3.1.4. AWS AI Services for Common Business Needs
💡 First Principle: AWS AI services are pre-trained, fully managed models that require zero ML expertise. They exist because many business problems—text extraction, translation, sentiment analysis, image recognition, transcription—have been solved well enough that training custom models is unnecessary. The exam tests whether you recognize when an AI service is sufficient and when custom training or a foundation model is needed.
| AI Service | What It Does | No Custom Model Needed For |
|---|---|---|
| Amazon Rekognition | Image/video analysis | Face detection, object labeling, content moderation, celebrity recognition |
| Amazon Textract | Document text extraction | OCR, form data extraction, table extraction from scanned documents |
| Amazon Comprehend | NLP analysis | Sentiment, entities, key phrases, language detection, PII detection |
| Amazon Comprehend Medical | Medical NLP | Medical entity extraction, relationship detection, ICD-10 codes |
| AWS HealthLake | FHIR R4 health data store | Storing, querying and sharing health records; integrated medical NLP over unstructured clinical text |
| Amazon Translate | Language translation | Real-time and batch text translation between languages |
| Amazon Transcribe | Speech-to-text | Audio/video transcription, speaker identification, custom vocabularies |
| Amazon Polly | Text-to-speech | Natural-sounding speech generation from text |
| Amazon Lex | Conversational AI | Chatbots, voice assistants, IVR systems |
| Amazon Personalize | Recommendations | Product recommendations, personalized search, content ranking |
Chaining services: many business problems are pipelines of solved tasks. Recorded calls → Transcribe (speech to text) → Comprehend (sentiment). Scanned forms → Textract (text and tables) → Comprehend (entities). Direction matters: Transcribe turns speech into text, Polly turns text into speech.
Where foundation models fit: when the task is open-ended generation — drafting, summarizing, answering questions over documents — rather than a fixed analysis, the managed option is a foundation model on Amazon Bedrock, usually with retrieval. Section 3.4 covers that decision.
SageMaker Canvas — the no-code model builder (not a pre-trained API):
Canvas belongs beside the table above and is easy to confuse with it. Everything in the table is a pre-trained model you call through an API — you bring no data and train nothing. Canvas instead builds a model on your own data through a visual interface, selecting the algorithm for you. Both require zero ML expertise, which is why an exam scenario that says "no ML team" or "no data scientists" can land on either one.
| You need... | Reach for |
|---|---|
| A generic, already-solved task (sentiment, OCR, translation, transcription) | The matching AI service above — no data, no training |
| A model over your own tabular or time-series data, with no code written | SageMaker Canvas — AutoML behind a visual UI, with a model leaderboard |
| A model over your own data, with control of training and deployment | SageMaker AI built-in algorithms (XGBoost, DeepAR, Linear Learner) or Autopilot |
Canvas also fronts the AI services and Bedrock foundation models, so a business analyst can run Rekognition or a Bedrock model from the same interface without touching an SDK.
⚠️ Exam Trap: For time-series forecasting, read who is building it. A forecasting build with no ML team and no code is SageMaker Canvas; a team that wants control over training, hyperparameters and deployment is DeepAR, which also learns across many related series and so handles newly added items with little history.
⚠️ Exam Trap: If a question describes a standard NLP task (sentiment analysis, entity extraction) and the answer choices include both Amazon Comprehend and a custom SageMaker AI model, Comprehend is almost always correct—unless the question specifically mentions "domain-specific entities" or "custom categories" that Comprehend doesn't support. The exam penalizes over-engineering.
Reflection Question: A logistics company needs to: (1) extract delivery addresses from scanned invoices, (2) translate customer emails from Spanish to English, and (3) predict package delivery times based on historical route data. Which of these needs a custom model and which can use an AI service?