Jun 12, 2026

7 min read

AWS Machine Learning: Picking the Right Service for the SAA Exam

In the SAA exam, a machine learning question usually isn’t testing whether you can train a model — it’s testing whether you can recognize the input data type and the task the prompt describes. The question drops a scenario — “automatically generate subtitles for videos”, “analyze sentiment in customer comments”, “extract data from scanned invoices” — and almost always there’s exactly one service that’s the intended answer.

The key insight: nearly every service in this section is a pre-trained, fully managed AI service — you just call an API, hand it data, and get a result; you don’t build a model. The one exception is SageMaker, where you build/train/deploy your own model. You get roughly 90 seconds per question, so recognition has to be nearly reflexive. This article is that map: for each service I summarize what it does, the core features, the use case, and most importantly — the keywords that give it away in the prompt.

Important note: These are only basic cues for quickly picking an answer under exam pressure. In the real world, choosing an AI/ML service demands far more thought: accuracy on real data, cost per request, data residency and compliance requirements, whether you need to train on your own data, and whether a pre-trained service is good enough or you must build with SageMaker. A keyword rarely maps to exactly one production-correct choice the way it does on the exam.

1. Amazon Rekognition — Image and Video Analysis

Rekognition is a pre-trained computer vision service (computer vision — letting machines “see” and understand the contents of images/video) that analyzes images and video through an API.

Detects objects and scenes (labels for things and settings in an image).
Facial analysis (age, emotion, wearing glasses…), plus face comparison and recognition.
Celebrity recognition.
Content moderation — flags inappropriate/unsafe content (nudity, violence).
Reads text that appears in images (text in image), detects personal protective equipment (PPE).
Integrates with Kinesis Video Streams for real-time video analysis.

Use case: Analyze photo/video libraries, moderate user-uploaded content, face-based verification, count people in a frame.

Keywords: image / video analysis, face detection/recognition, object/scene detection, content moderation, inappropriate / unsafe content.

2. Amazon Transcribe — Speech to Text

Transcribe uses ASR to convert spoken audio into text.

Automatic speech recognition with high accuracy.
Automatically redacts sensitive PII from the transcript.
Automatic language identification, multi-language support.
Custom vocabulary (a dictionary for domain-specific terms) and speaker diarization (who said what).

Use case: Transcribe calls/meetings, generate subtitles/captions for video, turn voicemail into text.

Keywords: speech to text, transcribe audio, subtitles / captions, convert audio to text.

3. Amazon Polly — Text to Speech

Polly does the reverse of Transcribe: it uses TTS to turn text into lifelike speech.

Many natural-sounding voices (neural voices) across languages.
Control delivery with SSML (pauses, emphasis, speed).
Lexicons (custom pronunciation dictionaries) and Speech Marks (sync text with audio for lip-sync/highlight effects).

Use case: Give apps a voice, turn articles into audio, support the visually impaired, voice a virtual assistant.

Keywords: text to speech, generate voice / audio from text, lifelike speech, read text aloud.

4. Amazon Translate — Language Translation

Translate is neural machine translation (translating in the context of the whole sentence rather than word by word), converting text between languages naturally.

Translates across many language pairs while preserving context.
Localizes app/website content.
Batch or real-time translation, with custom terminology support.

Use case: Localize multi-language UI/content, translate user-generated text, translate documents in bulk.

Keywords: translate, language translation, localization, multilingual.

5. Amazon Lex + Amazon Connect — Chatbots and Cloud Contact Center

Lex is the service for building conversational chatbots, using the same ASR + NLU technology as Alexa. Connect is a cloud-based customer contact center that can use Lex to power voice/chat bots.

Lex: understands users via intents (what they want) and slots (the information to collect), calling Lambda for business logic.
Connect: a virtual contact center with no upfront infrastructure, integrating Lex to automate calls.
Lex + Connect together build an intelligent IVR.

Typical flow — booking an appointment by phone:

A customer dials Connect’s virtual contact-center number (call). Connect streams the call audio to Lex (stream); Lex recognizes the speech and identifies the “book an appointment” intent along with the slots it needs (date, time). Once the intent is clear, Lex calls a Lambda function (invoke) to run the business logic — check availability and write the booking into the CRM system (schedule). The whole thing runs automatically, with no human agent.

Use case: Customer-support chatbots, conversational booking/ordering assistants, automated phone lines.

Keywords: chatbot, conversational interface, contact center / call center, IVR.

6. Amazon Comprehend — Text Analysis (NLP)

Comprehend is a pre-trained NLP service that reads and extracts information from free-form text.

Sentiment analysis (positive/negative/neutral).
Entity and key-phrase extraction.
Language detection, PII detection.
Topic modeling — automatically groups documents by topic.

Use case: Analyze customer feedback, auto-classify/label documents by topic, measure sentiment on social media.

Keywords: sentiment analysis, NLP, extract entities / key phrases, find topics in text.

7. Amazon Comprehend Medical — NLP for Medical Text

Comprehend Medical is a version of Comprehend specialized for clinical text, understanding medical terminology that plain Comprehend often handles poorly.

Detects PHI in records.
Extracts medical entities: conditions, medications, dosages, symptoms.
Links to standard medical coding systems (ICD-10-CM for diagnoses, RxNorm for medications).

Typical flow — feeding Comprehend Medical:

Comprehend Medical has three common ingestion paths. Clinical documents are kept in the Amazon S3 object store and submitted for batch analysis (documents). Continuously generated data is streamed through Kinesis Data Firehose for near-real-time processing (real-time). And spoken input — a doctor dictating notes, or a patient conversation — is first run through Amazon Transcribe to turn it into text (text). All three converge on Comprehend Medical, which extracts the medical entities and PHI (extract).

Use case: Extract structured information from doctor’s notes, patient records, and prescriptions.

Keywords: medical / clinical text, PHI, healthcare NLP, doctor's notes / patient records.

8. Amazon SageMaker AI — Platform to Build Your Own ML Models

SageMaker is a fully managed platform for you to build, train, and deploy your own machine learning models — fundamentally different from the pre-trained services above.

Notebooks for data exploration, plus a library of built-in algorithms.
Training jobs and hyperparameter tuning.
Deploy a model as an endpoint for real-time inference, or batch transform for bulk processing.
Ground Truth for data labeling.

Use case: A data science team building a custom model for a problem no pre-trained service can solve — this is the only “build it yourself” service in the group.

Keywords: build / train / deploy your own model, custom ML model, data scientists, machine learning platform.

9. Amazon Kendra — Intelligent Document Search

Kendra is an ML-powered enterprise search service that lets you search an internal document store with natural-language questions instead of just keyword matching.

Understands natural-language questions and returns a direct answer extracted from the documents, not just a list of links.
Built-in connectors for many sources (S3, SharePoint, Salesforce…).
Incremental learning from user feedback.

Use case: Internal Q&A systems, searching a company’s document store / knowledge base.

Keywords: document search, enterprise search, knowledge base, natural language search.

10. Amazon Personalize — Real-Time Personalized Recommendations

Personalize is a recommendation engine using the same technology as Amazon.com, producing personalized recommendations in real time.

Recommends products/content tailored to each user.
Personalized ranking (re-orders a list by preference) and “similar items” suggestions.
Updates in real time based on new user behavior.

Use case: Product recommendations in e-commerce, content suggestions, personalized email marketing.

Keywords: recommendations, personalization, "recommended for you", personalized ranking.

11. Amazon Textract — Extract Data from Documents

Textract uses advanced OCR to extract not just text but structured data (key-value pairs in forms, tables, handwriting) from scanned documents, PDFs, and images.

Reads text, forms (key-value), and tables from documents.
Recognizes handwriting too.
Preserves the data structure so it can flow straight into a processing system.

Use case: Process invoices, IDs, forms, and financial/tax/medical documents.

Keywords: extract text from documents, scanned documents / PDF, forms & tables, OCR, process invoices / IDs.

Tips & Tricks — Recognize the Keyword and Pick the Service

This is the most important part for the exam. Read the prompt, catch the keyword, map it straight to a service.

By input data type

When the prompt says…	Pick
Images, video, faces, objects	Rekognition
Spoken audio → need text out	Transcribe
Text → need speech out	Polly
Text that needs translating to another language	Translate
Text to analyze for sentiment / entities / topics	Comprehend
Medical / clinical text	Comprehend Medical
Scanned documents / PDF / forms / tables	Textract
Conversation / chatbot / contact center	Lex (+ Connect)
A search query over a document store	Kendra
Behavior history → need recommendations	Personalize

Pre-trained or build your own

When the prompt says…	Pick
”data scientists”, “train your own model”, “custom model”, “custom algorithm”	SageMaker
A common AI task (image, speech, translation, NLP…) via a ready API	The matching pre-trained service

Wrapping Up

One line to remember:

Recognize the data type and the task first, and the service falls out.

In the exam, the keyword → service reflex saves you precious time. But in real life, the right question isn’t “which service matches the keyword”, but whether a pre-trained service is accurate enough for the problem, what it costs per request, what compliance constraints the data carries, and when it’s worth training your own model with SageMaker.