AWS AI Practitioner (AIF-C01) Sample Questions
12 original AIF-C01 practice questions across all five exam domains, each with the correct answer and why the other options are wrong.
In this article
- AIF-C01 exam facts at a glance
- Domain 1: Fundamentals of AI and ML
- Domain 2: Fundamentals of GenAI
- Domain 3: Applications of Foundation Models
- Domain 4: Guidelines for Responsible AI
- Domain 5: Security, Compliance, and Governance for AI Solutions
- How to score yourself
- How these questions were written
- Practice more AIF-C01 questions
Below are 12 original practice questions for the AWS Certified AI Practitioner (AIF-C01) exam, with the correct answer and a short explanation for each, including why the other options are wrong. They cover all five official domains in roughly the same proportion as the real exam, so you can find your weak areas before exam day.
Try each question before reading the answer. Guessing first and then checking is what makes practice stick; reading the answer straight away mostly feels like learning.
Key takeaways
- AIF-C01 has 65 questions (50 scored) in 90 minutes, and you pass at a scaled score of 700 out of 1,000.
- Domain 3, Applications of Foundation Models, is 28% of the score, so 4 of the 12 questions below come from it.
- Most AIF-C01 questions test whether you can match a business need to the right concept or AWS service, not whether you can build a model.
- When two answers both sound right, the cheaper, managed, less custom option is usually the one AWS wants for a foundational-level scenario.
AIF-C01 exam facts at a glance
These facts come from the official AIF-C01 exam guide and the AWS Certified AI Practitioner page, checked on 2026-10-21. For the full list of facts, what each domain covers and a study plan, see our AIF-C01 exam prep page.
| Fact | Official value |
|---|---|
| Questions | 65 (50 scored, 15 unscored) |
| Time | 90 minutes |
| Passing score | 700 (scaled score 100–1,000) |
| Price | 100 USD |
| Question types | Multiple choice, multiple response, ordering, matching |
| Domain | Weight | Questions below |
|---|---|---|
| 1. Fundamentals of AI and ML | 20% | 1–2 |
| 2. Fundamentals of GenAI | 24% | 3–4 |
| 3. Applications of Foundation Models | 28% | 5–8 |
| 4. Guidelines for Responsible AI | 14% | 9–10 |
| 5. Security, Compliance, and Governance for AI Solutions | 14% | 11–12 |
The current exam guide is version 1.1, published on April 30, 2026. It added agentic AI, context engineering, Model Context Protocol (MCP) and Amazon Bedrock AgentCore to the objectives, so older practice sets may miss those topics.
Domain 1: Fundamentals of AI and ML
Question 1
A retail company wants to group its customers by purchasing behavior. It has no existing labels for the groups. Which type of machine learning fits this task?
- A. Supervised learning with a classification model
- B. Unsupervised learning with clustering
- C. Reinforcement learning with a reward function
- D. Supervised learning with a regression model
Answer: B. Clustering is unsupervised: it finds groups in data that has no labels, which is exactly the situation described.
- A and D are wrong because supervised learning needs labeled examples (known groups or known numeric targets), and the company has none.
- C is wrong because reinforcement learning trains an agent through rewards for actions in an environment; there is no agent or action here.
Question 2
A model scores 98% accuracy on its training data but only 71% on new data it has never seen. What is the most likely problem?
- A. Underfitting
- B. Overfitting
- C. Data drift in the training set
- D. The inference endpoint is too small
Answer: B. Overfitting means the model learned the training data too closely, including its noise, so it does not generalize. A large gap between training and validation performance is the classic sign.
- A is wrong because an underfit model performs poorly on the training data too.
- C is wrong because drift describes production data changing over time, after training.
- D is wrong because instance size affects speed and cost, not accuracy.
Domain 2: Fundamentals of GenAI
Question 3
A team wants to build a chatbot that calls foundation models from several providers through one API, without provisioning or managing any servers. Which AWS service fits best?
- A. Amazon SageMaker AI with a custom training job
- B. Amazon Bedrock
- C. Amazon EC2 with GPU instances
- D. Amazon Comprehend
Answer: B. Amazon Bedrock is a fully managed service that gives API access to foundation models from Amazon and other model providers, so there is no infrastructure to run.
- A is wrong because a training job is for building or tuning your own model, which the team does not need.
- C is wrong because self-hosting models on EC2 means managing servers, the opposite of the requirement.
- D is wrong because Amazon Comprehend is a natural language processing service (entities, sentiment, classification), not a way to call general-purpose foundation models.
Question 4
A company notices that its generative AI costs rise sharply when users paste long documents into the prompt. What explains this?
- A. On-demand model pricing is usually based on the number of input and output tokens.
- B. Long prompts force the model to be retrained.
- C. Pricing is based on the number of API users.
- D. Long prompts always use a larger model.
Answer: A. Text models are typically billed per token, and a token is a chunk of text (often part of a word). A long pasted document adds many input tokens to every request, so cost grows with prompt length.
- B is wrong because inference does not retrain the model.
- C is wrong because the number of users does not set per-request cost.
- D is wrong because the model is chosen by the application, not by the prompt length.
Domain 3: Applications of Foundation Models
Question 5
A bank uses a foundation model to fill in a fixed report template. Answers vary too much between identical requests. Which inference parameter change helps most?
- A. Increase the temperature
- B. Decrease the temperature
- C. Increase the maximum output length
- D. Remove the stop sequences
Answer: B. Temperature controls randomness when the model picks the next token. A lower temperature makes output more focused and consistent, which suits template filling.
- A is wrong because a higher temperature makes output more varied and creative, the opposite of what the bank wants.
- C is wrong because output length limits how long a response can be, not how consistent it is.
- D is wrong because stop sequences only end generation at a chosen point.
Question 6
A company wants an internal assistant that answers questions from its HR policies. The policies change every month, and answers must reflect the current version. Which approach is most cost-effective?
- A. Pre-train a new foundation model on the policies
- B. Fine-tune a model every month on the new policies
- C. Use Retrieval Augmented Generation (RAG) with Amazon Bedrock Knowledge Bases
- D. Paste all policies into a few-shot prompt as examples
Answer: C. RAG retrieves the relevant passages from the current documents at question time and gives them to the model as context. Updating the knowledge base is enough when policies change, with no retraining.
- A is wrong because pre-training is by far the most expensive customization option.
- B is wrong because monthly fine-tuning costs time and money, and fine-tuning is better at teaching style or a task than at keeping facts current.
- D is wrong because pasting every policy into every prompt raises token cost and can exceed the context window; examples teach a format, not a document set.
If RAG is new to you, our guide to retrieval-augmented generation explains the pattern step by step.
Question 7
A prompt asks a model to classify support tickets into one of five categories, but the model keeps using its own category names. What is the simplest prompt engineering fix?
- A. Add a few labeled examples that use the exact category names (few-shot prompting)
- B. Fine-tune the model on 10,000 tickets
- C. Raise the temperature
- D. Switch to a larger model
Answer: A. A few worked examples with the exact labels show the model the expected output, and listing the allowed categories in the instruction helps too. It costs nothing to try and is the first step before heavier options.
- B is wrong because fine-tuning may work but is far more expensive than fixing the prompt first.
- C is wrong because a higher temperature makes output less predictable.
- D is wrong because a larger model may still invent labels if the prompt never shows the expected format.
Our post on zero-shot vs few-shot prompting covers when examples help and when they backfire.
Question 8
A team wants an automatic metric to compare model-generated summaries with reference summaries written by people. Which metric is designed for this?
- A. ROUGE
- B. Mean squared error
- C. F1 score for image classification
- D. Latency
Answer: A. ROUGE (Recall-Oriented Understudy for Gisting Evaluation) measures the overlap between generated text and reference text and is commonly used for summarization. The exam guide also lists BLEU (often used for translation), BERTScore and LLM-as-a-judge.
- B is wrong because mean squared error measures numeric prediction error in regression.
- C is wrong because the task is text summarization, not image classification.
- D is wrong because latency measures speed, not quality.
Domain 4: Guidelines for Responsible AI
Question 9
A lender wants to check whether its loan-approval model treats applicants of different age groups unfairly, both in the training data and in the model's predictions. Which AWS capability is built for this?
- A. Amazon SageMaker Clarify
- B. AWS CloudTrail
- C. Amazon Textract
- D. Amazon Bedrock Knowledge Bases
Answer: A. SageMaker Clarify detects potential bias in data and in model predictions, and it also provides feature-attribution explanations of how inputs affect predictions.
- B is wrong because CloudTrail logs API activity; it does not analyze model fairness.
- C is wrong because Textract extracts text and data from documents.
- D is wrong because Knowledge Bases provide retrieval for RAG applications.
Question 10
Regulators require a company to explain to each customer why the model rejected their application. Which choice best supports this requirement?
- A. Use the most accurate deep neural network available, regardless of interpretability
- B. Use a more interpretable model, or add explainability tooling that shows which features drove each decision
- C. Hide the model's inputs to protect intellectual property
- D. Increase the size of the training dataset
Answer: B. Responsible AI on AWS includes transparency and explainability. A simpler, interpretable model, or feature-attribution explanations, lets the company tell a customer which factors led to the decision. The trade-off between performance and interpretability is a recurring exam theme.
- A is wrong because a model nobody can explain may fail the regulatory requirement even if it is accurate.
- C is wrong because hiding inputs reduces transparency.
- D is wrong because more data can improve accuracy but does not explain individual decisions.
Domain 5: Security, Compliance, and Governance for AI Solutions
Question 11 (choose two)
A company's customer-facing chatbot built on Amazon Bedrock must block attempts to override its instructions and must not reveal customers' personal data in responses. Which two measures address these requirements? (Select TWO.)
- A. Configure Amazon Bedrock Guardrails with a prompt attack filter
- B. Configure Amazon Bedrock Guardrails sensitive information filters to mask or block personal data
- C. Increase the model's temperature
- D. Store the system prompt in a public Amazon S3 bucket
- E. Turn off logging so attackers cannot read it
Answer: A and B. Guardrails can detect prompt attacks such as jailbreaks and prompt injection, and its sensitive information filters can block or mask personally identifiable information in inputs and outputs. Together they cover both requirements.
- C is wrong because temperature changes randomness, not security.
- D is wrong because a public bucket exposes the system prompt to anyone.
- E is wrong because logs are needed for audit and incident response; the fix is to restrict access to them, not to turn them off.
More background on the attack itself is in our post prompt injection explained.
Question 12
An auditor asks the company to show who called the Amazon Bedrock APIs and when, over the last 90 days. Which AWS service provides this record?
- A. AWS CloudTrail
- B. Amazon Inspector
- C. AWS Trusted Advisor
- D. Amazon Macie
Answer: A. CloudTrail records API calls made in an AWS account, including who made each call, when and from where, which makes it the audit trail for this kind of question.
- B is wrong because Amazon Inspector scans workloads for software vulnerabilities.
- C is wrong because Trusted Advisor gives best-practice recommendations on cost, security and limits.
- D is wrong because Amazon Macie discovers sensitive data in Amazon S3.
How to score yourself
Count your correct answers out of 12 and look at which domains you missed rather than the total. Twelve questions are too few to predict an exam result; a single miss in Domain 3 matters more than one in Domain 4 only because Domain 3 has twice the weight on the real exam.
- 10–12 correct: move on to timed, full-length practice and keep reviewing the questions you miss.
- 7–9 correct: go back to the objectives of the domains you missed in the exam guide.
- 6 or fewer: start with the concepts. Our prompt engineering learning path covers the GenAI ideas behind Domains 2 and 3.
Questions you miss are worth more than ones you get right. Write down why you chose the wrong answer, then try a similar question a few days later.
How these questions were written
TokIQ Editorial wrote these questions from the task statements in the AIF-C01 exam guide (version 1.1). They are not real exam questions and are not taken from any question dump; AWS keeps its exam content confidential. Exam facts were checked against the AWS pages linked above on 2026-10-21. TokIQ is not affiliated with or endorsed by Amazon Web Services.
Studying for Microsoft's AI fundamentals exam as well? See our AI-901 practice questions, which also explain what replaced AI-900.
Practice more AIF-C01 questions
If you want a short set like this every day, TokIQ (on Google Play, coming soon to iPhone) has an AIF-C01 prep mode: a daily session that brings back the questions you missed with spaced repetition, leans toward your weakest domain, and includes mock exams. Each question cites the official AWS source it is based on, and there is an explanation after every answer, in your own language.
Ready for the associate level? See the MLA-C02 sample questions, or compare options in which AI certification to take first.
Frequently asked questions
Are these real AIF-C01 exam questions?
No. All 12 questions were written by TokIQ Editorial for practice, based on the objectives in the official AWS exam guide. Real exam questions are confidential: candidates accept a non-disclosure agreement, so sets sold as exam dumps are a risk, not a shortcut. Use them to check your understanding, not to memorize answers.
How many questions are on the AWS Certified AI Practitioner exam?
The AIF-C01 exam has 65 questions in 90 minutes. Fifty questions count toward your score and 15 are unscored questions that AWS is testing; you can't tell which ones they are. Unanswered questions count as wrong and there is no penalty for guessing (AWS exam guide, checked 2026-10-21).
What score do you need to pass AIF-C01?
You need a scaled score of 700 on a scale of 100 to 1,000. The scoring is compensatory, so you pass or fail on the exam as a whole and don't need to pass each domain separately (AWS exam guide, checked 2026-10-21).
Which AIF-C01 domain should I practice most?
Applications of Foundation Models is the largest domain at 28% of scored content, followed by Fundamentals of GenAI at 24% and Fundamentals of AI and ML at 20%. Responsible AI and Security, Compliance and Governance are 14% each. Spend practice time roughly in that proportion.
What question types are on the AIF-C01 exam?
The exam guide lists four types: multiple choice (one correct answer out of four), multiple response (two or more correct out of five or more, all needed for credit), ordering, and matching. Question 11 above is a multiple-response example.
- #AIF-C01
- #AWS Certified AI Practitioner
- #practice questions
- #certification prep