Free AIF-C01 Cheat Sheet 2026 — Topic Weights, Exam Traps & Quick Rules

Core Framework

AIF-C01 Domain Map (exam guide v1.2)

Short business scenarios with one-line answers: which learning type, service, metric or customization method fits. The October 2026 v1.2 guide includes agentic AI, MCP, context engineering, Strands Agents, Bedrock AgentCore, Kiro and Amazon Quick, so older courses miss tested material.

01
Fundamentals of AI and ML (20%) — learning types, regression vs. classification vs. clustering, inference modes, metrics
02
Fundamentals of GenAI (24%) — tokens, embeddings, hallucination, agents, MCP, AgentCore, Bedrock pricing options
03
Applications of Foundation Models (28%) — prompting, RAG, fine-tuning vs. pre-training, inference parameters, evaluation metrics
04
Guidelines for Responsible AI (14%) — bias, fairness, explainability, Clarify, Model Cards, Guardrails, human review
05
Security, Compliance and Governance for AI (14%) — shared responsibility, Macie, KMS, PrivateLink, CloudTrail, Config, Artifact

Scenario Traps

Wrong instinct vs correct approach

A fraud model shows 99% accuracy on data where fraud is under one percent of records

✕ Wrong instinct

Approve it because the accuracy is very high

✓ Correct approach

A model predicting 'not fraud' every time scores near 99% while catching nothing. Use recall when misses are costly, precision when false alarms are, and F1 when both matter. Check performance per subgroup for fairness too.

Marketing wants customer segments discovered from purchase history that has no labels

✕ Wrong instinct

Train a classification model on the customer table

✓ Correct approach

No labels and unknown groups mean unsupervised learning, specifically clustering. Classification needs predefined labeled categories, and reinforcement learning learns from rewards by acting in an environment.

A three-month Bedrock pilot has unpredictable usage and no budget commitment

✕ Wrong instinct

Buy Provisioned Throughput to guarantee capacity

✓ Correct approach

On-demand pricing charges per token with no commitment, which suits light or irregular use. Provisioned Throughput bills reserved capacity whether used or not; batch inference is cheaper only when results can wait.

A team wants a foundation model for a chatbot with no infrastructure to manage

✕ Wrong instinct

Deploy a model from SageMaker JumpStart

✓ Correct approach

JumpStart deploys models onto SageMaker AI endpoints the team operates. Amazon Bedrock is the serverless API to foundation models, so it meets a no-infrastructure requirement directly.

A question asks who must secure prompts and training data used with Amazon Bedrock

✕ Wrong instinct

AWS secures everything because Bedrock is fully managed

✓ Correct approach

Shared responsibility still applies: AWS secures the service infrastructure, while the customer controls IAM access, data, encryption choices and application prompts. The Generative AI Security Scoping Matrix shows customer responsibility growing from scope 1 to scope 5.

Quick Rules

Know these cold

▸ AI contains ML, ML contains deep learning, deep learning contains generative AI
▸ Number output is regression, known category is classification, unlabeled groups are clustering
▸ Real-time for instant answers, batch for offline tables, async for large payloads, serverless for sporadic traffic
▸ Overfitting means strong on training data and weak on new data — igh variance
▸ ROUGE scores summaries, BLEU scores translation, BERTScore credits paraphrased meaning
▸ Lower temperature gives more consistent output; it does not supply missing facts
▸ Few-shot and chain-of-thought change the prompt only; no model weights are updated
▸ Continued pre-training uses unlabeled domain text; fine-tuning uses labeled task examples
▸ Bedrock is serverless model access; JumpStart gives more control with more operations
▸ Agents plan, call tools and act; MCP connects them to tools; AgentCore runs them

Self Check

Can you answer these without checking your notes?

In this scenario: "A fraud model shows 99% accuracy on data where fraud is under one percent of records" — what should you do first?

A model predicting 'not fraud' every time scores near 99% while catching nothing. Use recall when misses are costly, precision when false alarms are, and F1 when both matter. Check performance per subgroup for fairness too.

In this scenario: "Marketing wants customer segments discovered from purchase history that has no labels" — what should you do first?

No labels and unknown groups mean unsupervised learning, specifically clustering. Classification needs predefined labeled categories, and reinforcement learning learns from rewards by acting in an environment.

In this scenario: "A three-month Bedrock pilot has unpredictable usage and no budget commitment" — what should you do first?

On-demand pricing charges per token with no commitment, which suits light or irregular use. Provisioned Throughput bills reserved capacity whether used or not; batch inference is cheaper only when results can wait.

Failure Patterns

Common Exam Mistakes — What candidates get wrong

Studying from a course that predates the agentic AI additions

The 2026 guide revisions added agentic AI, Model Context Protocol, context engineering, Strands Agents and Amazon Bedrock AgentCore. Know that MCP is a connection protocol, Strands builds agents in code, and AgentCore runs them with memory, identity, gateway and observability.

Choosing pre-training or fine-tuning when prompting or RAG would do

Customization cost climbs from in-context learning to RAG to fine-tuning to continued pre-training to pre-training. A brand-voice change needs examples or fine-tuning, fresh documents need RAG, and pre-training from scratch is almost never the answer.

Confusing SageMaker Clarify, SageMaker Model Cards and Bedrock Guardrails

Clarify detects bias and explains predictions with feature attributions. Model Cards only document purpose, data and limits; they measure nothing. Guardrails filter prompts and responses at runtime. Distractors swap these three deliberately.

Picking the audit or compliance service whose name sounds right

CloudTrail records who called which API. Config tracks resource configuration compliance. Inspector scans for vulnerabilities. Artifact provides AWS's own compliance reports. Trusted Advisor recommends but never enforces. Macie discovers PII in S3; KMS encrypts but cannot see content.

Mixing up Transcribe, Polly, Translate, Comprehend and Lex

Transcribe turns speech into text, Polly turns text into speech, Translate converts languages, Comprehend extracts sentiment, entities and PII from text, and Lex builds conversational bots. Expect these before any custom SageMaker AI model.

Continue your prep

Resources to help you pass AWS Certified AI Practitioner on your first attempt

Practice Questions

Strengthen weak areas with exam-style practice questions and detailed explanations.

Mock Exam

Simulate the real exam experience and assess your readiness under timed conditions.

Study Guide

Review key concepts, objectives, and exam topics in one place.

AI Tutor

Get personalized explanations, learning recommendations, and instant answers.

AIF-C01 Training

Follow a structured learning path designed to help you prepare efficiently.