2026.06.13 2026.06.13

AIF-C01 Practice Questions: 30 Items by Domain Weighting Plus a Case Study

swiftwand

This set gives you 30 AIF-C01 practice questions distributed by the official domain weighting (Domain 1: 6, Domain 2: 7, Domain 3: 9, Domain 4: 4, Domain 5: 4), followed by a case-study exercise and full answers. Work through it under time pressure — about 1.4 minutes per question — to simulate the real 65-question, 90-minute exam.

Before you start — three cross-domain solving principles
30 practice questions
Case-study exercise — one scenario, three questions
Answers and explanations
Scoring and shoring up weak points
Summary — toward a full-format final rehearsal
References

忍者AdMax

Before you start — three cross-domain solving principles

Pick the smallest sufficient solution. When options range from prompt engineering to continued pre-training, the cheapest method that meets the need is usually correct.
Read for the keyword. “Current or private knowledge” points to RAG; “consistent style” points to fine-tuning; “take actions” points to agents.
Eliminate deterministic-required cases. If a task must be exact every time, generative AI or ML is the wrong tool.

30 practice questions

Domain 1: Fundamentals of AI and ML (6)

Which best describes the relationship between AI, ML, and deep learning? (a) They are unrelated. (b) Deep learning contains ML contains AI. (c) AI contains ML contains deep learning. (d) ML and AI are identical.
A bank must compute exact interest owed on each account. Is ML appropriate? (a) Yes, use regression. (b) Yes, use classification. (c) No, the result must be deterministic. (d) Yes, use clustering.
Grouping customers into segments with no labels is an example of: (a) supervised learning, (b) unsupervised learning, (c) reinforcement learning, (d) transfer learning.
Which service converts speech to text? (a) Amazon Polly, (b) Amazon Transcribe, (c) Amazon Translate, (d) Amazon Rekognition.
On a highly imbalanced fraud dataset, which metric is most misleading on its own? (a) Precision, (b) Recall, (c) Accuracy, (d) F1.
Which is the correct early order in the ML lifecycle? (a) Train, then frame the business goal. (b) Frame the business goal, then collect data. (c) Deploy, then prepare data. (d) Evaluate, then frame the problem.

Domain 2: Fundamentals of Generative AI (7)

What is an embedding? (a) A compressed image, (b) a numeric vector capturing meaning, (c) a database index, (d) a prompt template.
Which model type is the basis of most image generators? (a) LLM, (b) diffusion model, (c) regression model, (d) decision tree.
“The same prompt can return different answers” describes which property? (a) hallucination, (b) non-determinism, (c) overfitting, (d) bias.
Which correctly orders the retrieval preparation flow? (a) embedding, chunking, document, vector. (b) document, chunking, embedding, vector. (c) vector, document, chunking, embedding. (d) chunking, vector, document, embedding.
Which AWS service gives managed access to foundation models from multiple providers via one API? (a) Amazon Q, (b) Amazon Bedrock, (c) Amazon Comprehend, (d) Amazon Textract.
A confident but factually wrong answer is called: (a) drift, (b) hallucination, (c) leakage, (d) variance.
Which is a no-code playground for building generative-AI apps? (a) PartyRock, (b) SageMaker Studio, (c) CloudFormation, (d) Athena.

Domain 3: Applications of Foundation Models (9)

A chatbot must answer from your company’s latest internal documents. Lowest-cost approach? (a) continued pre-training, (b) fine-tuning, (c) RAG, (d) train a new model.
To make output more deterministic, you should: (a) raise temperature, (b) lower temperature, (c) increase top-k, (d) add more tokens.
Order the four customization methods from lowest to highest cost: (a) RAG, prompt engineering, fine-tuning, continued pre-training. (b) prompt engineering, RAG, fine-tuning, continued pre-training. (c) fine-tuning, RAG, prompt engineering, continued pre-training. (d) continued pre-training, fine-tuning, RAG, prompt engineering.
Which lets a model call tools and complete multi-step tasks? (a) embeddings, (b) agents, (c) guardrails, (d) tokenization.
Including a few examples in the prompt is: (a) zero-shot, (b) few-shot, (c) chain-of-thought, (d) fine-tuning.
Malicious instructions hidden inside user input describe: (a) prompt injection, (b) drift, (c) overfitting, (d) chunking.
Which metric best evaluates summarization quality? (a) BLEU, (b) ROUGE, (c) RMSE, (d) accuracy.
The managed AWS service that wires up the RAG pipeline is: (a) Bedrock Knowledge Bases, (b) CloudTrail, (c) Macie, (d) Polly.
Which is a valid vector store for Bedrock Knowledge Bases? (a) Amazon OpenSearch Serverless, (b) Amazon SQS, (c) Amazon Route 53, (d) AWS Lambda.

Domain 4: Guidelines for Responsible AI (4)

Which AWS service detects bias in data and models and explains feature importance? (a) SageMaker Clarify, (b) CloudWatch, (c) Amazon Q, (d) Textract.
Most model bias originates in: (a) the inference endpoint, (b) the training data, (c) the API gateway, (d) the prompt length.
Which documents a model’s purpose, data, and limitations? (a) SageMaker Model Cards, (b) IAM policy, (c) S3 bucket policy, (d) CloudFormation template.
Routing low-confidence predictions to human reviewers is done by: (a) Amazon A2I, (b) Amazon Macie, (c) AWS Config, (d) Amazon Polly.

Domain 5: Security, Compliance, and Governance (4)

Which service discovers sensitive data such as PII in Amazon S3? (a) Amazon Macie, (b) Amazon Rekognition, (c) Amazon Comprehend, (d) AWS Shield.
To keep traffic to Bedrock on the private AWS network, use: (a) AWS PrivateLink, (b) an internet gateway, (c) a public subnet, (d) Route 53.
Which provides an audit log of API activity for governance? (a) AWS CloudTrail, (b) Amazon SNS, (c) Amazon EBS, (d) AWS Batch.
To filter unsafe content from a generative-AI app, apply: (a) Bedrock Guardrails, (b) a NAT gateway, (c) Spot Instances, (d) S3 lifecycle rules.

Case-study exercise — one scenario, three questions

Scenario: A small e-commerce team wants an assistant that answers customer questions using their constantly changing product catalog and policy documents, without exposing customer data to model training, on a tight budget.

Best customization approach for the changing catalog? (a) fine-tune weekly, (b) RAG over the catalog, (c) continued pre-training, (d) a rules engine.
How to keep customer data out of base-model training? (a) use Amazon Bedrock, where your data is not used to train base models, (b) email AWS to opt out, (c) it is impossible, (d) disable encryption.
How to prevent the assistant from producing unsafe responses? (a) raise temperature, (b) apply Bedrock Guardrails, (c) remove IAM, (d) use a larger model.

Answers and explanations

Domain 1: 1-(c) AI contains ML contains deep learning. 2-(c) exact interest must be deterministic, so ML is unsuitable. 3-(b) clustering with no labels is unsupervised. 4-(b) Transcribe is speech to text. 5-(c) accuracy misleads on imbalanced data. 6-(b) frame the business goal before collecting data.

Domain 2: 7-(b) an embedding is a meaning vector. 8-(b) diffusion models generate images. 9-(b) varying output is non-determinism. 10-(b) document, chunking, embedding, vector. 11-(b) Amazon Bedrock. 12-(b) a confident wrong answer is a hallucination. 13-(a) PartyRock is the no-code playground.

Domain 3: 14-(c) RAG for current internal documents. 15-(b) lower temperature for determinism. 16-(b) prompt engineering, RAG, fine-tuning, continued pre-training. 17-(b) agents call tools. 18-(b) few-shot. 19-(a) prompt injection. 20-(b) ROUGE for summarization. 21-(a) Bedrock Knowledge Bases. 22-(a) OpenSearch Serverless.

Domain 4: 23-(a) SageMaker Clarify. 24-(b) bias originates in training data. 25-(a) Model Cards. 26-(a) Amazon A2I for human review.

Domain 5: 27-(a) Amazon Macie. 28-(a) AWS PrivateLink. 29-(a) AWS CloudTrail. 30-(a) Bedrock Guardrails.

Case study: 31-(b) RAG handles a constantly changing catalog without retraining. 32-(a) on Amazon Bedrock your data is not used to train the base models. 33-(b) Bedrock Guardrails filter unsafe responses.

Scoring and shoring up weak points

Score yourself by domain, not just overall. Because the real exam is compensatory and weighted, a low score in Domain 3 (28%) hurts far more than the same gap in Domain 4 (14%). If you missed two or more in any domain, reread that domain guide before moving on. Aim for 80%+ on a timed run before booking the exam.

Summary — toward a full-format final rehearsal

These 30 questions plus the case study mirror the exam’s weighting and its newer formats. Repeat them until the cross-domain principles — smallest sufficient solution, RAG for private knowledge, deterministic-required means no ML — feel automatic. Then take a full timed practice exam as your final rehearsal, covered in the exam-strategy guide.