A large share of AWS Certified AI Practitioner (AIF-C01) questions never mention machine learning math at all. Instead they describe a business problem — “extract data from scanned invoices,” “detect the sentiment of customer reviews,” “build a chatbot,” “turn an article into audio” — and ask you to pick the right managed AWS AI service. These are pure recognition questions, and they’re free points if you know what each service does and, just as importantly, where two similar services differ.
This guide walks through the AWS pre-trained, API-driven AI services that show up on the AWS Certified AI Practitioner (AIF-C01) exam. These are the services you call with an API — no model training, no infrastructure — which AWS positions as the top layer of its “AI/ML stack.” We’ll cover what each does, the signal words that point to it, and the head-to-head distinctions the exam loves to test.
If you want the full exam blueprint first, read the AWS AI Practitioner Exam Guide 2026 and the AIF-C01 domains breakdown. For the generative-AI layer, the Amazon Bedrock guide is the companion to this post.
The Three Layers of the AWS AI/ML Stack
AWS frames its AI and ML offerings as three layers, and understanding the layering helps you answer “which service?” questions correctly.
| Layer | What it is | Who uses it | Examples |
|---|---|---|---|
| AI Services | Pre-trained, API-driven, no ML expertise needed | Developers, business teams | Rekognition, Comprehend, Textract, Lex, Polly |
| ML Services | Build, train, and deploy your own models | Data scientists, ML engineers | Amazon SageMaker |
| ML Frameworks & Infrastructure | Low-level compute and frameworks | Advanced ML practitioners | EC2 (GPU), Trainium, Inferentia, TensorFlow/PyTorch |
The defining trait of the AI Services layer — the focus of this article — is that the models are already trained by AWS. You don’t bring data to train them; you call an API and get a result. The signal words “no ML expertise,” “pre-trained,” “fully managed API,” and “without building a model” all point at this layer. When a scenario instead says “train a custom model on our own data,” that’s SageMaker, the ML layer.
Vision: Amazon Rekognition
Amazon Rekognition is the managed computer vision service. It analyzes images and video and can:
- Detect objects, scenes, and activities (labels).
- Perform facial detection, analysis, and comparison (and face search against a collection).
- Read text in images (text-in-image, e.g., a sign or license plate).
- Flag unsafe or inappropriate content (content moderation).
- Detect personal protective equipment (PPE).
Signal words: “image,” “video,” “faces,” “object/scene detection,” “content moderation.” If the input is a picture or a video frame, think Rekognition.
The classic trap is Rekognition text-in-image vs. Textract. Rekognition reads short bits of text in a natural scene; Textract is for documents (see below). If the question is about photos, it’s Rekognition; if it’s about forms or scanned pages, it’s Textract.
Documents: Amazon Textract
Amazon Textract goes beyond simple OCR. It extracts:
- Printed and handwritten text from scanned documents.
- Forms as key-value pairs (e.g., “Name: Jane Doe”).
- Tables with structure preserved.
- Specialized data via analyzers for invoices, receipts, and identity documents.
Signal words: “scanned document,” “invoice,” “form,” “table extraction,” “key-value pairs,” “PDF.” Textract understands document structure, which plain OCR does not. Pair it mentally with Rekognition: Textract = documents, Rekognition = photos/video.
Text and Language: Amazon Comprehend
Amazon Comprehend is the managed natural language processing (NLP) service. Given text, it can extract:
- Sentiment (positive, negative, neutral, mixed).
- Entities (people, places, organizations, dates).
- Key phrases and the dominant language.
- PII detection and redaction.
- Topic modeling across a document set.
- Custom classification and custom entity recognition (you can train these on your labels).
Signal words: “sentiment,” “entities,” “analyze customer feedback/reviews,” “detect PII in text,” “what language is this.” There’s also Amazon Comprehend Medical for extracting medical information (conditions, medications, dosages) from clinical text — watch for the word “medical/clinical.”
Speech: Amazon Transcribe and Amazon Polly
These two are mirror images, and the exam tests the direction:
| Service | Direction | What it does | Signal words |
|---|---|---|---|
| Amazon Transcribe | Speech → text | Converts audio/voice to written transcripts; supports speaker identification and medical transcription | ”transcribe,” “captions,” “call recordings to text” |
| Amazon Polly | Text → speech | Converts written text into lifelike spoken audio (many voices/languages) | “text to speech,” “read aloud,” “voice response,” “audio narration” |
Remember the arrows: Transcribe turns talking into text; Polly (think “parrot”) turns text into talking. Amazon Transcribe Medical is the healthcare-tuned variant.
Translation: Amazon Translate
Amazon Translate is neural machine translation between languages. Signal words: “translate,” “localize content,” “multilingual.” It’s frequently combined with others in a pipeline — for example, Transcribe (audio→text) → Translate (language A→B) → Polly (text→audio) to build real-time translated voice. Recognizing such service chains is a common exam pattern.
Conversational AI: Amazon Lex
Amazon Lex builds conversational interfaces — chatbots and voice bots. It provides the same technology that powers Alexa: automatic speech recognition (ASR) and natural language understanding (NLU) to recognize intents and slots (parameters). Lex integrates with AWS Lambda for fulfillment logic and with Amazon Connect for contact centers.
Signal words: “chatbot,” “virtual agent,” “intent,” “conversational interface,” “voice or text bot.” If the scenario is about building a bot that understands user requests, it’s Lex. (Contrast with Amazon Q, the generative-AI assistant — covered with Bedrock in the generative-AI material.)
Personalization and Forecasting
| Service | Purpose | Signal words |
|---|---|---|
| Amazon Personalize | Real-time recommendations (products, content) using the same tech as Amazon.com | ”recommendation engine,” “personalized suggestions,” “you might also like” |
| Amazon Forecast | Time-series forecasting (demand, inventory, finance) | “predict future demand,” “time-series,” “forecast sales/inventory” |
Note: AWS has shifted some forecasting and recommendation capabilities into other tooling over time, but for the AIF-C01 exam, Personalize = recommendations and Forecast = time-series prediction remain the expected mappings.
Search and Fraud
| Service | Purpose | Signal words |
|---|---|---|
| Amazon Kendra | Intelligent enterprise search with natural-language queries across internal documents | ”search internal documents,” “natural-language enterprise search,” “knowledge base search” |
| Amazon Fraud Detector | Detect potentially fraudulent online activity (payments, new accounts) | “fraud detection,” “fake accounts,” “suspicious transactions” |
| Amazon Q | Generative-AI assistant for business and builders | ”AI assistant,” “ask questions about your data,” “generative” |
Kendra is worth a second look because it pairs naturally with generative AI: it’s a common retrieval source in Retrieval-Augmented Generation (RAG) pipelines, where Kendra finds relevant documents and a Bedrock model generates the answer. The prompt engineering guide and Bedrock post cover the generation half.
The Decision Table: Pick the Service Fast
This is the single most valuable thing to memorize for the AI-services questions. Map the scenario to the service:
| The scenario describes… | The service is… |
|---|---|
| Analyzing images or video, detecting faces/objects | Amazon Rekognition |
| Extracting text/forms/tables from scanned documents | Amazon Textract |
| Sentiment, entities, or PII in text | Amazon Comprehend |
| Medical entities from clinical notes | Amazon Comprehend Medical |
| Speech to text (transcripts, captions) | Amazon Transcribe |
| Text to speech (audio narration) | Amazon Polly |
| Translating between languages | Amazon Translate |
| Building a chatbot / voice bot | Amazon Lex |
| Product/content recommendations | Amazon Personalize |
| Time-series forecasting | Amazon Forecast |
| Enterprise search over internal docs | Amazon Kendra |
| Detecting online fraud | Amazon Fraud Detector |
| Building/training a custom ML model | Amazon SageMaker |
| Generative AI with foundation models | Amazon Bedrock |
Commonly Confused Pairs (Memorize These)
The exam manufactures difficulty by putting two near-neighbors in the same question. Drill these distinctions:
- Rekognition vs. Textract — photos/video vs. documents/forms.
- Transcribe vs. Polly — speech→text vs. text→speech.
- Comprehend vs. Kendra — analyze text (sentiment/entities) vs. search text (find documents).
- Lex vs. Polly — build a conversational bot vs. just speak text aloud.
- AI Services vs. SageMaker — call a pre-trained API vs. train your own model.
- AI Services vs. Bedrock — task-specific managed models vs. generative foundation models.
A quick code-shaped example of how “API-driven” feels in practice — you call a service, you don’t train it:
import boto3
comprehend = boto3.client("comprehend")
resp = comprehend.detect_sentiment(
Text="The onboarding was smooth and the support team was fantastic.",
LanguageCode="en",
)
print(resp["Sentiment"]) # POSITIVE
print(resp["SentimentScore"]) # confidence per classYou won’t write code on the AIF-C01 (it’s multiple choice), but recognizing this “call an API, get a result, no training” shape is exactly what the exam rewards.
How These Services Map to the Exam Domains
The AI services thread through several AIF-C01 domains rather than living in just one:
| Domain | Weight | Where AI services appear |
|---|---|---|
| Fundamentals of AI and ML | 20% | Identifying AI use cases and matching them to managed services |
| Fundamentals of Generative AI | 24% | Bedrock, Amazon Q, SageMaker JumpStart (separate from this layer) |
| Applications of Foundation Models | 28% | Kendra as a RAG retrieval source; service chaining |
| Responsible AI | 14% | Bias and fairness considerations when using managed models |
| Security, Compliance, Governance | 14% | IAM, data privacy, and monitoring for AI workloads |
The AIF-C01 study plan sequences these domains, and the responsible AI guide covers the fairness and governance angle that often attaches to service-selection questions (“which service, and how do we use it responsibly?”).
Exam-Day Strategy for Service Questions
- Read for the input and output. Most of these questions resolve the moment you identify the data type: image → Rekognition, document → Textract, audio in → Transcribe, audio out → Polly.
- Eliminate by layer first. If the scenario says “no ML expertise” or “fully managed,” cross off SageMaker immediately and stay in the AI-services layer.
- Watch for service chains. When a question describes multiple steps (transcribe, then translate, then speak), the answer is often a combination, and the trick is ordering them correctly.
- Don’t overthink generative vs. task-specific. “Summarize/generate/converse” → generative (Bedrock/Q). “Detect/extract/classify a specific thing” → an AI service.
Practice the Recognition Reflex
Service-selection questions are won on speed and confidence. The difference between candidates who pass comfortably and those who run out of time is how instantly “extract key-value pairs from a scanned invoice” maps to Textract — without re-reading the options twice. That reflex only comes from working through many exam-style scenarios across every service.
Sailor.sh’s AWS Certified AI Practitioner (AIF-C01) Mock Exam Bundle gives you eight full-length, timed mock exams with detailed explanations that mirror the real exam’s style and difficulty — including the AI-services scenarios and the commonly-confused pairs covered here. Use it alongside the AIF-C01 study plan and the AWS AI certification path guide to map your next steps, and gauge your readiness with the AWS AI Practitioner practice questions. If you’re deciding where to start in AWS, the AI Practitioner vs Cloud Practitioner comparison will help.
Frequently Asked Questions
What’s the difference between Amazon Rekognition and Amazon Textract?
Rekognition is for images and video — detecting objects, scenes, faces, and short text in natural scenes, plus content moderation. Textract is for documents — extracting printed/handwritten text, forms (key-value pairs), and tables from scanned pages, invoices, and receipts. Rule of thumb: photo/video → Rekognition; document/form → Textract.
When should I use Amazon Comprehend versus Amazon Kendra?
Use Comprehend to analyze text — sentiment, entities, key phrases, language, and PII. Use Kendra to search text — natural-language queries that find relevant documents across an enterprise knowledge base. Comprehend tells you about a piece of text; Kendra finds the right text for you.
What’s the difference between Amazon Transcribe and Amazon Polly?
They’re opposites. Transcribe converts speech to text (audio in, transcript out). Polly converts text to speech (text in, spoken audio out). A common pattern chains them with Translate to build multilingual voice applications.
Are AWS AI services the same as Amazon SageMaker?
No. AI services (Rekognition, Comprehend, Textract, Lex, etc.) are pre-trained, API-driven services that need no ML expertise — you call an API and get a result. SageMaker is the ML platform for building, training, and deploying your own models on your own data. On the exam, “no ML expertise / pre-trained / fully managed API” points to AI services; “train a custom model” points to SageMaker.
Do I need to write code or train models for the AIF-C01 exam?
No. The AWS Certified AI Practitioner is a foundational, multiple-choice exam. You won’t write code or train models. You need to recognize what each service does, match services to business scenarios, and understand core AI/ML and generative-AI concepts. See the AIF-C01 exam guide for the format and the difficulty breakdown for what to expect.
Which AWS service do I use to build a chatbot?
Amazon Lex. It provides the automatic speech recognition and natural language understanding (intents and slots) to build conversational text or voice bots, and integrates with AWS Lambda for fulfillment. For a generative-AI assistant over your own data, consider Amazon Q or a Bedrock-based application instead — see the Amazon Bedrock guide.