AIF-C01 Cheat Sheet: Key Facts to Review

The short version

This sheet is a last-pass review, written in our own words. It does not cover everything on the exam, and AWS says its own guide is not a comprehensive list either. Use it after you have studied, then check gaps against the AWS exam guide. It follows guide version 1.2 (October 29, 2026).

Exam facts at a glance

Item	Fact
Time and questions	90 minutes, 65 questions (50 scored, 15 unscored)
Passing score	700 on a scaled 100 to 1,000
Scoring	Compensatory: only the overall result counts
Guessing	No penalty; unanswered counts as wrong
Question types	Multiple choice, multiple response, ordering, matching
Weightings	AI/ML 20%, GenAI 24%, Foundation Model applications 28%, Responsible AI 14%, Security and governance 14%
Retake	14 calendar days after a fail

More detail: exam format and renewal and retakes.

Domain 1: Fundamentals of AI and ML (20%)

Key concepts

AI is the broad field. ML learns patterns from data. Deep learning uses neural networks with many layers. Generative AI creates new content. Agentic AI adds goals, tools, and multi-step action.
Learning types: supervised (labeled data), unsupervised (finding structure), reinforcement (reward and penalty).
Training builds the model. Inference uses it. Inference can be batch, real-time, asynchronous, or serverless.
Metrics: accuracy, precision, recall, F1. Business metrics: cost per user, ROI, customer feedback.

Confusable pairs

Classification predicts a category. Regression predicts a number.
Precision asks “of the items I flagged, how many were right?” Recall asks “of the real cases, how many did I find?”
Overfitting fits training data too closely. Underfitting misses the pattern even in training data.
Transcribe turns speech into text. Polly turns text into speech. Translate converts languages. Comprehend analyzes text. Textract pulls text and data from scanned documents. Lex builds chatbots.

Decision rules

Result needed right now for each request: real-time. Large pile of data, no hurry: batch. Long-running requests or large payloads: asynchronous. Spiky, unpredictable traffic: serverless.
Missing a case costs more than a false alarm (screening, fraud): favor recall. A false alarm costs more: favor precision.
Rules or exact calculations will do, or the outcome must be fully deterministic: AI may not be the right tool.
Traditional ML vs foundation model (new in 1.1): think of regulation, explainability, and operating constraints.

Domain 2: Fundamentals of GenAI (24%)

Key concepts

A token is a chunk of text, not a word. Pricing and limits are counted in tokens, so longer prompts and outputs cost more and take longer.
Embeddings are numeric vectors that capture meaning; similar meanings sit close together. Chunking splits documents so they can be embedded and retrieved.
Foundation models are large pretrained models. Multimodal models handle more than one kind of input. Diffusion models are common for images.
Hallucination: fluent output that is wrong. Nondeterminism: the same prompt can give different answers.
Context engineering is the work of deciding what information goes into the model’s context. Agentic concepts include tools, memory, and the Model Context Protocol (MCP), which connects agents to external systems.

Confusable pairs

Prompt (what you send at run time) vs training data (what shaped the model before).
Token limit on input vs on output.
Amazon Bedrock (access to foundation models through one service) vs Amazon SageMaker AI (build, train, and host models).

Decision rules

Scenario mentions search by meaning over documents: embeddings and a vector store.
Scenario mentions cost rising with long prompts: token-based pricing.
Scenario says answers are sometimes false: hallucination, then grounding.

Domain 3: Applications of Foundation Models (28%)

Key concepts

Choose a model by cost, latency, modality, languages, size, customization options, and input and output length.
Inference parameters: temperature changes how varied the output is. Top-p and top-k limit which tokens can be picked. Max length and stop sequences control where output ends.
RAG retrieves relevant material and adds it to the prompt. Vector stores hold embeddings.
Customization ladder, from cheapest to costliest: prompt engineering, RAG, fine-tuning, distillation, pre-training from scratch.
Prompt techniques: zero-shot, few-shot, chain-of-thought. Prompt versioning is handled with Amazon Bedrock Prompt Management.
Evaluation: ROUGE and BLEU compare overlap with a reference text, BERTScore compares meaning, LLM-as-a-judge uses a model to score outputs, human evaluation covers what metrics miss.

Confusable pairs

RAG (adds fresh knowledge at run time) vs fine-tuning (changes model behavior using labeled examples).
Zero-shot (no examples) vs few-shot (a few examples in the prompt).
Low temperature (steadier output) vs high temperature (more varied).
Pre-training (learn from huge unlabeled data) vs fine-tuning (adapt with a smaller labeled set).

Decision rules

Data changes daily: RAG, not retraining.
Needs a consistent style or format across many outputs: consider fine-tuning.
Small budget, quick start: prompt engineering first.
Summaries or translations compared with a reference: ROUGE or BLEU. Meaning matters more than wording: BERTScore.

Domain 4: Guidelines for Responsible AI (14%)

Key concepts

Responsible AI features include fairness, inclusivity, robustness, safety, and veracity.
Bias in training data produces biased outputs. Diverse, balanced, curated data helps.
Transparency is about how the system is built and used. Explainability is about why an output happened.
Tools: Amazon SageMaker Model Cards (document a model), SageMaker Clarify, Amazon Bedrock Guardrails (filter inputs and outputs), Amazon Bedrock Model Evaluations.
Human-centered design includes user-feedback mechanisms.

Decision rules

A group gets worse results: check dataset balance and fairness first.
Great on training data, poor on new data: overfitting.
Need documentation of a model’s purpose and limits: model cards.
Need to block topics or harmful content in a chatbot: guardrails.

Domain 5: Security, Compliance, and Governance (14%)

Key concepts

Shared responsibility: AWS secures the cloud itself; you secure your data, access, and configuration.
IAM roles and policies grant least privilege. Encryption protects data at rest and in transit. PrivateLink keeps traffic on the AWS network.
Prompt injection tries to override instructions. Data leakage exposes sensitive content. Output filtering and validation, grounding, and logging reduce risk.

Confusable pairs

CloudTrail records who called which API. Config tracks resource configuration and changes. Macie finds sensitive data in S3. Inspector scans for vulnerabilities. Artifact provides compliance reports. Trusted Advisor gives best-practice checks.

Decision rules

“Who did this and when”: CloudTrail.
“Is there personal data in this bucket”: Macie.
“Need an audit report for a regulator”: Artifact.
“Keep model calls off the public internet”: PrivateLink.

Common traps

Picking the familiar service name before reading the constraint in the question.
Choosing fine-tuning when RAG or better prompts are enough.
Reading “temperature” as “intelligence”.
Mixing up transparency and explainability.
Forgetting that multiple response, ordering, and matching questions give credit only for fully correct answers.

Last-day checklist

I can state the five weightings from memory.
I can explain each confusable pair above in one sentence.
I have done at least one full timed set of 65 questions.
I checked the exam guide revision history for new topics.
ID, booking time, and (if online) room and equipment are ready. See the exam day guide.
I will answer every question and flag the long ones for later.

More guides for AIF-C01

Next step

Revisit domains explained for the reasoning behind these rules, and test yourself with our AIF-C01 practice tests, which report results for each domain.

Try these practice exams