Most Cloud Practitioner candidates can name EC2, S3, RDS and Lambda in their sleep. Then the exam describes a company that wants to “automatically extract the total from scanned invoices” or “convert customer service calls into searchable text”, and four unfamiliar service names appear. These questions come from Task Statement 3.7 of the CLF-C02 exam guide — identify AWS artificial intelligence and machine learning (AI/ML) services and analytics services — and they are among the cheapest marks on the paper. Each service needs one sentence of knowledge. The catch is that the sentences are easy to confuse with each other, and the exam is built around exactly those confusions.
This guide covers the AI/ML half of that task statement, one service at a time, with the recognition cue the exam uses, the distractor it is usually paired with, and the current state of each service — because the list has changed since the exam guide was written. The analytics half (Athena, Glue, Kinesis, Redshift, EMR and QuickSight) has its own analytics services guide, and the other small categories in the same domain are in the business apps, end-user computing, frontend and IoT guide.
Where This Sits on the CLF-C02
Domain 3, Cloud Technology and Services, is the largest domain at 34% of scored content. Task Statement 3.7 asks for two skills:
- Understanding the different AI/ML services and the tasks that they accomplish (the exam guide’s own examples are Amazon SageMaker, Amazon Lex and Amazon Kendra).
- Identifying the services for data analytics (Athena, Kinesis, Glue, QuickSight).
Depth is recognition only. You will not be asked how to train a model, what an epoch is, or how a transformer works — that is the territory of the AI Practitioner exam, compared in AIF-C01 vs CLF-C02. A typical CLF-C02 question is one or two sentences describing a business need followed by four service names, or a service name followed by four descriptions. For how Domain 3 is weighted against the rest of the exam, see the CLF-C02 domains breakdown.
The in-scope list has moved
Here is a detail that catches people who study from an old PDF. The exam guide (version 1.0) lists nine Machine Learning services: Comprehend, Kendra, Lex, Polly, Rekognition, SageMaker, Textract, Transcribe and Translate. AWS now maintains the in-scope list as a web page, and the current page differs:
| Service | Exam guide PDF (v1.0) | Current in-scope web page |
|---|---|---|
| Amazon Comprehend | Listed | Listed |
| Amazon Kendra | Listed | Not listed |
| Amazon Lex | Listed | Listed |
| Amazon Polly | Listed | Listed |
| Amazon Q | Not listed | Listed |
| Amazon Rekognition | Listed | Listed |
| Amazon SageMaker | Listed as “Amazon SageMaker” | Listed as “Amazon SageMaker AI” |
| Amazon Textract | Listed | Listed |
| Amazon Transcribe | Listed | Listed |
| Amazon Translate | Listed | Listed |
Both lists say they are non-exhaustive and subject to change, so the practical rule is: know all ten. Kendra is still a plausible distractor and still appears in the task statement’s own example. Amazon Q is the newest addition and the one most candidates have never read about. Amazon Bedrock is on neither list, but it is the service AWS talks about most, so a one-line definition is cheap insurance.
The Three-Layer Mental Model
Before the individual services, fix the structure that every AWS AI/ML question is drawn from. AWS organises its offerings in three layers, and the exam’s distractors are almost always from a different layer than the correct answer:
| Layer | What it is | Who uses it | Services on the list |
|---|---|---|---|
| Pre-trained AI services | Ready-made models behind an API. You send data in, you get a prediction out. No training, no ML expertise. | Application developers with no data science background | Rekognition, Textract, Comprehend, Translate, Transcribe, Polly, Lex, Kendra |
| ML platform | Tools to build, train, tune, deploy and monitor your own models. | Data scientists and ML engineers | SageMaker AI |
| Generative AI assistants and foundation models | Large language models that generate text, code and answers, packaged either as an assistant or as a model API. | Everyone, from developers to business users | Amazon Q (and, conceptually, Amazon Bedrock) |
When a question says “without machine learning expertise”, “no model training” or “ready to use”, the answer is a pre-trained AI service. When it says “build and train a custom model”, “data scientists” or “Jupyter notebooks”, the answer is SageMaker AI. When it says “generative AI assistant”, “answer questions about company documents” or “help developers write code”, the answer is Amazon Q.
Vision: Rekognition and Textract
Amazon Rekognition
Amazon Rekognition analyses images and videos. Point it at a photo or a video stream and it returns what it sees: objects and scenes (labels), faces (detection, comparison and search against a collection), celebrities, text that appears inside an image such as a street sign, personal protective equipment, and unsafe or inappropriate content for moderation. Custom Labels lets a business train it to recognise its own objects — a specific product on a shelf, a logo — with a small set of example images and no ML expertise.
Recognition cue: images, video, faces, objects, content moderation.
Amazon Textract
Amazon Textract extracts text and structured data from documents — scanned forms, PDFs, invoices, receipts, identity documents. It goes beyond classic optical character recognition (OCR): it understands forms as key-value pairs (“Invoice Number: 4471”), tables as rows and columns, signatures, and it can answer natural-language queries such as “what is the total amount due?” without you building templates.
Recognition cue: documents, forms, tables, invoices, receipts, OCR.
The pair the exam loves
Rekognition and Textract both “read images”, so the exam pairs them constantly. The split is simple:
| The question mentions… | Answer |
|---|---|
| A photo, a video, a face, a product on a shelf, inappropriate content | Rekognition |
| A scanned document, a form, a table, an invoice, a PDF, “extract the fields” | Textract |
| Text visible in a street scene or on a sign in a photo | Rekognition (text-in-image detection) |
| Text that is the document | Textract |
Language: Comprehend, Translate and Kendra
Amazon Comprehend
Amazon Comprehend is natural language processing (NLP): it reads text and extracts meaning. Given a block of text it returns the language, the entities (people, places, organisations, dates), key phrases, sentiment (positive, negative, neutral, mixed), and it can detect and redact personally identifiable information (PII). Custom classification and custom entity recognition let you train it on your own categories without ML expertise. Comprehend Medical does the same for clinical text.
Recognition cue: sentiment, entities, key phrases, PII in text, “understand what customers are saying”.
Amazon Translate
Amazon Translate is neural machine translation between languages, in real time or in batch, with custom terminology so that product names and brand terms are translated the way you want. It is the only service on the list whose job is to produce the same content in another language.
Recognition cue: translate, localise, multiple languages, “serve customers in their own language”.
Amazon Kendra
Amazon Kendra is intelligent enterprise search. It indexes documents from many sources — SharePoint, S3, databases, wikis — and answers natural-language questions (“how many days of leave do I get?”) with a specific passage rather than a list of keyword matches, honouring the document permissions of the user who asks.
Recognition cue: search across company documents, natural-language questions, intranet search.
Current state: AWS has closed Amazon Kendra to new customers, effective 30 July 2026, and points new workloads to Amazon Bedrock Knowledge Bases. Existing customers can keep using it. For the exam, Kendra is still the name attached to “enterprise search” and still appears as an example in the exam guide, so recognise it — but do not be surprised if a newer question describes the same need and expects a generative AI answer such as Amazon Q.
Comprehend vs Kendra vs Textract
Three services touch “text” and the exam uses all three as distractors for each other:
| Need | Service |
|---|---|
| Get the text out of a scanned document | Textract |
| Understand what a piece of text means — sentiment, entities, PII | Comprehend |
| Find the right document or answer across a corpus | Kendra (or Amazon Q for a generative, conversational answer) |
A common pipeline question chains them: Textract extracts the text from scanned contracts, Comprehend finds the entities and redacts PII, Kendra or a knowledge base makes the result searchable. If a question describes that pipeline, every stage maps to one service.
Speech and Conversation: Transcribe, Polly and Lex
Amazon Transcribe
Amazon Transcribe converts speech to text. It handles recorded audio files and live streams, separates speakers, applies custom vocabularies for industry jargon, redacts PII from the transcript, and has a Call Analytics mode for contact centers and a medical variant for clinical dictation.
Recognition cue: transcribe, captions, subtitles, “convert recorded calls into text”, speech-to-text.
Amazon Polly
Amazon Polly converts text to speech. It turns written text into lifelike spoken audio in dozens of voices and languages, with the Speech Synthesis Markup Language (SSML) for control over pronunciation, pauses and emphasis. Typical uses are voice for applications, audio versions of articles, accessibility, and the spoken side of IVR systems.
Recognition cue: text-to-speech, lifelike voice, “read the article aloud”, audio output.
Amazon Lex
Amazon Lex builds conversational interfaces — chatbots and voice bots — using the same technology that powers Alexa. It combines automatic speech recognition (so it can take voice input) with natural language understanding (so it can work out the intent behind “I’d like to change my flight” and collect the slots such as the date and destination). Business logic is typically fulfilled by AWS Lambda, and Lex is the natural bot engine inside Amazon Connect contact centers.
Recognition cue: chatbot, voice bot, virtual agent, intents, “let customers book by chatting”.
The speech triangle
Transcribe, Polly and Lex are the most heavily tested trio in this category because the descriptions sound alike. Anchor on direction and purpose:
| Direction | Service |
|---|---|
| Speech in, text out | Transcribe |
| Text in, speech out | Polly |
| Two-way conversation with intent handling | Lex |
| Text in one language, text in another | Translate |
A complete customer-service flow uses all of them: Transcribe turns the call into text, Comprehend scores the sentiment, Translate renders it for an overseas agent, Polly speaks the reply, and a Lex bot handles the simple cases before a human is involved. Lex vs Amazon Connect is another favourite pairing: Connect is the contact center (phone numbers, queues, agents); Lex is the bot that can answer inside it.
The Platform: Amazon SageMaker AI
Amazon SageMaker AI is the fully managed service for building, training and deploying your own machine learning models. It covers the whole lifecycle so that data scientists do not have to manage servers: hosted notebooks and an integrated development environment (SageMaker Studio), data labelling, built-in and custom training algorithms, managed training jobs on CPU or GPU fleets, automatic model tuning, and deployment to managed inference endpoints — real-time, serverless, asynchronous or batch. JumpStart offers a catalogue of pre-trained and foundation models to start from, and Canvas provides a no-code interface for business analysts to build models by pointing and clicking.
Two naming facts matter for 2026 exams:
- At re:Invent 2024, AWS made “Amazon SageMaker” the umbrella name for a broader data and AI platform (the next generation of SageMaker, with Unified Studio and a lakehouse). The classic ML service — the one the Cloud Practitioner exam means — was renamed Amazon SageMaker AI. Questions may use either name; they refer to the same capability.
- Several SageMaker AI sub-features entered maintenance for new customers on 30 July 2026, including Ground Truth (labelling), Clarify (bias detection) and Model Monitor (drift). The exam does not test these sub-features; it tests the one-line definition of SageMaker AI itself.
Recognition cue: build, train, deploy custom models; data scientists; notebooks; “we have our own training data and need a custom model.”
The platform-versus-service distinction is the most common Domain 3 AI question. If the scenario can be solved by a pre-trained API — “detect faces”, “transcribe calls”, “translate the website” — the answer is not SageMaker AI, even though SageMaker AI could technically do it. SageMaker AI is the answer when the organisation needs a model of its own for a problem no pre-trained service solves, such as predicting equipment failure from its own sensor history.
Generative AI: Amazon Q (and Bedrock)
Amazon Q is AWS’s family of generative AI assistants. The members you should recognise:
| Member | What it does |
|---|---|
| Amazon Q Developer | An assistant for building on AWS: generates and explains code in the IDE and command line, answers AWS questions in the Management Console, helps troubleshoot and modernise applications |
| Amazon Q Business | A workplace assistant that answers questions, summarises and generates content from a company’s own documents and connected systems, respecting user permissions |
| Amazon Q in QuickSight | Natural-language questions and narrative generation inside the BI service |
| Amazon Q in Connect | Real-time suggested responses and actions for contact center agents |
Recognition cue: generative AI assistant, “help developers write code”, “answer employee questions from internal documents”, chat interface.
Current state: Amazon Q Business is no longer available to new customers; AWS positions Amazon Quick (the Quick Suite, launched in late 2025) as its successor for workplace assistants, and existing Q Business customers can continue or bring their index into Quick. Amazon Q Developer and Amazon Q in Connect continue. Interestingly, the current in-scope list already spells the BI service “Amazon Quick Sight”. For exam purposes, “Amazon Q” remains the name on the in-scope list, and the business-assistant scenario is still answered with Amazon Q.
Amazon Bedrock is not on the in-scope list, but it is the service AWS describes everywhere, and a newer question could use it as the correct answer or a distractor. One line is enough: Bedrock is a fully managed service that gives API access to foundation models from Amazon and third-party model providers, so that developers can build generative AI applications — chat, summarisation, retrieval-augmented generation with Bedrock Knowledge Bases — without training or hosting the models themselves. The distinction from SageMaker AI: Bedrock is use a foundation model through an API; SageMaker AI is build, train and host your own model.
Machine Learning Hiding Inside Other Services
The exam guide’s Appendix A lists “Machine learning” as a concept, not only a category, and several services outside Task 3.7 use ML as their defining feature. These appear in Domain 2 and Domain 4 questions, so recognise the ML angle:
| Service | What it does with ML | Domain |
|---|---|---|
| Amazon Macie | Discovers and classifies sensitive data (PII, credentials) in S3 | Security — see the security and compliance guide |
| Amazon GuardDuty | Threat detection from CloudTrail, VPC Flow Logs and DNS logs | Security |
| AWS Cost Anomaly Detection | Learns spending patterns and alerts on unusual spend | Billing — see the pricing and billing guide |
| AWS Compute Optimizer | Recommends right-sized instances from utilisation history | Management and governance |
A question that says “uses machine learning to identify sensitive data in S3 buckets” is asking for Macie, not Comprehend — the data is in S3 and the job is discovery, not NLP on a string you send. A question about “unusual spending” is Cost Anomaly Detection, not a SageMaker model.
Scenario-to-Service Cheat Sheet
| The question says… | Answer |
|---|---|
| Detect inappropriate images uploaded by users | Rekognition |
| Identify which employees appear in security footage | Rekognition (face search) |
| Pull the invoice number and total out of scanned invoices | Textract |
| Extract rows from a table in a scanned PDF | Textract |
| Determine whether product reviews are positive or negative | Comprehend |
| Find and mask credit-card numbers in support tickets | Comprehend (PII detection) |
| Offer the website in Spanish, German and Japanese | Translate |
| Let employees ask natural-language questions across SharePoint and S3 | Kendra (or Amazon Q Business for a generative answer) |
| Generate captions for recorded training videos | Transcribe |
| Convert call-center recordings into searchable text | Transcribe |
| Read news articles aloud in a natural voice | Polly |
| Build a chatbot that books appointments | Lex |
| A contact center needs a self-service voice bot | Lex (inside Amazon Connect) |
| Data scientists need to build and train a custom demand-forecasting model | SageMaker AI |
| A business analyst wants to build a model without writing code | SageMaker AI (Canvas) |
| Help developers generate code and answer AWS questions in the IDE | Amazon Q Developer |
| Give employees a chat assistant over internal documents | Amazon Q Business |
| Access foundation models from multiple providers through one API | Bedrock |
| Discover sensitive data in S3 using machine learning | Macie |
Common Mistakes to Avoid
- Choosing SageMaker AI for everything. It is the right answer only when the organisation must build its own model. For any described task that a pre-trained service covers, pick the pre-trained service.
- Swapping Transcribe and Polly. Transcribe listens, Polly speaks. Say “Transcribe transcribes” and “Polly talks” until it is reflexive.
- Choosing Rekognition for documents. Rekognition can read text in a photo of a sign; Textract reads forms, tables and PDFs. “Invoice”, “form” and “scanned” always mean Textract.
- Confusing Comprehend with Kendra. Comprehend analyses a piece of text you give it; Kendra searches a corpus to find the right piece.
- Confusing Lex with Connect. Connect is the contact center; Lex is the bot engine inside it.
- Treating Amazon Q as a single product. Q Developer helps build on AWS; Q Business answers from company data. The scenario tells you which.
- Studying only the PDF. The live in-scope list adds Amazon Q, renames SageMaker to SageMaker AI and drops Kendra. Know all ten names.
- Over-studying. You need the one-line purpose, the direction (in/out) and the distractor partner. Architecture, pricing and API details are beyond the exam.
How to Study This Category Efficiently
Write the ten services on one page in three columns — pre-trained AI services, platform, generative AI — and under each name write a single verb phrase: sees images, reads documents, understands text, translates, searches, hears speech, speaks, converses, builds models, assists. Then drill the pairs: Rekognition/Textract, Transcribe/Polly, Comprehend/Kendra, Lex/Connect, SageMaker AI/Bedrock. Twenty minutes of that beats two hours of documentation, because the exam only ever asks the pair question.
Then test recognition under time pressure. The CLF-C02 gives you 65 questions in 90 minutes, and these items should take fifteen seconds each. The Sailor.sh CLF-C02 mock exam bundle phrases these services the way the real exam does — mixed into full-length papers alongside compute, storage, security and billing questions, with explanations that name the distractor pair each item was built on. Use it to find the one or two pairs you still hesitate on, and reread those sections above. For the broader plan, the CLF-C02 exam guide and the study plan cover the whole syllabus, and the practice questions show the question style across all four domains.
If this category interests you more than the rest of the exam, the AWS Certified AI Practitioner goes one level deeper into the same services plus generative AI concepts, and is the natural next certification.
Frequently Asked Questions
How many AI/ML questions are on the CLF-C02?
AWS does not publish per-task counts. Domain 3 is 34% of scored content and contains eight task statements, so expect a handful of AI/ML items — typically two to four on a 65-question paper. They are quick marks if the service names are reflexive.
Do I need to know how machine learning works?
No. The Cloud Practitioner tests which service accomplishes which task. Concepts such as supervised learning, training data and model evaluation are on the AI Practitioner exam, not this one.
Is Amazon Bedrock on the CLF-C02?
It is not on the official in-scope list for CLF-C02, but the list is non-exhaustive, and Bedrock is central to how AWS describes generative AI. Know the one-line definition — managed API access to foundation models — and the distinction from SageMaker AI.
Is Amazon Kendra still relevant if AWS closed it to new customers?
Yes for the exam. It appears as an example in the task statement and was on the exam guide’s list, so it remains a plausible answer or distractor for “enterprise search”. In practice, new search and retrieval workloads go to Bedrock Knowledge Bases, and the workplace assistant role has moved to Amazon Quick.
What is the difference between Amazon SageMaker and Amazon SageMaker AI?
Since re:Invent 2024, “Amazon SageMaker” names a broader unified data and AI platform, and “Amazon SageMaker AI” is the classic service for building, training and deploying models. Exam questions may use either name for the ML platform; treat them as the same answer.
Which service turns text into speech, and which turns speech into text?
Polly turns text into speech. Transcribe turns speech into text. Lex does both as part of a two-way conversation, but its defining feature is understanding intent, not conversion.
Which service would I use to extract data from a scanned form?
Amazon Textract. It identifies key-value pairs and tables in forms and documents without templates, which is the capability Rekognition does not have.
Conclusion
Task Statement 3.7 rewards a small, well-organised vocabulary. Eight pre-trained AI services each do one job — Rekognition sees, Textract reads documents, Comprehend understands text, Translate translates, Kendra searches, Transcribe hears, Polly speaks, Lex converses. SageMaker AI is where you build your own model when none of those fit. Amazon Q is the generative AI assistant, with Developer and Business variants, and Bedrock is the foundation-model API underneath the generative AI story. Learn the direction of each service, the partner it is confused with, and the three-layer model that separates pre-trained services from the platform from generative AI, and every AI/ML question on the Cloud Practitioner exam becomes a fifteen-second decision.