The AI services shortcut: Textract, Rekognition, Comprehend
Lesson 3 of 5 in Choosing a Modeling Approach.
The cheapest model is the one AWS already trained. The AI services are pre-trained models behind plain APIs — no training data, no training jobs, no endpoints to size. The exam's recurring pattern: a stem describes a common perception or language task, one option names the matching AI service, and the distractors describe building it yourself. Unless a requirement forces customization, don't build what AWS sells pre-trained.
| Service | Modality | It does | Classic exam cue |
|---|---|---|---|
Textract | Documents | Extracts text, forms, tables from scans/PDFs | "scanned invoices," "extract fields from forms" |
Rekognition | Images & video | Objects, faces, text-in-image, content moderation | "detect objects in photos," "moderate uploads" |
Comprehend | Text | Sentiment, entities, key phrases, PII detection | "customer review sentiment," "redact PII" |
Comprehend Medical | Clinical text | Medical entities, PHI detection | "clinical notes," "extract medications" |
Transcribe | Speech | Speech-to-text (incl. call audio) | "call recordings to text" |
Translate | Text | Language-to-language translation | "support 12 languages" |
Polly | Text → speech | Lifelike text-to-speech | "read articles aloud" |
Lex | Conversation | Chat/voice bots with intents and slots | "self-service chatbot" |
Personalize | Interactions | Recommendations and personalized ranking | "recommend products," no ML team |
Interactive sorting exercise: Buy, don't build: match each task to the pre-trained service that already does it.
When does the shortcut run out?
When the task is unique to your domain (the service has no feature for it), when the provided customization knobs can't reach your quality bar, or when you must own the model and its weights. Then step down one altitude: a Bedrock FM, or SageMaker with your own data.
Textract or Rekognition for documents?
Textract is purpose-built for documents — text, key-value pairs, tables, layouts. Rekognition reads text in photos (signs, labels) but does not understand forms. Scanned paperwork → Textract, always.
Comprehend or a Bedrock FM for text analysis?
If the need matches Comprehend's fixed menu (sentiment, entities, PII), take the service. If the need is open-ended — summarize, draft, reason over documents — that is generative territory: a foundation model on Bedrock.
Where does Personalize sit?
It is the pre-trained answer to recommendation: you supply interaction data, it trains and serves recommenders without you touching a model. Building a recommender on SageMaker is the answer only when the stem demands algorithmic control Personalize does not expose.
Interactive checkpoint quiz (2 questions) — open this page in a browser to take it.