AWS AI services: match the service to the task

AWS AI services are pre-trained, API-based services for common problems (images, documents, text, speech, conversation, recommendations) that need no model training. The decision rule: name the input and the output, and pick the service whose job that is; build a custom model in SageMaker AI only when no managed service covers the task or the exam scenario demands a custom algorithm. Most exam questions on this topic are one-line matching problems, so the skill is knowing the input and output of each service cold.

The input-to-output cheat sheet

The AWS documentation phrases these as follows. Rekognition analyses images and videos stored in Amazon S3 and also covers face comparison and liveness. Textract detects typed and handwritten text and offers AnalyzeExpense for invoices and receipts, AnalyzeID for identity documents, and a Queries feature to ask for specific fields. Comprehend recognises entities, key phrases, language, sentiment and PII. Transcribe converts audio to text either in real time (streaming) or from files in S3 (batch). Lex V2 builds voice and text bots with automatic speech recognition and natural language understanding, integrated with Lambda.

Worked example: a call-centre pipeline

A company records support calls in Spanish and wants English summaries of customer sentiment, with sensitive data flagged.

1. Transcribe turns the recorded audio in S3 into Spanish text (batch, because the calls are already recorded; streaming would be for live calls). Transcribe can redact PII content, which costs extra.
2. Translate converts the Spanish transcript to English.
3. Comprehend returns the sentiment of each transcript and detects PII entities.
4. Polly is only added if the English summary must be read aloud, for example in an accessibility feature.

Each step has one job and one output that becomes the next step’s input. No model was trained. A distractor answer that proposes training a sentiment model in SageMaker AI is the classic wrong option at this level: Bedrock vs SageMaker AI covers when custom building is justified, and the SageMaker AI components guide covers what building involves.

Availability changes the exam guides have not caught up with

AWS has been closing services to new customers. A service in “maintenance” cannot be onboarded by new customers, existing customers keep using it, and AWS does not add features. “Sunset” has a timeline after which AWS ends operation. Verified on October 9, 2026:

The rest of Comprehend and Rekognition, and Transcribe, Translate, Polly, Lex, Textract and Personalize, showed no availability notice on their documentation pages when checked. Note the scope of the two partial changes: they affect named features, not the whole service. The current AIF-C01 and MLA-C02 outlines still name Comprehend, Transcribe, Translate, Lex, Polly, Textract and Rekognition, but do not list Forecast, Fraud Detector, Kendra or Lookout. If an older practice question says “use Amazon Forecast”, treat the item as dated: the concept (a managed time-series forecasting service) is still valid, but a new account cannot use that service.

Which exams test this

How the questions are built

The scenario describes a business need in plain words, and the answer is the service whose input and output match. Underline the data type first. “Scanned invoices” is Textract, not Rekognition, even though both read images; Rekognition detects text in a scene, Textract extracts structured fields from documents. “Customer reviews” is Comprehend. “Recorded meetings” is Transcribe, “live captions” is Transcribe streaming. “Product recommendations from purchase history” is Personalize. “Chatbot that books appointments” is Lex.

Constraint words decide the rest: “no ML expertise” or “least development effort” favours the managed service; “proprietary categories” points to the service’s custom variant (Comprehend custom classification, Rekognition Custom Labels) before SageMaker AI. The two classic confusions are Textract vs Rekognition for text, and Lex vs Comprehend: Lex runs a dialogue, Comprehend analyses text that already exists. For generative needs, the answer shifts to Bedrock; see RAG explained for the grounded-answer pattern that replaced search products like Kendra in AWS’s own guidance, and AWS IAM basics for the role that lets a Lambda function call any of these services.

Practice questions

Go deeper

• What is Amazon Rekognition? and What is Amazon Textract?, for the exact capabilities.
• What is Amazon Comprehend?, covering insights and custom models.
• AWS lifecycle changes, the definitions of maintenance, sunset and full shutdown.
• Related: AWS IAM basics and the AIF-C01 vs MLA-C02 comparison.