An overview of AWS’s AI and Machine Learning services — why they exist, how they differ, and when to use each.
Visual Overview
Why Does AWS Have So Many AI Services?
AWS offers AI/ML services at multiple levels of abstraction to serve different users and use cases:
| User Type | Need | AWS Approach |
|---|---|---|
| Developers | Add AI features without ML expertise | Pre-built AI APIs (Rekognition, Comprehend, Polly) |
| Data Scientists | Build, train, and deploy custom models | Full ML platform (SageMaker) |
| GenAI Builders | Use foundation models (LLMs) in apps | Managed access to FMs (Bedrock) |
| Enterprises | Productivity tools powered by AI | AI assistants (Amazon Q) |
Key Insight: AWS gives you a spectrum — from “just call this API” to “build everything from scratch” — so you pick the right level for your skills and requirements.
The Three Layers of AWS AI
1. Managed AI Services (Pre-built APIs)
For: Developers who want AI features without building models
Plug-and-play APIs for common AI tasks. No ML knowledge required.
| Service | What It Does |
|---|---|
| Rekognition | Image/video analysis — face detection, object recognition, content moderation |
| Comprehend | NLP — sentiment analysis, entity extraction, language detection |
| Polly | Text-to-speech with natural voices |
| Transcribe | Speech-to-text transcription |
| Translate | Real-time language translation |
| Textract | Extract text and data from documents (OCR++) |
| Lex | Build conversational bots (powers Alexa) |
| Kendra | Intelligent enterprise search |
| Personalize | Real-time recommendations |
| Forecast | Time-series forecasting |
2. ML Platform (SageMaker)
For: Data scientists and ML engineers who build custom models
Amazon SageMaker is a complete ML platform for the full lifecycle:
| Stage | SageMaker Capability |
|---|---|
| Prepare | Data labeling (Ground Truth), data wrangling, feature store |
| Build | Notebooks, built-in algorithms, bring your own code |
| Train | Managed training, distributed training, hyperparameter tuning |
| Deploy | Real-time endpoints, batch inference, multi-model endpoints |
| Monitor | Model monitoring, bias detection, explainability |
2024 Updates:
- SageMaker Unified Studio — single interface for data, analytics, and ML
- SageMaker Lakehouse — access data lakes and warehouses directly
- Native integration with foundation models (LLMs)
When to use SageMaker: You need full control, custom algorithms, or training on proprietary data. ML expertise required.
3. Generative AI (Bedrock)
For: Builders who want to use foundation models without managing infrastructure
Amazon Bedrock provides serverless access to foundation models from multiple providers — now offering ~100 models via Bedrock Marketplace:
| Provider | Models Available |
|---|---|
| Amazon | Titan (text, embeddings, image), Nova (multimodal) |
| Anthropic | Claude 4.x family (Claude 4.5 Sonnet, etc.) |
| Meta | Llama 4 (Maverick, Scout) |
| Mistral | Mistral Large 3, Ministral 3 series |
| Cohere | Command, Embed, Rerank |
| AI21 Labs | Jamba models |
| Stability AI | Stable Diffusion, SDXL |
Key Bedrock Features:
- Single API to access all models
- Knowledge Bases for RAG (Retrieval-Augmented Generation)
- Agents for building autonomous workflows
- Guardrails for content filtering and responsible AI
- Fine-tuning to customize models on your data
- Model evaluation to compare outputs
When to use Bedrock: You want to build GenAI apps without training models. Choose a foundation model, call the API, optionally customize with your data.
Bedrock vs SageMaker — When to Use Which
| Criteria | Use Bedrock | Use SageMaker |
|---|---|---|
| Goal | Use pre-trained foundation models | Train custom models from scratch |
| ML Expertise | Low to medium | High |
| Data | Your data for RAG/fine-tuning | Your data for training |
| Infrastructure | Serverless, fully managed | Managed but more control |
| Customization | Fine-tuning, prompt engineering | Full algorithm control |
| Best for | Chatbots, content generation, summarization | Fraud detection, custom NLP, proprietary algorithms |
Can you use both? Yes! Many enterprises use Bedrock for GenAI features and SageMaker for custom ML models. They complement each other.
Other AWS AI/ML Tools
| Service | Category | Purpose |
|---|---|---|
| Amazon Q | GenAI Assistant | AI assistant for business (Q Business) and developers (Q Developer) |
| Amazon Quick Suite | Agentic Workspace | Unifies QuickSight + Q Business for research, flows, automation |
| PartyRock | GenAI Playground | No-code tool to experiment with Bedrock-powered apps |
| AWS Trainium / Inferentia | Custom Chips | AWS-designed chips for ML training and inference (cost savings) |
TL;DR
AWS AI services exist at three levels:
- Managed AI Services (Rekognition, Comprehend, etc.) — Pre-built APIs for common tasks. No ML required.
- SageMaker — Full ML platform for building, training, and deploying custom models. For data scientists.
- Bedrock — Serverless access to foundation models (Claude, Llama, Titan, etc.). For GenAI apps.
Pick based on your needs: API for quick features → Bedrock for GenAI → SageMaker for custom ML.
Subfolders (Detailed Notes)
- AWS Managed AI Services — Rekognition, Textract, Transcribe, Polly, Translate, Comprehend, Lex, Kendra
- AWS GenAI — Bedrock, Amazon Q, PartyRock
- AWS Machine Learning — SageMaker AI deep dive
Resources
AWS Machine Learning
Official hub for all AWS AI/ML services.Amazon Bedrock
Overview of foundation model access and features.Amazon SageMaker
Complete ML platform documentation.