The short version
The exam has five domains. Applications of Foundation Models is the largest at 28%, and generative AI as a whole (domains 2 and 3) is 52% of scored content. Weightings below come from AWS exam guide version 1.2, published October 29, 2026 (the same as version 1.1, published April 30, 2026), and apply to scored content only.
| Domain | Weighting |
|---|---|
| 1. Fundamentals of AI and ML | 20% |
| 2. Fundamentals of GenAI | 24% |
| 3. Applications of Foundation Models | 28% |
| 4. Guidelines for Responsible AI | 14% |
| 5. Security, Compliance, and Governance for AI Solutions | 14% |
AWS stresses that the guide does not list everything that can appear. Use it as a map, not a fence.
Domain 1: Fundamentals of AI and ML (20%)
What it covers. Basic terms and how AI, ML, deep learning, generative AI, and agentic AI relate. Types of data and types of learning. When AI helps and when a plain rule or calculation is better. The ML lifecycle and how a model reaches production. Metrics such as accuracy, precision, recall, and F1, plus business metrics like cost per user and return on investment. Managed AWS services for common tasks.
In practice. If a company wants call recordings turned into text, that is Amazon Transcribe. If it wants to find key phrases in written reviews, that is Amazon Comprehend. If the goal is a house price as a number, that is regression, not classification.
Typical traps.
- Mixing up services that sit next to each other: Polly speaks, Transcribe listens, Translate converts languages, Comprehend analyzes existing text, Textract pulls text from scanned documents.
- Missing the “AI is not appropriate” clue, for example when the business needs a guaranteed exact outcome rather than a prediction.
- Confusing batch, real-time, asynchronous, and serverless inference. The clue is usually how fast results are needed and how traffic arrives.
What to practice. Matching a short scenario to a learning type, a technique, or a service. Say out loud why the other options do not fit.
Domain 2: Fundamentals of GenAI (24%)
What it covers. Tokens, chunking, embeddings, vectors, prompt engineering, transformers, foundation models, multi-modal models, and diffusion models. The foundation model lifecycle from data selection through deployment and feedback. Benefits (flexibility, conversation) and drawbacks (hallucinations, inaccuracy, nondeterminism). How token-based pricing affects cost. Context engineering and basic agentic concepts, including the Model Context Protocol, memory, and tools. AWS services for building generative applications.
In practice. A chatbot that sometimes states wrong facts is showing hallucination. A team that wants to add searchable company documents needs chunking and embeddings.
Typical traps.
- Treating “tokens” as words. A token is a piece of text, and pricing and limits are counted in tokens.
- Putting a lifecycle activity in the wrong stage, such as calling user feedback a pre-training step.
- Choosing a model on one factor when the question stresses another, such as latency or cost.
What to practice. Definitions you can explain in one sentence, plus the lifecycle in order.
Domain 3: Applications of Foundation Models (28%)
What it covers. How to select a model (cost, modality, latency, languages, size, customization, input and output length). The effect of inference parameters such as temperature. Retrieval augmented generation and the AWS services that can store embeddings. The cost and effort ladder of prompting, RAG, fine-tuning, distillation, and pre-training. Prompt techniques and their risks, including prompt versioning with Amazon Bedrock Prompt Management. Training and fine-tuning data. Evaluation methods and metrics such as ROUGE, BLEU, BERTScore, and LLM-as-a-judge.
In practice. If a company’s product facts change daily, RAG usually fits better than retraining. If the team needs a model to follow a house style across thousands of outputs, customization may be worth it.
Typical traps.
- Choosing fine-tuning where RAG or better prompting is enough. The cheaper approach often wins when the question hints at changing data or a small budget.
- Mixing up evaluation metrics: ROUGE and BLEU compare word overlap, while BERTScore compares meaning.
- Treating temperature as a setting that makes the model “smarter”. It changes how varied the output is.
What to practice. Given a constraint (data changes often, low latency, small team), pick an approach and justify it. This domain has the most scenario questions.
Domain 4: Guidelines for Responsible AI (14%)
What it covers. Features of responsible AI such as fairness, inclusivity, robustness, safety, and veracity. Legal and business risks of generative AI. Dataset qualities like diversity and balance. Bias and variance, overfitting and underfitting. Transparency and explainability, including tools such as SageMaker Model Cards and Amazon Bedrock Guardrails, and human-centered design.
In practice. A hiring model that does worse for one group needs a look at the training data balance. A model that scores well on training data and poorly on new data is overfitting.
Typical traps.
- Mixing up transparency (how the system is built and used) and explainability (why an output happened).
- Assuming more data from the same biased source fixes bias.
What to practice. Naming the problem from a symptom, then the fix.
Domain 5: Security, Compliance, and Governance for AI Solutions (14%)
What it covers. IAM roles and policies, encryption, Amazon Macie, AWS PrivateLink, the shared responsibility model, and Guardrails. Data lineage and source documentation. Secure data handling. Prompt injection, data leakage, output filtering, and logging. Grounding to improve accuracy. Governance and compliance services such as AWS Config, Amazon Inspector, AWS Artifact, AWS CloudTrail, and AWS Trusted Advisor.
In practice. CloudTrail records who called which API. Macie finds sensitive data in S3. Artifact provides compliance reports.
Typical traps.
- Picking a service that sounds right but does a different job, such as Macie for logging or CloudTrail for finding personal data.
- Forgetting that the customer is responsible for security in the cloud, including their data and access settings.
What to practice. A one-line job description for each security and governance service in the exam guide.
More guides for AIF-C01
- AIF-C01 Exam Format, Scoring and Retakes
- AIF-C01 Sample Questions With Answers
- Is the AIF-C01 Exam Hard?
- AIF-C01 Prerequisites and Experience
- AIF-C01 Renewal and Retakes Explained
- Is AIF-C01 Worth It? An Honest Look
- AIF-C01 Cheat Sheet: Key Facts to Review
- AIF-C01 vs MLA-C02: Which First?
- AWS AIF-C01 vs AI-901: Which First?
Next step
Put these into a schedule with the study plan. To test yourself by domain, try our AIF-C01 practice tests, which report your results for each of the five domains.