazon Bedrock also assists within the AWS Console to reduce bot development time:
Descriptive bot builder – You describe your ideal bot using natural language, and Bedrock automatically generates the required Lex intents, slots, and utterances.
Assisted slot resolution – If a customer inputs a variation of a slot that your bot does not strictly recognize, Bedrock uses an LLM to infer the meaning. For example, if the slot expects "Large Pizza" and the caller says, "Give me the biggest size you have," the LLM maps that phrase to the "Large" slot value.
How the voice pipeline works
When a user calls your IVR or talks to your voicebot, the interaction flows through a three-step pipeline:
Speech-to-text (STT) – Amazon Lex takes the incoming audio from the caller and passes it through its built-in speech recognition engine to convert spoken words into text.
LLM processing – Lex takes that text string, along with any prompt context or instructions, and passes it to Amazon Bedrock via API. The LLM analyzes the text and generates a text-based response.
Text-to-speech (TTS) – The LLM's text output is sent back to Lex, which pushes it to Amazon Polly to convert back into audio for the caller.
Important: In the traditional pipeline, Amazon Lex transcribes audio to text before any data is sent to an LLM in Amazon Bedrock. This means the LLM cannot hear background noise, detect tone or sarcasm, or sense urgency directly from the audio. It receives only the transcribed words. With Nova Sonic integration, speech-to-speech processing bypasses this step.
Third Party STT and TTS Integration
Amazon Connect no longer requires you to use only AWS-native models for speech processing. You can now integrate third-party AI providers for both speech-to-text and text-to-speech in self-service interactions.
Deepgram for speech to text
Deepgram can serve as an alternative to Amazon Transcribe for converting customer speech into text. Organizations might choose Deepgram for:
Specific accuracy advantages in certain accents or domains
Faster processing speed for particular use cases
Existing enterprise agreements or familiarity
ElevenLabs for text to speech
ElevenLabs can serve as an alternative to Amazon Polly for generating spoken responses. Organizations might choose ElevenLabs for:
Highly natural, expressive voice synthesis
Custom voice cloning for brand consistency
Specific voice characteristics not available in native options
Choosing between native and third party
Expanded Language Support
The wait and continue functionality in Amazon Lex, which keeps customers engaged during processing delays, now supports 10 additional languages.
The following languages are now supported for Lex wait and continue interactions:
Chinese (Mandarin)
Japanese
Korean
Cantonese
Spanish
French
Italian
Portuguese
Catalan
German
This expansion enables organizations to build multilingual self-service experiences that maintain engagement across a broader range of customer populations without requiring separate bot configurations for each language.
Introduction to Generative AI Self Service with Amazon Q in Connect
The Power of a Unified AI Assistant
Rather than building separate generative AI tools for agents and customers, Amazon Q in Connect offers a unified approach. The same AI assistant that helps your agents find information and resolve customer issues can now interact directly with your customers through your IVR and digital channels. Your carefully curated knowledge base, established integrations, and refined configurations for agent assistance are valuable assets. These resources can be used to create exceptional self-service experiences for your customers.
Think about your current contact center operations. When a customer calls, they might start in your IVR system, potentially transfer to a virtual agent, and sometimes need assistance from a human agent. Traditionally, each of these touch points might use different technologies and access different knowledge sources. With Amazon Q in Connect, you are creating a seamless experience powered by a single, intelligent AI assistant that maintains context and consistency throughout the customer journey.
Customer Self Service Interactions with Amazon Q in Connect
How Amazon Q in Connect Works for Self Service
At its core, Amazon Q in Connect uses large language models (LLMs) to process and understand customer inquiries. When implemented for self-service, Amazon Q in Connect works much like it does for your agents. Amazon Q in Connect provides a customer-facing interface through your IVR or digital experience for self-service. Amazon Lex with Amazon Q in Connect transforms how these interactions happen by using the knowledge articles stored in your knowledge base. The same knowledge base that is established for your agents.
Watch the following video to see how Amazon Lex and Amazon Q in Connect create natural, helpful conversations for your customers.
Video: Video - Customer Self-service with QiC.mp4
Transcript Customer Self Service Interactions with Amazon Q in Connect
Let's follow Sofía, as she explores home insurance options with AnyCompany insurance.
She starts on the AnyCompany website, and opens the embedded chat interface. Because Sophia visited before and provided her information, she's greeted by her name.
Sofía starts by saying that she recently bought a home and wants to know how she can get the right coverage.
The answers provided are powered by a generative AI assistant for customer service, Amazon Q in Connect. Amazon Q in Connect is powered by your knowledge articles to provide real-time assistance to both customers and agents across chat interactions like you see here, as well as voice calls.
Amazon Q in Connect can be turned on in your self-service experience in just a few steps and can provide the right answers while carrying on a conversation.
Here, Amazon Q in Connect can provide an overview about the available home insurance coverages, while being able to dive into follow-up questions about the differences.
Amazon Q in Connect can be personalized, to give you control over its tone and behavior.
You can also incorporate data to augment your knowledge articles and tailor the conversation in real time.
Amazon Q in Connect also has the ability to implement guardrails. Amazon Q in Connect can acknowledge when a licensed agent is needed to provide a definitive answer and will provide that option to Sofía.
Sofía isn't quite ready to talk to someone, but she does want to make sure that when she is ready, she doesn't have to go through this process again. So Sofía fills out some key information through this self-service experience provided by Amazon Connect step-by-step Guides.
AnyCompany insurance can securely collect and process sensitive information such as Social Security numbers, credit card details, addresses, and more. This ensures that the agent doesn't have to be exposed to this kind of data.
Details that have already been provided through this conversation are prefilled but editable. And Sofía just has to add the additional information about either herself or her property.
Sofía gets to the last section, but she gets interrupted by an incoming call. Distracted, Sofía doesn't fully complete the form.
But that's not a problem. This information will be stored and ready to continue when Sofía comes back to get the quote that she's looking for.
The conversation flows naturally because Amazon Q in Connect understands context and maintains conversation history. It can handle follow-up questions and even proactively offer related information about delivery options or return policies.
When to use Amazon Lex versus Amazon Q in Connect
Amazon Lex and Amazon Q in Connect serve different purposes in your self-service architecture. Understanding this distinction helps you select the right tool for each customer need.
Q&A and knowledge assistance (searching documents, generating answers)
How it operates
Intent-based: relies on designed flows to gather specific data slots
Search-based: reads unstructured data and summarizes answers dynamically
Primary user
The end-customer (acting as IVR voicebot or web chatbot)
The human agent (populating their screen) or the customer for open FAQs
Strengths
Strict transactional operations, data validation, deterministic routing
Open-ended questions, synthesizing information from multiple sources
Limitations
Requires a developer to code intents for each scenario
Not designed for strict transactional workflows requiring validation
AMAZON.QinConnectIntent, which passes the question to Amazon Q in Connect
Think of Amazon Lex as an automated front desk clerk that excels at executing specific tasks in sequence. Think of Amazon Q in Connect as an expert research assistant that excels at parsing extensive documentation and synthesizing natural language answers.
Key benefits of using Amazon Q in Connect for self-service:
Improved customer experience: Customers receive immediate, accurate responses without waiting for agents to become available.
24/7 availability: Self-service options remain available around the clock.
Reduced contact center load: By handling routine inquiries automatically, Amazon Q in Connect frees up human agents for more high-value interactions.
Consistent information: Using the same knowledge base that your agents use, customers receive the same accurate information at all times.
Cost efficiency: Automated handling of routine inquiries can significantly reduce operational costs.
Scalability: Build your knowledge base once and use it across your entire contact center environment.
By implementing Amazon Q in Connect for customer self-service interactions, you are creating a seamless, intelligent experience that mirrors the quality of human agent assistance. This powerful solution uses your existing knowledge base to deliver consistent responses 24 hours a day, while maintaining natural conversation flow and context awareness. This results in improved customer satisfaction through instant access to information, reduced wait times, and personalized interactions that can seamlessly transition to human agents when needed. As demonstrated in Sofía's journey, customers can engage naturally, receive comprehensive answers, and even pause and resume their experience. Your organization benefits from reduced operational costs, increased scalability, and the ability to handle more complex inquiries with your human agents. Amazon Q in Connect transforms self-service from a basic FAQ system into an intelligent, conversational experience.
Understanding when to use Amazon Lex versus Amazon Q in Connect
Amazon Lex and Amazon Q in Connect serve different primary purposes within Amazon Connect self-service. Understanding the distinction helps you select the right tool for each customer scenario and avoid building duplicate solutions.
Amazon Lex
Amazon Q in Connect
Amazon Lex is a conversational engine for structured, intent-driven actions. Think of it as an automated front desk clerk that follows a defined sequence of steps to complete a specific task.
Primary role: Execute deterministic self-service transactions such as booking appointments, resetting passwords, checking account balances, and routing calls.
How it operates: Lex relies on configured intents, slots, and fulfillment logic. A customer says "I want to check my balance." Lex matches the CheckBalance intent, prompts for the account number slot, verifies the PIN slot, calls a backend system via AWS Lambda, and reads the result back to the caller.
Strengths: Strict transactional workflows, data validation, slot capture, and predictable routing logic.
Limitation: If a customer asks an open-ended question like "What is your return policy for damaged goods bought during a holiday sale?" Lex cannot answer unless a developer specifically coded an intent for that scenario.
Amazon Q in Connect
Amazon Q in Connect is a generative AI knowledge assistant that retrieves information and resolves issues through search and reasoning. Amazon Q in Connect directly helps customers resolve their issues through automated self-service interactions.
Primary role: Answer open-ended questions, search unstructured knowledge sources, and provide contextual guidance without requiring pre-built intent flows.
How it operates: Amazon Q in Connect reads your connected knowledge sources (documentation, policy files, product catalogs) and synthesizes natural language answers in real time. It can also perform actions defined through tools, such as checking order status, processing returns, or scheduling callbacks.
Strengths: Handling ambiguous questions, synthesizing information from multiple documents, maintaining conversational context across follow-up questions, and adapting responses based on customer profiles.
Customer-facing capabilities: When exposed directly to customers, Amazon Q in Connect handles unscripted scenarios. A customer can ask "Can I return a laptop I bought three weeks ago if I lost the receipt?" and receive a contextual answer drawn from your return policy documentation.
When to use which, and how they work together. You do not have to choose one over the other. Amazon Connect supports both in a unified experience. Use Amazon Lex for structured transactions where you need predictable data capture and deterministic fulfillment (booking, payments, password resets, and call routing). Use Amazon Q in Connect for knowledge retrieval, open-ended questions, and scenarios where the customer's request does not fit a predefined intent flow.
In practice, the two services complement each other. Lex handles the structured front door: identify the customer, capture account details, and route to the correct workflow. When a customer asks something that falls outside configured intents, Lex can seamlessly hand off to Amazon Q in Connect, which searches your knowledge base and provides the answer. Throughout this process, full conversation context transfers between services, so the customer never repeats information. When the self-service experience cannot fully resolve an issue, both Lex and Amazon Q in Connect support seamless handoff to a human agent with the complete interaction history preserved.
Understanding Natural Customer Conversations
Natural Language Understanding
Think about how customers naturally ask questions. Someone might ask, "How do I return my new laptop?" and another might say, "I need to send back an electronic item that I bought." Though phrased differently, both customers need the same information. Amazon Q in Connect understands these variations in human communication, and interprets the true intent behind customer questions rather than simply matching keywords.
Amazon Q in Connect processes these questions using advanced natural language understanding and generates responses that feel natural and helpful. Instead of receiving a rigid, prewritten script, customers get conversational responses that directly address their specific concerns. For example, when asked about electronics returns, Amazon Q in Connect might respond like this:
"I can help you with returning your electronic item. Our electronics return window is 30 days from purchase, and you will need the original packaging and receipt. Would you like me to guide you through the return process or explain our testing requirements for electronic returns?"
Example Customer Interaction
Example customer interaction with Amazon Q in Connect
Review the following diagram to see how a customer might interact with Amazon Q in Connect when needing to change a previously booked flight.
Intent recognition
Amazon Q in Connect understands variations in how customers phrase their needs.
Natural responses
Responses feel conversational rather than scripted.
Conversational memory
The system maintains context throughout the interaction.
Follow up capability
The system can handle follow-up questions and clarifications.
Understanding how customers naturally communicate is fundamental to creating effective self-service experiences. Amazon Q in Connect excels at bridging the gap between human expression and automated assistance. By interpreting intent rather than relying on exact keyword matches, Amazon Q in Connect delivers truly conversational interactions that feel intuitive and helpful to your customers. The result is a self-service experience that customers find genuinely useful and engaging. It understands not just what they are asking, but what they really need to accomplish. This leads to faster resolutions and higher satisfaction rates across all customer touchpoints.
Optimizing Your Knowledge Foundation
Knowledge Base Optimization
For these conversations to work effectively, Amazon Q in Connect draws information from your organization's knowledge base. The same knowledge base of carefully curated information that helps your agents can now drive customer self-service interactions. This might include your internal documentation, policy details, product specifications, and lessons learned from successful customer interactions.
The key is to ensure that this information is well-organized, up to date, and relevant for the type of interactions that you receive. When you update a policy or add new product information, both your agents and the self-service system immediately have access to the latest details.
Confirm up to date content
Ensure that your knowledge base contains accurate, up-to-date information.
Structure your content
Well-structured content helps guide Amazon Q in Connect to provide your agents with more refined responses.
Consider breaking up long sections to shorter focused paragraphs centered on one idea.
Break down complex processes into clear steps by using bulleted or numbered lists when appropriate.
Simplify language and be specific
Be sure to write clearly and simply without complex vocabulary or sentence structures and steer away from making broad, vague statements.
Plain, straightforward writing is easier for Amazon Q in Connect and your agents to comprehend
Supplement visuals with descriptions
Amazon Q in Connect can analyze and extract meaningful information from unstructured or semistructured documents with advanced parsing enabled by default.
When using PDF files, it involves breaking down the document into its constituent parts. Such as text, tables, images, and metadata, and identifying the relationships between these elements.
All other files types you should provide descriptive text to supplement visuals like images, charts, and graphs that are referenced in your documentation.
It's important to add captions and summaries directly to your documents to help Amazon Q in Connect better interpret this content.
Define terms and reference unknown terminology
Include definitions for acronyms, abbreviations, and uncommon terms unique to your business on first use in each article.
Connect these terms to the phrases that your customers use to aid the understanding of Amazon Q in Connect.
Use examples
Always illustrate concepts with specific use cases by using examples.
By providing these details liberally, you ground how well Amazon Q in Connect understands your business, as well as the agent's expected responses and actions.
Stay agile, stay informed, and let your knowledge base be a dynamic evolving repository that your agents and generative AI-powered capabilities like Amazon Q in Connect can use
Responsible AI and Continuous Improvement
Responsible AI in self-service:
Use AI guardrails: Configure guardrails to prevent inappropriate responses and ensure accuracy.
Enable citations: Configure your AI assistant to cite sources when providing information.
Personalize responses: Pass customer context to Amazon Q in Connect to provide personalized responses.
Add conversation bridges: Create smooth transitions when the system needs to hand off to a human agent.
Monitor and refine: Regularly review transcripts of Q&A interactions to identify improvement opportunities.
Continuous improvement
Success with Q&A capabilities comes from continuous refinement. By reviewing conversation transcripts, you can identify areas where the system excels and where it needs improvement. Look for patterns in customer questions that might indicate gaps in your knowledge base or opportunities to make responses more helpful. Remember, the goal is to create conversations that feel natural and helpful while providing accurate information. As you implement and refine these capabilities, focus on how well the system understands customer intent and how naturally it can engage in back-and-forth dialogue. This attention to conversational quality helps ensure that your customers receive the kind of experience that builds trust and satisfaction with your self-service options.
Actions for Amazon Q in Connect Self Service
What Are Actions
What are Actions for Amazon Q in Connect Self service
Actions in Amazon Q in Connect allow the AI assistant to do more than just answer questions. They facilitate self-service experiences for customers. Actions in Amazon Q in Connect are defined through tools. Actions can intelligently gather information and orchestrate the customer journey through your contact flow.
Think of actions as intelligent instructions that help guide customers through your self-service experiences. When a customer expresses an intent like "I need to change my appointment," Amazon Q in Connect can collect and validate the necessary information. It then routes to the appropriate self-service flow or Amazon Lex bot that handles rescheduling.
The power lies in how Amazon Q in Connect uses natural conversation to do the following:
Understand customer intent more accurately.
Gather required information naturally.
Make smart routing decisions.
Hand off seamlessly to existing self-service workflows.
Tool Types in Amazon Q in Connect
Actions are defined through tools. Amazon Q in Connect comes with the default system tools that you can use immediately. For more specialized tasks, Amazon Q in Connect default system tools can also be extended to create custom tools.
Default System Tools
Amazon Q in Connect comes with the following built-in tools:
QUESTION: Provides answers and gathers relevant information when no other tool can directly address the query
ESCALATION: Automatically transfers to an agent when customers request human assistance
CONVERSATION: Engages in basic dialog when there's no specific customer intent
COMPLETE: Concludes the interaction when customer needs are met
FOLLOW_UP_QUESTION: Enables more interactive and information-gathering conversations with customers. For more information about using this tool, see the Follow Up Question Tool tab.
To learn more about default system tools for Amazon Q in Connect, see the Use Generative AI-Powered Self-Service with Amazon Q in Connect section in the Amazon Connect Administrator Guide.
Follow Up Question Tool
The FOLLOW_UP_QUESTION tool enhances Amazon Q in Connect self-service capabilities by enabling more interactive and information-gathering conversations with customers. This tool works alongside the default and custom tools. It helps collect necessary information before determining which action to take.
The FOLLOW_UP_QUESTION tool complements your defined tools by enabling Amazon Q in Connect to gather necessary information before deciding which action to take. It's particularly useful for the following:
Intent disambiguation — When the customer's intent is unclear, use this tool to ask clarifying questions before selecting the appropriate action.
Information gathering — Collect required details for completing a task or answering a question.
To learn more about custom actions for Amazon Q in Connect, visit the FOLLOW_UP_QUESTION Tool section in the Amazon Connect Administrator Guide.
Custom Tools
Custom tools allow you to extend the Amazon Q in Connect capabilities by creating tools that handle specific tasks and can hand off to Amazon Lex bots relevant to your business.
You can customize a default system tool to be specific for checking order status, to ensure that it is configured to know what details you might need to accomplish this action.
For example, a customer wants to check their order status. You might want to collect specific information before finding their order status or passing to a deterministic bot that will continue to handle that self-service interaction. Amazon Q in Connect might give this response:
"I can help you check your order status. Could you please provide your order number or the email address used for the purchase?"
Typical Actions that custom tools are created for are as follows:
Appointment scheduling and rescheduling: Allow customers to book, view, or change appointments.
Order management: Enable checking order status, modifying orders, or initiating returns.
Account updates: Permit customers to update contact information or preferences.
Payment processing: Facilitate making payments or checking payment history.
To learn more about custom actions for Amazon Q in Connect, see the Custom Actions for Self-Service section in the Amazon Connect Administrator Guide.
Implementation Strategies
Implementation Strategy Best Practices
Successful implementation of Amazon Q in Connect requires thoughtful planning and a strategic approach. Consider the experience of a telecommunications provider who started their journey with a focused implementation strategy. They began by identifying their highest-volume customer inquiries: technical support for internet connectivity issues. By starting with this specific use case, they could refine their approach before expanding to more complex scenarios.
Complex Scenarios
When implementing Amazon Q in Connect for more sophisticated use cases, organizations should follow several key strategies to ensure success:
Start small and expand: Begin with low-complexity interactions or with deterministic self-service experiences that have higher volume to allow expansion.
Focus on customer journey integration: Ensure that Amazon Q in Connect is part of a cohesive customer experience.
Prioritize knowledge management: Invest time in organizing and optimizing your knowledge base.
Personalize when possible: Use customer attributes in flows with Amazon Q in Connect to personalize generated responses.
Monitor, learn, and adapt: Use analytics to identify areas for improvement.
By applying these strategies systematically, you can gradually expand the capabilities of Amazon Q in Connect while maintaining quality and reliability. The most successful implementations typically begin with specific use cases and grow organically based on performance data and customer feedback.
Key Business Impact Metrics
To truly understand the value that Amazon Q in Connect brings to your organization, you need to measure its performance. This measurement should use targeted metrics that align with your business objectives. Organizations that successfully implement these AI-powered solutions typically track the following indicators:
Containment rate: This is the percentage of inquiries fully resolved through self-service.
Cost per contact: Compare the cost of self-service interactions against agent-handled contacts.
Customer satisfaction: Measure satisfaction specifically for self-service interactions.
Agent efficiency: Track how agent handling times and job satisfaction improve.
Resolution speed: Compare how quickly issues are resolved through self-service instead of traditional methods.
A methodical approach to these metrics provides a clear picture of your return on investment. By establishing these measurements early in your implementation process, you can quantify success and identify opportunities for continuous optimization.
Agentic Self Service with Amazon Q in Connect
From Scripted Flows to Autonomous Agents
The self-service capabilities you have learned about so far, Amazon Polly for voice, Amazon Lex for intent recognition, and Amazon Q in Connect for generative AI responses, have evolved into something more powerful. Amazon Q in Connect now supports agentic self-service: AI agents that can reason through complex problems, remember context across a conversation, and take action to resolve customer requests end-to-end.
Previously, self-service systems followed a linear path: recognize intent, gather information, provide a response or transfer to an agent. Agentic self-service breaks this pattern. The AI agent can dynamically decide what steps to take, in what order, based on the customer's actual situation.
How agentic self-service works
Powered by Amazon Nova Sonic, AI agents interact with customers through natural voice conversations. The agent performs the following tasks:
Listens and understands the customer's request (including complex, multi-part problems)
Reasons about what information it needs and what actions to take
Uses tools via Model Context Protocol (MCP) to retrieve data and complete transactions
Responds naturally, adapting its approach based on what it discovers
This happens in real time, with the customer experiencing a fluid conversation rather than a series of prompts.
What this means for customers
Customers can describe their problem in their own words, without navigating menus or simplifying their request. The AI agent handles the complexity:
"I got charged twice for my order last week, and I also need to change my delivery address for the replacement." A single AI agent interaction resolves both issues.
"My internet has been slow since the storm. I've already rebooted the router." The agent skips troubleshooting steps the customer has already tried.
The role of Nova Sonic
Amazon Nova Sonic is the voice-native LLM that powers these conversations. Because it processes speech directly (rather than converting speech to text and back), interactions feel natural and responsive. Customers hear an expressive, conversational voice, not a robotic one.
Enhanced Capabilities
AI agents need access to information and systems to resolve customer requests. Several enhancements make this more powerful and flexible.
Multiple knowledge bases
You can now bring your own Amazon Bedrock Knowledge Bases and connect multiple knowledge bases to a single AI agent. This means an agent can draw from:
Product documentation
Policy and procedure guides
FAQ databases
Troubleshooting knowledge
Each knowledge base can be maintained independently, and the agent determines which source is most relevant for each query.
Model Context Protocol MCP tools
MCP provides a standardized way for AI agents to use tools, such as retrieving information, calling APIs, and completing actions in external systems. When a customer asks to reschedule a delivery, the agent uses MCP to connect to the scheduling system, check availability, and make the change.
This standardized protocol means new tools and integrations can be added without custom development for each one.
Stream messages
AI agent responses are now shown to customers as they are generated, rather than waiting for the complete response before displaying anything. This reduces perceived wait times, particularly for complex queries that require the agent to reason through multiple steps.
For voice interactions, this means the agent begins speaking as soon as it has enough context to respond, creating a more natural conversational cadence.
Automated self service evaluations
The system can automatically evaluate the quality of self-service interactions. It checks whether the AI agent resolved the request correctly, maintained appropriate tone, and followed organizational guidelines. This provides continuous quality monitoring without requiring supervisors to manually review every interaction.
AI Agent Performance Metrics
The Eight Metrics
The following eight metrics provide a complete picture of AI agent performance. They are grouped into three categories: outcome metrics, quality metrics, and customer feedback metrics.
Outcome metrics
These measure whether the AI agent accomplished what it set out to do:
Goal success rate — Did the AI agent successfully resolve the customer's request? This is the primary measure of agent effectiveness. A high goal success rate means customers are getting their problems solved without needing to escalate to a human agent.
Hand-off rate — How often does the AI agent transfer the interaction to a human? Some hand-offs are appropriate (complex situations requiring human judgment), but a high rate may indicate the agent needs tuning.
Quality metrics
Customer feedback metrics
These measure how well the AI agent performs its work:
Faithfulness score — Are the agent's responses grounded in factual, source-backed information? This metric detects contextual hallucinations — cases where the agent generates plausible but unsupported answers. A low faithfulness score is a serious quality issue.
Tool selection accuracy — Did the agent choose the right tools to accomplish the task? When an agent has access to multiple tools via MCP (order lookup, scheduling, refund processing), this metric evaluates whether it selected the appropriate one for each situation.
Conversation turns — How many exchanges does it take to resolve an issue? Fewer turns generally indicate a more efficient agent, though complex issues naturally require more.
Tool use accuracy — Did the agent use the selected tools correctly with proper parameters? This measures execution quality after the correct tool is chosen.
Quality metrics
These measure how well the AI agent performs its work:
Customer feedback metrics
These capture the customer's direct experience:
Invocation success rate — The percentage of AI agent invocations that complete without error.
Completeness score — Did the AI agent fully address all aspects of the customer's request?
Faithfulness Scoring in Depth
Faithfulness scoring deserves closer attention because it addresses one of the most significant risks of autonomous AI agents: hallucination.
When a customer asks, "What's your return policy for electronics?" the AI agent should respond with information from the organization's actual return policy documentation — not generate a plausible-sounding policy from its training data.
Faithfulness scoring evaluates each response against the source material available to the agent. It answers the question: "Is this response supported by the agent's knowledge bases and retrieved data?"
How it works
The system compares the AI agent's response against:
The documents retrieved from connected knowledge bases
Data returned by tool calls (order details, account information)
The conversation context established by the customer
Each response receives a score indicating how well it is grounded in these sources. Responses that contain claims not present in any source material score lower.
What to do with low scores
Low faithfulness scores indicate the agent is generating unsupported information. Actions to take:
Review the specific interactions flagged with low scores
Check whether the knowledge base contains the information the agent should have used
Identify gaps in documentation that force the agent to extrapolate
Adjust agent guardrails to reduce hallucination in specific topic areas
Setting thresholds
Organizations can define minimum acceptable faithfulness scores. Interactions falling below the threshold can be:
Flagged for human review
Automatically escalated to a live agent
Logged for agent improvement training
Accessing the Metrics
These metrics are accessible through multiple channels depending on your role and needs.
Access method
Best for
Capabilities
AI Agent Performance dashboard
Supervisors, day-to-day monitoring
Visual trends, real-time scores, alerts,filtering by agent name, escalation status, or intent
Contact details page
Investigators, issue-level analysis
Step-by-step traces showing each reasoning step, model used, latency, tool parameter inputs, and model output
GetMetricDataV2 API
Developers, custom reporting
Programmatic access, integration with external tools
The dashboard provides the quickest path to understanding agent performance. You can filter by agent name, escalation status, or intent to narrow down specific performance patterns. For deeper investigation of individual interactions, the contact details page allows you to drill into step-by-step traces that show how the AI agent reasoned, which tools it selected, and the latency of each step during self-service voice interactions. The API enables automation — for example, triggering alerts when faithfulness scores drop below a threshold. The data lake supports deeper analysis for teams optimizing agent behavior over time.
RECALL & APPLY
Practice what you learned
Try answering in your own words, then compare with the suggested answer. These activities are for self-study and are not automatically graded.
01What makes a request suitable for an initial self service deployment?
Topic: AI self service in the contact center
Reveal suggested answer
A clear customer need, dependable information or action, manageable exceptions, and a measurable definition of success.
02What is the difference between recognizing words and completing a request?
Topic: IVR and conversational self service
Reveal suggested answer
Recognition produces an interpretation. Completion also requires data collection, validation, fulfillment, and handling of the result.
03Which capability answers an open ended policy question?
Topic: Polly, Lex, and Amazon Q roles
Reveal suggested answer
Knowledge retrieval and generation can support the answer. The workflow must use the approved policy content and handle uncertainty appropriately.
04Why would an account balance prompt use text to speech?
Topic: Amazon Polly prompts
Reveal suggested answer
The value changes by customer and time. Dynamic text can provide the current value without recording every possible message.
05Why can a long written paragraph make a poor voice prompt?
Topic: Designing prompts for listening
Reveal suggested answer
Listeners cannot scan the text or easily revisit an earlier point. Spoken prompts must manage sequence, memory, and interruptions.
06What would you test when a product code sounds wrong?
Topic: Voice selection and SSML
Reveal suggested answer
Test the text and supported SSML interpretation, including whether characters should be spoken individually and whether the selected voice supports the chosen tags.
07In an appointment bot, what is the difference between the intent and a slot?
Topic: Amazon Lex core concepts
Reveal suggested answer
The intent is booking an appointment. Slots hold details such as date, time, location, or appointment type.
08Why use an alias instead of hardcoding a bot version?
Topic: Bot structure and deployment
Reveal suggested answer
An alias lets the integration select a published version through a stable reference. Version changes can then follow a controlled deployment process.
09Does collecting transfer details move money by itself?
Topic: Intents, slots, and confirmation
Reveal suggested answer
No. The conversation captures and confirms data. An authorized backend fulfillment integration must perform and confirm any real transfer.
10Why should a voice bot be tested through a real call path?
Topic: Bot testing and telephony input
Reveal suggested answer
Telephony audio, noise, timing, and caller behavior can expose issues that do not appear in typed tests or clean microphone recordings.
11A customer changes from a balance inquiry to a transfer. What should the design preserve?
Topic: Conversational interface principles
Reveal suggested answer
Preserve useful context, recognize the new goal, collect the required details, and confirm the consequential action.
12Should the same long answer be used unchanged in voice and chat?
Topic: Voice and text design
Reveal suggested answer
Usually not. Voice needs concise sequential delivery, while text can use short paragraphs, lists, and links where appropriate.
13What should be checked before adding the bot to a flow?
Topic: Building a Lex bot in Amazon Connect
Reveal suggested answer
Confirm that the language builds successfully, intent names are consistent, slots capture the intended data, and fallback behavior works.
14What happens if the flow’s intent name does not match the bot?
Topic: Connecting the bot to a flow
Reveal suggested answer
The intended routing can fail. Use the exact configured intent name and test the corresponding branch before deployment.
15What evidence proves that the conversational test succeeded?
Topic: Testing the complete experience
Reveal suggested answer
Correct intent recognition, captured values, confirmation behavior, and expected flow routing. A successful conversation does not prove a backend transaction occurred.
16Why should generated utterances still be reviewed?
Topic: Generative AI for Lex development
Reveal suggested answer
They may be irrelevant, ambiguous, or inconsistent with the intended task. Evaluate them against real customer language and neighboring intents.
17How does assisted slot resolution differ from building a new intent?
Topic: Generative Lex capabilities
Reveal suggested answer
It helps interpret a value for an existing slot. It does not by itself define a new business goal or fulfillment workflow.
18Which path fits retrieving a policy answer versus coordinating several tools?
Topic: Knowledge and agent integrations
Reveal suggested answer
Knowledge question answering fits retrieving and presenting approved information. Agent orchestration fits a task that requires multiple operations and context gathering.
19What does a text based model receive in a speech to text pipeline?
Topic: NLU and voice architecture choices
Reveal suggested answer
It receives the transcribed words and supplied context. It does not directly receive all acoustic information unless the architecture provides it.
20What would you compare when evaluating a speech integration?
Topic: Speech providers and language support
Reveal suggested answer
Use representative customer audio, relevant languages, latency, error recovery, operational requirements, and the supported integration path.
21What should transfer with the customer when self service cannot resolve the issue?
Topic: Amazon Q for customer self service
Reveal suggested answer
The relevant issue, information already gathered, actions attempted, and current state, according to the implemented handoff design.
22Why might one customer interaction use both services?
Topic: Lex and Amazon Q in a shared journey
Reveal suggested answer
A structured workflow can collect required details, while a knowledge assistant answers an unexpected question before the customer returns to the task.
23A customer says “send it back.” What should the assistant establish?
Topic: Natural customer conversations
Reveal suggested answer
Determine the item and intended action from context or a clarifying question. Do not assume the relevant order or policy.
24Why might an accurate document still retrieve poorly?
Topic: The knowledge foundation
Reveal suggested answer
Its structure, terminology, or missing explanatory text may make the relevant information hard to find or interpret.
25What should you do when the assistant repeatedly fails on the same question?
Topic: Responsible self service improvement
Reveal suggested answer
Review the affected conversations, locate the knowledge or configuration gap, make a targeted change, and retest the behavior.
26Why is asking a follow up question sometimes the best action?
Topic: Actions and system tools
Reveal suggested answer
The assistant may lack the information needed to select or execute a tool correctly. Clarification can prevent an incorrect operation.
27What information might an order status tool need?
Topic: Follow up questions and custom tools
Reveal suggested answer
An appropriate order or customer identifier, plus the access and validation required by the organization before returning customer specific information.
28What should a first deployment prove before it expands?
Topic: Implementation strategy and outcomes
Reveal suggested answer
It should reliably address the intended requests, handle exceptions, and meet agreed quality and operational goals.
29A customer reports a duplicate charge and an address change. What makes this more complex than a single FAQ?
Topic: Agentic self service
Reveal suggested answer
It involves separate goals, customer specific data, consequential actions, and a need to track which parts have actually completed.
30Does starting a response quickly prove the task is complete?
Topic: Knowledge, tools, and streamed responses
Reveal suggested answer
No. Response timing and successful task completion are separate. Confirm the tool result and whether the full request was addressed.
31What is the difference between selecting the right tool and using it correctly?
Topic: AI agent outcome and quality metrics
Reveal suggested answer
Selection chooses the appropriate capability. Correct use supplies valid parameters and handles its result properly.
32Can a faithful answer still be incomplete?
Topic: Reliability, completeness, and faithfulness
Reveal suggested answer
Yes. It can accurately reflect a source while failing to answer part of the customer’s request.
33What should a team avoid when a score falls below its threshold?
Topic: Investigating low faithfulness
Reveal suggested answer
Avoid treating the score alone as a complete diagnosis. Inspect the interaction, evidence, and configuration before choosing the correction.
34Choose one use case and describe its rollout plan.
Topic: Performance review and course application
Reveal suggested answer
A strong answer defines the customer need, data and knowledge, tools, human handoff, test cases, outcome metrics, and ongoing review owner.
Capture a key idea, an example, or a question for your instructor.
REFERENCE
A–Z topic index
Choose a topic to jump directly to its explanation in the eBook.