AWS Certified AI Practitioner (AIF-C01) Exam Guide (2026)

If you can explain when to use Bedrock versus SageMaker, what a prompt injection looks like, and why a retrieval-augmented answer is safer than a raw foundation-model guess, you are in the right exam.

AIF-C01 is AWS’s foundational AI certification. It is not a coding test and it is not Machine Learning Engineer. It sits next to Cloud Practitioner: concepts, AWS service choices, and responsible-use judgment.

This guide is for people who will take the current AIF-C01 outline, including the 2026 objective updates around agentic AI, context, and output validation.

Who this exam is for

Take it if you are in product, sales, solutions, data, or early engineering and you need a shared vocabulary for AWS AI. Skip it if you already live in SageMaker training jobs and want a professional-level ML cert — that is MLA-C01 or the specialty path.

You do not need to train models from scratch. You do need to know what training, inference, fine-tuning, and RAG each change in a system.

Exam shape

AWS currently weights five domains:

Domain	Weight	What they actually test
Fundamentals of AI and ML	20%	Terms, problem types, supervised vs unsupervised, overfitting, bias
Fundamentals of generative AI	24%	Tokens, context windows, embeddings, limits of GenAI, agentic patterns
Applications of foundation models	28%	Largest slice. Prompts, RAG, customization, evaluation, Bedrock vs alternatives
Responsible AI	14%	Fairness, transparency, human oversight, toxicity, hallucination handling
Security, compliance, and governance	14%	IAM, encryption, data leakage, logging of model use, shared responsibility

Scroll horizontally to see all columns

Domain 3 is where people fail if they only memorized definitions. The items look like product decisions: freeze the base model and retrieve company docs, or fine-tune, or use a guardrail.

What changed in 2026

AWS expanded the outline rather than replacing AIF-C01. Study materials from 2024–2025 are still useful for ML basics. They are thin on:

Agentic AI and multi-agent workflows
Model Context Protocol (MCP) as a way models talk to tools
Context engineering, not just “write a better prompt”
Hallucination detection, grounding, and output validation
Newer Bedrock pieces (AgentCore, guardrails, identity) and names such as Amazon Q

If your notes stop at “prompt + temperature + Bedrock Knowledge Bases,” add a week for agents, tools, and how you prove the model stayed on-policy.

How to study without wasting a month

Week 1 — vocabulary that shows up in stems. Classification vs regression vs clustering. Training vs inference. Bias vs variance. Embeddings as numeric meaning, not “AI magic.” Generative vs discriminative.

Week 2 — AWS map. Bedrock for consuming foundation models. SageMaker when you train or host a custom model. Amazon Q for assistant-style work in AWS and the workplace. Rekognition, Textract, Transcribe, Comprehend, Polly, Kendra for single-purpose AI. Know when a purpose-built API is cheaper and safer than an LLM.

Week 3 — Domain 3 scenarios. For each story, force a choice: prompt only, RAG, fine-tune, continued pre-training, or a non-LLM API. Write one sentence for cost, data gravity, and who can see the prompts.

Week 4 — responsible AI and security. Prompt injection, sensitive data in logs, IAM for invoke permissions, encryption, Guardrails, human review on high-risk outputs. Shared responsibility still applies: AWS secures the service, you secure the prompts, plugins, and data you send.

Traps that look like easy elimination

Easy banks teach the wrong habit. Real AIF items often have two AWS services that both “do AI.”

Sentiment on support tickets is often Comprehend, not Bedrock.
“Answer from our 10,000 PDFs” is RAG, not fine-tuning.
“Make the model speak in our brand voice on generic tasks” can be prompt or a small customization; it is not automatically a new trained model.
A chatbot that must not invent refund policy is a guardrail and retrieval problem, not a larger context window.

If you can delete the scenario and still pick the answer from the wording alone, the question is too easy. The live exam will not do that.

How this maps to CloudFluently

We already have AWS AI Practitioner practice exam sets on the site. Treat those as a first pass. The hard sets we are writing next put two good answers in every item so you cannot eliminate your way to a pass.

Free stays free: notes, first exam banks, and this guide. The curated hard packs and timed mocks are the part we will put behind a checkout later.

Official outline: AWS Certified AI Practitioner (AIF-C01). Use AWS’s PDF for task statements. Use this page for how to study and what changed.