AWS AI services have expanded faster than almost any other part of the platform. Writing a comprehensive feature list would be out of date within weeks. This post takes a different approach: covering the four tiers of AWS AI and how to decide which one fits your problem. The tiers do not change even when the individual services do.
The four tiers
Pre-built AI services are managed services that solve a specific, well-defined problem without requiring any machine learning expertise. You call an API, you get an answer.
Bedrock provides access to foundation models via API. You are not running or training a model. You are choosing from a catalogue of pre-trained models and using them to build applications.
SageMaker is a full ML platform for teams that need to train, fine-tune, or deploy their own models. Requires ML expertise to use effectively.
Amazon Q is an AI assistant designed to be connected to your own data and tools. It exists in several forms depending on the use case.
Pre-built AI services
These are the right starting point for the majority of use cases. The problem is already solved, the model is already trained, and you pay per API call. No infrastructure to manage.
Textract
Document extraction. Where traditional OCR fell apart on structured forms and tables, Textract understands context. Feed it a PDF or image of an invoice, a mortgage application, or a patient form and it returns the structured data. Common in financial services, healthcare, and legal workflows.
Transcribe
Speech to text. Supports real-time and batch transcription, speaker identification, custom vocabularies for domain-specific terminology, and PII redaction built in. The redaction capability matters for any regulated workload where transcripts cannot contain card numbers or personal identifiers.
Comprehend
Natural language processing. Entity recognition, sentiment analysis, key phrase extraction, and PII detection on text. Useful downstream of Transcribe for extracting insight from transcripts, or directly on any text corpus you want to analyse at scale.
Rekognition
Image and video analysis. Face detection and comparison, content moderation, object and scene detection, text in images. Content moderation use cases often reach for this first because it handles the heavy lifting of flagging unsafe content before human review.
Polly
Text to speech. Realistic voice synthesis with multiple voices and languages. Useful when you need to generate audio at scale rather than maintaining audio files manually.
Translate
Language translation. Supports real-time and batch translation across a large number of language pairs. Integrates cleanly with the rest of the pre-built services.
Bedrock
Bedrock is the right tier when the pre-built services do not cover your use case and you want to build on top of a foundation model without managing one yourself.
You choose from a catalogue of models including Amazon’s own Nova and Titan families, Anthropic’s Claude, Meta’s Llama, Mistral, and others. Different models suit different tasks, languages, and cost points. You can switch models without changing your application architecture.
Bedrock is not just model access. The platform includes:
Knowledge Bases connect a foundation model to your own data via retrieval-augmented generation (RAG). The model answers questions grounded in your documents rather than its training data alone. This is the pattern to reach for when you need an AI that knows about your business, not just the internet.
Agents enable multi-step reasoning where the model can call APIs and take actions to complete a task. You define the tools it can use; the model figures out the sequence of steps.
Guardrails provide content filtering and PII redaction applied consistently across your Bedrock usage. Useful for any customer-facing application where you need to control what the model will and will not respond to.
SageMaker
SageMaker is the right tier when Bedrock does not give you what you need: either because the available foundation models are not suitable for your problem, you need to fine-tune a model on proprietary data, or you need full control over the training and deployment pipeline.
It is a significant platform in its own right. Expect to invest time in it and have ML engineering capability available. For teams that need it, it covers everything from data labelling through training, evaluation, and production deployment.
For most AWS application teams, SageMaker is not the right starting point. Start with the pre-built services or Bedrock and escalate to SageMaker only when those tiers cannot meet your requirements.
Amazon Q
Amazon Q comes in several forms that serve different purposes.
Q Developer is a code assistant integrated into IDEs and the CLI. It generates code, explains existing code, and can handle larger tasks like upgrading Java applications or transforming code between frameworks. It is AWS-aware, so it understands CloudFormation, CDK, and AWS service patterns.
Q Business connects to your internal data sources, documents, and tools to answer employee questions in natural language. Think of it as a search and Q&A layer across your organisation’s knowledge. It respects the access controls of the underlying data sources.
Q in QuickSight adds natural language querying to your BI dashboards. You describe the chart or analysis you want rather than building it manually.
Which tier should I use?
| Problem | Tier |
|---|---|
| Document extraction, transcription, image analysis, NLP, translation | Pre-built services |
| Building an application on top of an LLM | Bedrock |
| Fine-tuning or training your own model | SageMaker |
| AI assistant connected to your own data | Q Business |
| Code assistance in the IDE | Q Developer |
Start at the top of that list and only go further down when the tier above cannot meet the requirement. The pre-built services are the fastest path to production, the lowest operational burden, and the lowest risk. Foundation models via Bedrock are the right choice for open-ended language tasks. SageMaker is powerful but carries real complexity cost.
The services within each tier will continue to evolve. The tiers themselves are stable.
