AI services, Bedrock, or SageMaker — choosing your altitude
Lesson 4 of 4 in The ML Lifecycle on AWS.
Choosing a modeling approach — and half the scenario questions everywhere else — comes down to one recurring decision: at what altitude do you solve this problem? Fly high with a pre-trained AI service and you write an API call but control almost nothing. Fly low with custom SageMaker training and you control everything but own every consequence. The exam rewards the lowest-effort option that still meets every stated requirement, judged along four axes: cost, control, data, and latency.
| Altitude | You bring | You control | It wins when… |
|---|---|---|---|
AI services — Textract, Rekognition, Comprehend, Transcribe, Translate, Polly, Lex, Personalize | An API call (sometimes a custom vocabulary or custom labels) | Almost nothing — AWS owns the model | The task is a commodity (OCR, speech, translation, sentiment, recommendations) and out-of-box quality meets the bar |
Bedrock FMs | Prompts, documents for RAG, optionally fine-tuning data | Model choice, prompting, grounding, safety policy | The task is generative or language-heavy and "no infrastructure to manage" is in the stem |
SageMaker built-ins — XGBoost, Linear Learner, DeepAR, k-means… | Your structured data and hyperparameters | Algorithm choice and tuning — not the algorithm code | Classic prediction on your own data: churn, fraud, forecasting — no custom network needed |
SageMaker custom training | Your own training script or container (via ECR) | Everything: architecture, loss, framework, serving stack | Requirements no managed option meets — novel architectures, strict latency on self-hosted models, full IP control |
Keep the AI-service roster as reflexes: Textract reads documents and forms, Rekognition sees images and video, Comprehend mines text (entities, sentiment, PII), Transcribe hears speech, Translate translates, Polly speaks, Lex runs conversational bots, Personalize recommends. When a scenario describes one of those tasks and signals no appetite for ML work, the AI service usually wins before Bedrock even enters the conversation — purpose-built, nothing to prompt, nothing to manage.
Pick your altitude
Interactive decision tree — outcomes:
- Use the pre-trained AI service
Highest altitude: an API call, no model to train or manage. Exam phrases that point here: "no ML expertise", "as quickly as possible", "minimal development effort".
- AI service + its customization hooks
Several AI services absorb domain tweaks without real ML work — custom vocabularies (Transcribe), custom entities and classifiers (Comprehend), custom labels (Rekognition). Still far less effort than training anything yourself.
- Bedrock + prompts (+ Knowledge Bases)
Serverless FM, pay per token, ground with Knowledge Bases when it must answer from your documents, wrap with Guardrails for policy. Exam phrases: "no infrastructure to manage", "reduce hallucinations with internal docs".
- Bedrock customization
Fine-tune a hosted FM on labeled examples, or Custom Model Import for compatible weights you already trained — same serverless API afterward. Watch the tiebreaker: if the stem stresses full control of the training loop, you've crossed into SageMaker territory.
- SageMaker AI custom training and hosting
Lowest altitude: maximum control over data, architecture, and latency — and maximum operational ownership. Correct when the requirements say so; a classic overkill distractor when they do not.
- SageMaker built-in algorithm
The managed middle: proven implementations where you supply data and hyperparameters (and let AMT tune them). The exam's default for tabular prediction.
Interactive checkpoint quiz (2 questions) — open this page in a browser to take it.