Applications of Foundation Models
What you will learn
In this domain, you learn the main AI/ML services that AWS provides. The central theme is knowing when to use Amazon Bedrock (a foundation model API), Amazon SageMaker (a machine learning platform), Amazon Q (a business-oriented assistant), and purpose-specific AI services (Comprehend, Rekognition, Transcribe, and others). It is the largest domain, accounting for 28% of the AIF-C01 exam.
Key points
- Amazon Bedrock - a service that lets you invoke multiple foundation models through a unified API. You can use models from vendors such as Anthropic and Meta, along with Amazon's own Nova, in a serverless manner
- Bedrock Prompt Management - saves prompts with variables, compares variants, and lets applications call them by specifying a version
- Amazon SageMaker - an ML platform that provides end-to-end support from data preparation through model development, training, and deployment
- Amazon Q - a business-oriented generative AI assistant. There are Q Developer (development support) and Q Business (internal document search)
- Purpose-specific AI - Comprehend (text analysis), Translate (translation), Transcribe (speech to text), Polly (text to speech), Rekognition (image and video analysis), Textract (extracting text from documents)
- Search, conversation, and review - Kendra (semantic search over internal documents), Lex (conversational bots), Personalize (recommendations), and A2I (human review of AI predictions)
- Bedrock Knowledge Bases - a Bedrock feature that makes it easy to implement RAG with your own data
- Model distillation - a customization that trains a small model (the student) on the answers of a large model (the teacher) to lower the cost and latency of inference
- Bedrock Agents - a feature that lets you build AI agents on Bedrock that autonomously execute multiple steps
- AI agents - break a request into steps, call tools, and autonomously carry the work through to the end. Strands Agents is one example of an SDK for building them, and Bedrock AgentCore is the foundation for running them in production
- Evaluating foundation models - in addition to benchmarks, evaluation in which people score answers (human-in-the-loop), and LLM-as-a-judge (Bedrock model evaluation), check alignment with business goals through task completion rate, satisfaction, and cost per interaction
- Bedrock Guardrails - a guardrail feature to prevent inappropriate statements and leakage of personal information
Terms and concepts
Amazon Bedrock
A fully managed service that lets you invoke foundation models from multiple vendors through a single API (as of August 2026, these include models from Anthropic, Meta, and Mistral AI, along with Amazon's own Nova, and the lineup keeps growing). It can be used in a serverless manner and provides additional features such as Knowledge Bases (RAG), Agents, and Guardrails.
Bedrock Prompt Management
Amazon Bedrock Prompt Management is a feature for creating, testing, and saving prompts that you reuse (as of October 2026). You can put variables in a prompt and swap in values for each use, and compare the output of multiple variants that differ in model or inference parameters to choose the better one. You save versions as you refine a prompt, and applications call a prompt by specifying a version, so you can keep using a validated version in production while trying out a new one.
Amazon SageMaker
A platform that supports the entire process of ML model development (data preparation, training, evaluation, deployment, operation). It provides development with SageMaker Studio, training with SageMaker Training, and production inference with SageMaker Endpoint. It is powerful when you want to build your own model from scratch.
→ ML Development with SageMaker / → Amazon SageMaker Glossary
Amazon Q (Q Developer / Q Business)
Amazon Q is a business-oriented generative AI assistant provided by AWS. Q Developer supports code generation, review, and AWS resource operations, while Q Business is an internal knowledge assistant that answers from internal documents using RAG. It is provided in a SaaS form and is ready to use right away.
→ Enterprise AI with Q Business / → Coding Assistance with Q Developer
When to use purpose-specific AI services
Comprehend (sentiment analysis, entity extraction), Translate (machine translation), Transcribe (speech to text), Polly (text to speech), Rekognition (face detection, object detection, content moderation), and Textract (extracting structured text from PDFs and images) are each APIs optimized for their specific purpose. They can be adopted easily, quickly, and at low cost.
→ NLP Analysis with Comprehend / → Image Analysis with Rekognition
Bedrock Guardrails and Knowledge Bases
Knowledge Bases is a Bedrock feature that builds RAG from data sources such as S3, automating the integration of search and prompts. Guardrails is a filter that blocks inappropriate content, personal information, specific topics, and the like, ensuring the safety needed for enterprise use.
Model distillation
Model distillation is a customization that uses the answers of a large, capable model (the teacher) to train a small, fast, inexpensive model (the student) for a specific use. In Amazon Bedrock model distillation, the teacher generates answers from prompts you provide, or the teacher's answers recorded in production invocation logs are used, and the student is fine-tuned on those answers (as of October 2026). In terms of customization cost, prompt engineering (in-context learning) is the cheapest, and the cost grows in the order of RAG, fine-tuning, and pre-training. Distillation takes training effort, but because the resulting model is small, it lowers the cost and latency of inference during operation.
The role of AI agents and how they are used in business
An AI agent is a mechanism that breaks a user's request into steps, gathers the information it needs, and calls APIs and tools to carry the work through to the end. In business, agents are used for answering inquiries and carrying out procedures, automating processing that spans multiple systems, and creating reports based on research into internal data. Some services, such as AWS Transform, let agents handle the migration of older applications and servers. For the tools for building agents and the foundation for running them in production, see "AWS tools for building generative AI" in "Fundamentals of Generative AI." There is also Amazon Bedrock Agents, which you build by configuring it within Bedrock (as of October 2026).
→ Mainframe Modernization with AWS Transform / → Amazon Bedrock Glossary
How to evaluate and align with business goals
The performance of a foundation model is measured by combining automatic evaluation on prepared benchmark datasets, evaluation in which people score answers (human-in-the-loop), and LLM-as-a-judge, which has a separate LLM do the scoring. As of October 2026, Amazon Bedrock model evaluation supports these three methods and can also evaluate retrieval and answers for RAG that uses Knowledge Bases. Good metric values do not necessarily mean the model is useful for the business. Also check metrics tied to business goals, such as task completion rate (the share of requests handled through to the end), user satisfaction, and cost per interaction.
The difference between Kendra and Q Business
Kendra is a machine-learning-based enterprise search engine that searches internal documents by meaning and returns the matching documents and passages. Q Business goes beyond search: its generative AI assembles the answer itself and returns it as a finished assistant. In short, Kendra is a search building block (usable as the retrieval side of RAG), while Q Business is a ready-made solution that bundles search plus generation - that is the contrast between the two.
→ Enterprise Search with Kendra / → Enterprise AI with Q Business
Personalize / Lex / A2I
Personalize is a recommendation service based on the same technology that powers Amazon.com's recommendations, enabling personalization such as "customers who viewed this item also viewed." Lex is a service for building voice and text chatbots using the same conversational technology as Alexa, commonly used for automated responses in contact centers. A2I (Augmented AI) is a service for building workflows that route only low-confidence AI predictions to human reviewers, and it is used in combination with Textract and Rekognition.
→ Recommendation Engines with Personalize / → Building Chatbots with Lex / → Textract and Human Review with A2I