Cheat Sheet: AWS AI Services

A fast pass over the managed AI services: one line each, and the pairs that look interchangeable but are not.

Services at a glance

Several names have come off the in-scope service list, and they are worth learning as a set. Amazon Kendra went into maintenance mode on 30 June 2026 and closed to new customers on 30 July. Existing indexes keep running. A new retrieval build resolves to Amazon Bedrock Knowledge Bases rather than a Kendra index. Amazon Forecast and Amazon Fraud Detector are closed to new customers and are not listed either. A forecasting or fraud-detection scenario resolves to Amazon SageMaker AI. The use case is still fair game even though the named service is not. Amazon Q Developer is not on the list, and it is the easiest of them to reach for by reflex. It does genuine developer work from the IDE, the CLI, and the console. Kiro is the developer tool that is listed, so a code scenario resolves there.

The GenAI build surface at a glance

The table above is the packaged half of the list, one fixed job per service. The generative half is where you build the job nobody has packaged yet, and the blocks stack in a predictable order.

The Bedrock surface at a glance

Bedrock is one service with several named features, and the names carry the distinction. Grounding answers in your own documents is the one that turns up most often.

One name has come off that list. The console-configured agent that Bedrock plans and runs is now Amazon Bedrock Agents Classic. It went into maintenance mode on 30 July 2026 and is closed to accounts with no prior usage, so a new agent build does not go there. An account already running one keeps it; everything else starts on Amazon Bedrock AgentCore, where the old action groups arrive as MCP tools behind AgentCore Gateway.

Responsible-AI documentation on these services

AWS publishes an AI Service Card for several of its AI services and models, Rekognition face matching, Amazon Titan Text, and the Amazon Nova models among them. Each card states intended use cases and limitations, responsible design considerations, and deployment guidance. It is AWS’s document about an AWS service. Your team’s equivalent for a model you built is an Amazon SageMaker Model Card, which records intended use, risk rating, training details, and evaluation results as a versioned document. Both serve transparency, and they differ in who wrote them and about what.

Two of the purpose-built services do responsible-AI work themselves. Rekognition moderates images and video, and Comprehend detects toxicity in text. Both cover the safety dimension for a team with no foundation model anywhere in the design.

Decision rules

• If the job is turning text into audio, that is Polly; if it is turning audio into text, that is Transcribe. The direction of the arrow decides which one.
• If the task is understanding existing text (sentiment, entities, PII), reach for Comprehend rather than asking a foundation model to do analysis a purpose-built service already does more cheaply.
• If you are building a defined-intent chatbot or voice interface, use Lex; an open-ended assistant over your own documents is Amazon Quick, not a Lex bot.
• If the input is an image or video and you need to know what is depicted, use Rekognition. If it is a document and you need its fields, tables, and key-value pairs, use Textract.
• If you need a retrieval layer feeding your own application, that is Amazon Bedrock Knowledge Bases; if you need a ready assistant with a UI, connectors, and permissions, that is Amazon Quick. Kendra held the first slot until it closed to new customers, so only an account that already runs an index still has it as an option.
• If the ask is “recommend items to this user”, that is Personalize, not a general-purpose model prompted with behavioural data.
• If a request is about writing code, that is Kiro; if it is about answering questions from company documents, that is Amazon Quick. Amazon Q Developer covers similar developer ground from the IDE, the CLI, and the console, which is why it gets named in the same breath as both, and it is not on the list.
• If a purpose-built AI service already solves the problem, prefer it over building the same capability on a foundation model; it is usually cheaper, faster, and more consistent for a narrow task.
• Capability descriptions cluster on six managed AI/ML services: Amazon SageMaker AI, Amazon Transcribe, Amazon Translate, Amazon Comprehend, Amazon Lex, and Amazon Polly. A scenario about what a managed service can do is nearly always one of those six.
• Sort Amazon Quick, Kiro, and Amazon Bedrock by who is asking. A business user with a question about company data gets Amazon Quick; a developer writing software gets Kiro; an application that needs a model behind its own code gets Bedrock.
• Bedrock or SageMaker AI comes down to who operates the endpoint. Bedrock is an API call to a model AWS runs and bills by the token; SageMaker AI is a model you deploy, size, and pay for while it is up.
• Amazon SageMaker JumpStart rather than Bedrock when the weights have to land on infrastructure you chose: your instance type, your VPC, your account.
• If answers have to come from the company’s own current documents, that is Amazon Bedrock Knowledge Bases and not fine-tuning. Retrieval keeps pace with documents that change; training does not.
• If the model has to take an action in another system, that is an agent, not a longer prompt.
• If a wording change has to be rolled back, that is prompt management in Amazon Bedrock, where prompts are versioned resources with input variables.
• For what AWS says the limits of a managed AI service are, that is the AI Service Card; for the limits of your own model, that is a SageMaker Model Card.

Traps

• Assuming Comprehend generates text. It analyses text that already exists; it does not write anything new.
• Confusing Rekognition and Textract because both process images. Rekognition answers “what is depicted”; Textract answers “what is the structure of this document” (fields, tables, key-value pairs).
• Treating Kendra as a chatbot. It is the retrieval layer other things are built on top of, not a conversational surface in itself.
• Picking Kendra for a new build. It went into maintenance mode on 30 June 2026 and closed to new customers a month later; existing indexes keep running, but a new retrieval build goes to a Bedrock knowledge base.
• Swapping the assistants in a sentence about “the AWS assistant”. Amazon Quick answers from company data with citations and no code written. Kiro is the listed developer tool, an agentic IDE for people writing software. Amazon Q Developer does comparable developer work from the IDE, the CLI, and the console, and is not on the list. A question about company documents is never answered by either developer tool.
• Reaching for a foundation model to do sentiment analysis or entity extraction when Comprehend already does it as a managed, purpose-built call.
• Picking Personalize for a one-off recommendation with no behavioural history to train on; it needs interaction data to be worth using over a simpler rule.
• Forgetting that Lex handles defined intents and slots; an open-ended, document-grounded conversation is a different architecture, not a Lex bot with more intents bolted on.
• Treating Strands Agents and Amazon Bedrock AgentCore as competing choices. Strands Agents is the framework the agent is written in; AgentCore is the managed runtime it can be deployed onto. Picking one does not rule out the other. Amazon Bedrock Agents Classic was a third shape again, an agent configured rather than written and run by Bedrock. It is closed to accounts with no prior usage, so a new build picks between the other two.
• Reading Bedrock Guardrails and a prompt instruction as the same control. A guardrail evaluates the input and the response against configured policies and blocks or masks what matches, whatever the prompt said. “Never discuss competitors” inside a prompt is wording the output may not follow.
• Confusing Amazon Bedrock Evaluations with SageMaker Clarify. Evaluations score what a model produced; Clarify, closed to new customers, measures bias and feature attribution.
• Answering a forecasting or fraud-detection scenario with Amazon Forecast or Amazon Fraud Detector. Both are closed to new customers and off the in-scope list, so that work goes to SageMaker AI.

Say it in one line

1. Polly turns text into speech; Transcribe turns speech into text. Pick by which way the conversion runs.
2. Comprehend analyses text that already exists: sentiment, entities, PII. It does not generate anything.
3. Lex builds defined-intent chatbots and voice bots from intents and slots; an open-ended assistant is a different build.
4. Rekognition reads images and video; Textract reads document structure, forms, and tables.
5. Kendra was the enterprise-search retrieval layer and is now maintenance-only and closed to new customers. Amazon Bedrock Knowledge Bases is where a new retrieval build goes. Amazon Quick is the ready assistant with connectors, permissions, and citations.
6. Personalize ranks and recommends from behavioural data.
7. Amazon Quick answers questions from company data with citations and no code; Kiro is the listed tool for people writing software. Amazon Q Developer does the same kind of developer work as Kiro and is not on the list, so a code scenario resolves to Kiro.
8. Prefer a purpose-built AI service over a foundation model when one already exists for the task.
9. Bedrock is a model AWS operates and bills by the token; SageMaker AI is a model you operate on endpoints you size and pay for; JumpStart is how a ready-made model gets onto them.
10. Knowledge Bases grounds answers in documents that change; fine-tuning does not keep up with them.
11. Guardrails evaluate input and response against policy and block or mask what matches; a prompt instruction only shapes the wording.
12. An AI Service Card is AWS’s account of an AWS service; a SageMaker Model Card is your team’s account of your model.