AWS AI Practitioner Fundamentals of GenAI Practice Questions (Domain 2, 24%)

Domain 2 is 24% of scored content. It covers tokens, chunking, embeddings and foundation models, token-based pricing, context engineering and agentic AI concepts such as MCP and multi-agent patterns, the strengths and limits of GenAI, and AWS services such as Amazon Bedrock, AgentCore and Strands Agents. Answer these 10 free questions with no signup, written to AWS's AIF-C01 exam guide v1.2.

Last updated November 11, 2026.

10 Free AIF-C01 Fundamentals of GenAI Questions with Answers

Question 1: Fundamentals of GenAI

Match each description to the multi-agent communication pattern it describes. 1. A coordinating agent assigns subtasks to worker agents and collects their results. 2. Each agent's output becomes the next agent's input, in a fixed order. 3. Agents message one another directly, with no central coordinator.

  1. 1 hierarchical, 2 peer-to-peer, 3 sequential
  2. 1 hierarchical, 2 sequential, 3 peer-to-peer
  3. 1 peer-to-peer, 2 sequential, 3 hierarchical
  4. 1 peer-to-peer, 2 hierarchical, 3 sequential

Correct answer: B. B is correct because a coordinator directing workers is hierarchical (supervisor), a fixed hand-off chain is sequential (pipeline), and direct agent-to-agent messaging without a coordinator is peer-to-peer. A is incorrect because a fixed order of hand-offs is sequential, while agents messaging freely without a set order is peer-to-peer. C is incorrect because a coordinator assigning work is the defining feature of a hierarchy, and the absence of any coordinator defines peer-to-peer. D is incorrect because a coordinator handing out subtasks rules out peer-to-peer, a fixed hand-off chain has no coordinator, and free messaging has no fixed order. Exam guide: Domain 2 (Fundamentals of GenAI), Task Statement 2.1.

Question 2: Fundamentals of GenAI

A support agent sends the full 40-turn chat history and the complete raw output of every tool call, often thousands of tokens of records it never uses, with each request. Its answers have started to drift, and its token costs are rising. Which combination of context-engineering changes best addresses both problems?

  1. Summarize older turns, and switch to a model with a larger context window
  2. Keep the full history, and fine-tune the model on examples of past tool outputs
  3. Keep only the needed tool fields, and raise the temperature for more variety
  4. Summarize older turns, and keep only the needed fields from each tool result

Correct answer: D. D is correct because summarizing history and trimming tool results to the needed fields both shrink the context to what matters, which improves focus and cuts input tokens. A is incorrect because summarizing helps, but a larger window does not remove the unused tool output that is causing drift and cost. B is incorrect because the full history keeps costs rising, and fine-tuning does not stop unused tool output from being sent on every request. C is incorrect because trimming tool output helps, but a higher temperature adds randomness, which makes drifting answers more likely. Exam guide: Domain 2 (Fundamentals of GenAI), Task Statement 2.1.

Question 3: Fundamentals of GenAI

A furniture design team wants to create new concept images from short text descriptions such as "walnut desk with brass legs". Which model type and process fit this goal?

  1. A diffusion model that removes noise step by step, guided by the text prompt
  2. A text-only LLM that predicts the next words of the description and then prints them
  3. An embedding model that finds the stored catalog image closest to the description
  4. A classification model that labels each design image with the right furniture style

Correct answer: A. A is correct because diffusion models generate images by repeatedly denoising random noise, with each step steered toward the text prompt, which is how they create new images from descriptions. B is incorrect because a text-only LLM generates text, not images, so it cannot produce a visual concept of the desk. C is incorrect because retrieving the nearest existing image does not create a new concept; it can only return something already in the catalog. D is incorrect because a classifier assigns labels to existing images; it does not generate new images from a description. Exam guide: Domain 2 (Fundamentals of GenAI), Task Statement 2.1.

Question 4: Fundamentals of GenAI

A retailer's service agent can issue refunds by calling a payments tool. Policy says any refund above $500 must be approved by a manager before it is paid. How should the agent's workflow meet this policy?

  1. Tell the agent in its instructions to be careful whenever a refund is large
  2. Give the agent long-term memory of every refund it has issued in the past
  3. Add a human approval step that pauses refunds above $500 until a manager agrees
  4. Lower the model's temperature so it is less likely to approve large refunds

Correct answer: C. C is correct because a human-in-the-loop step in the workflow stops the high-value action until a person approves it, which enforces the policy regardless of what the model decides. A is incorrect because an instruction to be careful does not create an approval step, so a large refund could still be paid without a manager. B is incorrect because remembering past refunds does not stop a new large refund from being paid without approval. D is incorrect because temperature only changes how random the output is; it cannot guarantee that a manager approves each large refund. Exam guide: Domain 2 (Fundamentals of GenAI), Task Statement 2.1.

Question 5: Fundamentals of GenAI

An online grocer adds a generative AI shopping assistant. Its stated goal is that more visitors who use the assistant go on to place an order. Which metric most directly measures progress toward this goal?

  1. Number of questions the assistant answers per day
  2. Model accuracy on a public reasoning benchmark
  3. Average number of tokens per assistant response
  4. Conversion rate of sessions that used the assistant

Correct answer: D. D is correct because conversion rate is the share of visits that end in a purchase, which is exactly the outcome the grocer wants the assistant to increase. A is incorrect because a high question count shows usage, not whether those visitors place an order. B is incorrect because benchmark scores describe the model in general, not its effect on the grocer's sales. C is incorrect because token count tracks cost and verbosity, not whether visitors go on to buy. Exam guide: Domain 2 (Fundamentals of GenAI), Task Statement 2.2.

Question 6: Fundamentals of GenAI

A lender wants to use an LLM to approve or decline small-business loans. Regulations require the lender to give each declined applicant the specific factors behind the decision, and to show that the same application gets the same outcome. Which approach best meets these requirements?

  1. Decide with an explainable, deterministic model and use the LLM to draft the letter
  2. Let the LLM decide and attach a sample of its training data to each letter
  3. Let the LLM decide and ask it to list its reasons in each decline letter
  4. Let the LLM decide at a temperature of zero so its decisions can be explained

Correct answer: A. A is correct because an explainable model gives consistent decisions with traceable factors, while the LLM adds value by writing the explanation in plain language without making the decision. B is incorrect because training examples do not explain how a specific decision was reached. C is incorrect because reasons an LLM writes about itself are generated text, not a reliable trace of how it decided, and its outputs can vary between runs. D is incorrect because a low temperature may reduce variation, but it does not make the model's reasoning interpretable for regulators. Exam guide: Domain 2 (Fundamentals of GenAI), Task Statement 2.2.

Question 7: Fundamentals of GenAI

A product team uses an LLM to brainstorm campaign themes. A team member complains that running the same prompt twice gives different ideas each time and asks for this to be fixed. What is the best response?

  1. Fix it by replacing the LLM with a rule-based theme generator
  2. Fix it by fine-tuning the model so that it always returns one theme list
  3. Keep it, since varied output is useful for brainstorming and adds more ideas
  4. Treat it as hallucination and add documents to ground every answer

Correct answer: C. C is correct because nondeterminism is a limitation when consistency matters, but for brainstorming, different ideas on each run are a benefit rather than a defect. A is incorrect because rules produce fixed, predictable output, which would lose the creative variety the team wants. B is incorrect because fine-tuning is costly and would work against the goal of generating a wide range of ideas. D is incorrect because different ideas on each run are variation, not invented facts, so grounding does not apply. Exam guide: Domain 2 (Fundamentals of GenAI), Task Statement 2.2.

Question 8: Fundamentals of GenAI

A five-person startup with no machine learning engineers wants to add a writing assistant to its app within a month. Which advantage of using a managed AWS generative AI service such as Amazon Bedrock matters most for this team?

  1. It lets the team skip testing, since managed models produce correct output
  2. It lowers the barrier to entry: no model training or hosting needed
  3. It requires the team to collect labelled data first, which improves quality
  4. It gives the team full control over the model's GPU instances and drivers

Correct answer: B. B is correct because managed foundation models are ready to use through an API, so a small team without ML skills can build features without collecting data, training models or running infrastructure. A is incorrect because managed models can still produce errors, so the team must test the feature as it would any other. C is incorrect because using a foundation model through Bedrock needs no labelled training data to get started. D is incorrect because a managed service hides the infrastructure; full hardware control is what self-hosting offers, at the cost of more work. Exam guide: Domain 2 (Fundamentals of GenAI), Task Statement 2.3.

Question 9: Fundamentals of GenAI

A publisher needs to summarize one million archived articles with a foundation model in Amazon Bedrock. Results are needed within a few days, and no user is waiting on any single summary. The team wants the lowest cost. Which approach fits best?

  1. Provisioned Throughput with a long-term commitment for the archive job
  2. A real-time SageMaker AI endpoint kept running until the job is done
  3. Bedrock batch inference, submitting the articles as one job at a lower price
  4. On-demand calls for each article, sent one at a time from a script

Correct answer: C. C is correct because batch inference processes large sets of prompts asynchronously at a lower price than on-demand calls, which suits a big job with no real-time need. A is incorrect because a commitment for reserved capacity suits steady ongoing traffic, not a one-time job that batch handles more cheaply. B is incorrect because an always-on real-time endpoint is designed for low-latency requests and adds hosting cost for work that needs no immediate answer. D is incorrect because on-demand pricing is higher than batch, and real-time responses add nothing when no one is waiting. Exam guide: Domain 2 (Fundamentals of GenAI), Task Statement 2.3.

Question 10: Fundamentals of GenAI

A retailer's product assistant on Amazon Bedrock handles steady, high traffic around the clock, and the business needs consistent throughput for the next six months. Finance asks whether Provisioned Throughput costs more than on-demand. What is the best answer?

  1. It costs the same as on-demand, since both charge for each token processed
  2. It always costs more than on-demand, so it is only worth it for custom models
  3. It always costs less than on-demand, since committed capacity is discounted
  4. It can cost less at sustained high volume, but idle capacity is still billed

Correct answer: D. D is correct because Provisioned Throughput charges for reserved model units over the term, so it pays off when traffic keeps them busy and guarantees throughput, but idle capacity is still billed. A is incorrect because Provisioned Throughput is billed for reserved capacity over time, not per token. B is incorrect because at sustained high volume it can be the cheaper option, and it also serves base models. C is incorrect because reserved capacity that sits idle is still billed, so it is cheaper only when utilization stays high. Exam guide: Domain 2 (Fundamentals of GenAI), Task Statement 2.3.

Fundamentals of GenAI Task Statements in the Exam Guide

TaskExam guide task statementDomainQuestions
2.1Explain the basic concepts of generative AIFundamentals of GenAI40
2.2Understand the capabilities and limitations of GenAI for solving business problemsFundamentals of GenAI26
2.3Describe AWS infrastructure and technologies for building GenAI applicationsFundamentals of GenAI28

More AWS AI Practitioner Practice

FlashGenius is not affiliated with or endorsed by Amazon Web Services. AWS and AWS Certified AI Practitioner are trademarks of Amazon.com, Inc. or its affiliates. These are original practice questions, not questions from the real exam. Your score here is practice accuracy, not an AWS score or a pass prediction.