Free AWS AI Practitioner Practice Exam (AIF-C01), Exam Guide v1.2

Start answering below: 10 questions across all five AIF-C01 domains, weighted like the exam, with no signup. You get your score by domain and an explanation for every answer.

Last updated November 11, 2026. Written to AWS's AIF-C01 exam guide v1.2 (October 29, 2026).

10 Free AWS AI Practitioner Practice Questions with Answers

Question 1: Fundamentals of AI and ML

A media company runs a model on uploaded video files of up to 1 GB. Each request takes up to 10 minutes to process, uploads arrive unpredictably throughout the day, and users need results within the hour rather than instantly. Which type of inference fits best?

  1. Batch inference that processes the whole day's uploads together each night
  2. Real-time inference on an always-on endpoint that answers each request within milliseconds
  3. Asynchronous inference that queues requests and returns results when each one completes
  4. Serverless inference that scales to zero between short, lightweight requests

Correct answer: C. C is correct because asynchronous inference is built for large payloads and long processing times, queues requests as they arrive, and can scale down when the queue is empty. A is incorrect because a nightly batch run would leave most users waiting many hours, which breaks the within-the-hour requirement. B is incorrect because real-time endpoints are meant for small, fast requests, and a 10-minute job on a 1 GB file exceeds what they are designed to handle. D is incorrect because serverless inference suits short requests with small payloads, not 1 GB files that take minutes to process. Exam guide: Domain 1 (Fundamentals of AI and ML), Task Statement 1.1.

Question 2: Fundamentals of AI and ML

A hospital screens scans for a rare but serious condition. A positive result sends the scan to a radiologist for review, which costs far less than a missed case. Model A has precision 0.92 and recall 0.70. Model B has precision 0.60 and recall 0.97. Which model should the hospital choose?

  1. Model A, because it has the better overall balance of precision and recall
  2. Model B, because high recall misses the fewest cases and extra positives only add reviews
  3. Model B, because high recall also means few false positives reach the radiologists
  4. Model A, because its high precision means that the cases it flags are rarely false alarms

Correct answer: B. B is correct because recall measures the share of true cases the model catches, and when a missed case is far costlier than a false alarm that only triggers a cheap review, recall is the metric to favor. A is incorrect because a balanced profile suits cases where both errors cost about the same, but here a missed case is far costlier than an extra review. C is incorrect because recall says nothing about false positives; Model B's 0.60 precision means about 4 in 10 flagged scans are false alarms, which is acceptable only because reviews are cheap. D is incorrect because precision reduces false alarms, but Model A would miss about 30% of real cases, which is the costliest error in this scenario. Exam guide: Domain 1 (Fundamentals of AI and ML), Task Statement 1.3.

Question 3: Fundamentals of GenAI

A company's AI agent must look up orders in an internal database and open tickets in a help desk tool. The team wants a standard way for agents to discover and call such tools without building a custom integration for each model. What should the team use?

  1. Model Context Protocol (MCP), an open standard for connecting agents to tools and data
  2. Fine-tuning the model on example database queries and help desk ticket formats
  3. Retrieval Augmented Generation, which adds retrieved documents to the model's prompt
  4. Prompt caching, which reuses repeated prompt prefixes across the agent's requests

Correct answer: A. A is correct because MCP defines a common way for agents to discover and call external tools and data sources, so one integration can serve many agents and models. B is incorrect because fine-tuning changes how the model writes, not how it connects to live systems, and it would need repeating for every model. C is incorrect because RAG supplies reference text to the model, but it does not give the agent a standard way to call tools that take actions, such as opening a ticket. D is incorrect because prompt caching reduces the cost and latency of repeated input; it does not connect the agent to any external tool. Exam guide: Domain 2 (Fundamentals of GenAI), Task Statement 2.1.

Question 4: Fundamentals of GenAI

A low-volume support assistant on Amazon Bedrock uses on-demand pricing and handles a few hundred requests a day. Each request sends a 6,000-token prompt containing the entire product manual plus the customer's question, and responses average 200 tokens. Costs are higher than expected. Which change most directly lowers the cost per request?

  1. Lower the maximum output length so each response stops at 100 tokens instead of 200
  2. Retrieve only the relevant manual sections so each request sends far fewer input tokens
  3. Raise the temperature so the model reaches each answer in fewer generation steps
  4. Move from on-demand pricing to Provisioned Throughput for the support assistant

Correct answer: B. B is correct because on-demand pricing charges per input and output token, and input is about 97% of the tokens in each request, so even though output tokens cost more per token, sending only relevant sections cuts cost the most. A is incorrect because output is a small share of each request's tokens, so halving it saves little and may cut off useful answers. C is incorrect because temperature changes how random the output is, not how many tokens the request consumes. D is incorrect because Provisioned Throughput reserves capacity for steady, high-volume use, so for a few hundred requests a day it would likely cost more. Exam guide: Domain 2 (Fundamentals of GenAI), Task Statement 2.1.

Question 5: Fundamentals of GenAI

A bank is piloting an LLM to draft plain-language explanations of loan decisions. Testers find that the same input sometimes produces differently worded explanations, and one explanation cited a fee policy the bank does not have. Which limitation poses the greater risk here, and what best reduces it?

  1. Nondeterminism; set the temperature to zero so every explanation is worded identically
  2. Hallucination; raise the temperature so that the model explores more policy options
  3. Nondeterminism; fine-tune the model on a larger set of past loan explanations
  4. Hallucination; ground responses in the bank's policy documents and check cited policies

Correct answer: D. D is correct because an invented policy in a customer explanation is a compliance and trust risk, and grounding the model in approved documents while validating what it cites targets that risk directly. A is incorrect because varied wording is low risk when the facts are right, and a temperature of zero does not stop the model from stating false facts. B is incorrect because a higher temperature makes output more varied and less predictable, which tends to increase invented content rather than reduce it. C is incorrect because varied wording is the lesser risk, and fine-tuning on past explanations does not ensure that every policy the model cites exists. Exam guide: Domain 2 (Fundamentals of GenAI), Task Statement 2.2.

Question 6: Applications of Foundation Models

A marketing team uses a foundation model to brainstorm product slogans. The suggestions are repetitive and very similar from one run to the next. Which inference parameter change should the team try first?

  1. Add stop sequences so that the model ends each response at a fixed point
  2. Reduce the maximum output length so the model generates each slogan faster
  3. Decrease the temperature so the model favors its most likely next words
  4. Increase the temperature so the model samples less likely words more often

Correct answer: D. D is correct because a higher temperature flattens the probability distribution over next tokens, which makes output more varied and creative. A is incorrect because stop sequences control where generation ends, not how varied the generated text is. B is incorrect because output length limits how long a response can be; it does not make responses more varied. C is incorrect because a lower temperature makes output more predictable, which would make the repetition worse. Exam guide: Domain 3 (Applications of Foundation Models), Task Statement 3.1.

Question 7: Applications of Foundation Models

An insurer wants an assistant that answers its agents' questions about policy rules. The rule documents are updated every week, and each answer must cite the document it came from. Which approach is the most cost-effective?

  1. Retrieval Augmented Generation with Amazon Bedrock Knowledge Bases over the policy documents
  2. Fine-tune a foundation model on the policy documents and repeat the fine-tuning weekly
  3. Use few-shot prompts that include three sample questions and answers in each request
  4. Run continued pre-training of a foundation model on the full policy library each quarter

Correct answer: A. A is correct because RAG retrieves the current documents at question time, so weekly updates only require re-syncing the data source, and the retrieved passages give the answer a citable source. B is incorrect because weekly fine-tuning is costly and slow to repeat, and a fine-tuned model cannot reliably cite which document an answer came from. C is incorrect because a few examples show the model a format but do not give it the current rules, so it cannot answer accurately or cite a source. D is incorrect because quarterly training leaves answers up to three months out of date, continued pre-training costs far more than RAG, and the model still cannot cite its sources. Exam guide: Domain 3 (Applications of Foundation Models), Task Statement 3.1.

Question 8: Applications of Foundation Models

A team wants a foundation model to sort support emails into five company-specific categories and return only the category label. With a zero-shot prompt, the labels come back inconsistent. What is the lowest-effort fix to try first?

  1. Raise the temperature so the model considers more of the label options
  2. Switch to a larger model that offers a longer context window
  3. Add a few labelled example emails and their categories to the prompt
  4. Fine-tune the model on several thousand labelled support emails

Correct answer: C. C is correct because few-shot prompting shows the model the categories and the exact output format by example, and it needs no training or model change. A is incorrect because a higher temperature makes output more random, which would make the labels less consistent, not more. B is incorrect because the prompt is not running out of space; the model lacks examples of the categories and the expected format. D is incorrect because fine-tuning may work, but it costs far more effort than adding examples to the prompt, which should be tried first. Exam guide: Domain 3 (Applications of Foundation Models), Task Statement 3.2.

Question 9: Guidelines for Responsible AI

A bank's customer chatbot is built on Amazon Bedrock. It must refuse to give investment advice and must remove account numbers from its responses. Which feature addresses both requirements?

  1. Amazon Macie, with scheduled scans of chat logs stored in Amazon S3
  2. Amazon Bedrock Guardrails, with a denied topic and sensitive information filters
  3. Amazon Bedrock Model Evaluation, with a toxicity job run on the chatbot's model
  4. Amazon Bedrock Knowledge Bases, with the bank's policies added as a data source

Correct answer: B. B is correct because Guardrails can block a defined topic such as investment advice and can detect and mask sensitive information such as account numbers in model responses. A is incorrect because Macie finds sensitive data in stored S3 objects after the fact; it cannot filter responses before they reach the customer. C is incorrect because an evaluation job scores a model's responses to test prompts; it does not block topics or mask data in live chatbot responses. D is incorrect because a knowledge base supplies reference content to the model; it does not stop the model from discussing a topic or exposing data. Exam guide: Domain 4 (Guidelines for Responsible AI), Task Statement 4.1.

Question 10: Security, Compliance, and Governance for AI Solutions

Before fine-tuning a model, a retailer must find which files in a large Amazon S3 training-data bucket contain customers' personal information. Which AWS service should the retailer use?

  1. AWS Trusted Advisor
  2. AWS Config
  3. Amazon Inspector
  4. Amazon Macie

Correct answer: D. D is correct because Macie uses machine learning and pattern matching to discover and report sensitive data, such as personal information, stored in Amazon S3. A is incorrect because Trusted Advisor checks account settings against best practices for cost, security and performance; it does not inspect file contents. B is incorrect because Config records and evaluates resource configurations, such as whether a bucket is public, not what data the objects contain. C is incorrect because Inspector scans workloads such as EC2 instances and container images for software vulnerabilities, not S3 data for personal information. Exam guide: Domain 5 (Security, Compliance, and Governance for AI Solutions), Task Statement 5.1.

Practice by AIF-C01 Domain

AWS AI Practitioner Exam at a Glance

The exam uses four question types: multiple choice, multiple response (pick two or more), ordering (put 3–5 steps in order) and matching (pair 3–7 prompts with answers).

Scoring is compensatory: you need 700 overall, not a pass in every domain. There is no penalty for guessing, so answer every question.

AWS publishes exam guide changes about a month before they reach the live exam. Version 1.2 was published on October 29, 2026.

What the Current AIF-C01 Exam Guide Covers

TopicWhat to know
Agentic AI conceptsMulti-agent patterns, Model Context Protocol (MCP) for connecting agents to external systems, agent communication patterns, memory, tool use and workflow orchestration (task 2.1).
Context engineeringDeciding what goes into a model's context window and what stays out, and why it affects cost, latency and answer quality (task 2.1).
AWS agent servicesAmazon Bedrock AgentCore, Strands Agents, Kiro and Amazon Quick, alongside Amazon Bedrock, SageMaker AI and SageMaker JumpStart (tasks 1.3 and 2.3).
Securing agentsAgentCore Identity, Policy in AgentCore and Amazon Bedrock Guardrails, next to IAM, encryption, Amazon Macie and AWS PrivateLink (task 5.1).
Products to stop studyingAWS closed SageMaker Clarify, SageMaker Model Monitor, SageMaker Ground Truth, Amazon Kendra and Amazon Q Business to new customers on July 30, 2026. Our questions test bias, monitoring and search as concepts instead.

5 Agentic AI Practice Questions (MCP, AgentCore, Context Engineering)

Question 1: Fundamentals of GenAI

An ordering agent must check live stock in a warehouse system, and the developers want that integration reusable by other agents built on different models. Orders over a set value must not be placed until a manager agrees. Which combination meets both needs?

  1. Expose the warehouse system through an MCP server, and add a human approval step for large orders
  2. Load a nightly stock export into the prompt, and add a human approval step for large orders
  3. Expose the warehouse system through an MCP server, and lower the agent's temperature for large orders
  4. Fine-tune each model on warehouse records, and give the agent long-term memory of past orders

Correct answer: A. A is correct because an MCP server gives any MCP-capable agent a standard way to call the warehouse tools, and a human-in-the-loop approval step stops large orders until a manager agrees. B is incorrect because approval is right, but a nightly export is not live stock, and pasting it into one agent's prompt is not a reusable integration. C is incorrect because the MCP server meets the reuse need, but a lower temperature only changes randomness and cannot require a manager's agreement. D is incorrect because fine-tuning bakes in stale stock data for each model separately, and remembering past orders does not require approval for new ones. Exam guide: Domain 2 (Fundamentals of GenAI), Task Statement 2.1.

Question 2: Fundamentals of GenAI

A consulting firm's research assistant must produce a market analysis, a legal analysis and a technical analysis of a proposal. None of the three depends on the others, and the combined report is needed as quickly as possible. Which multi-agent pattern fits best?

  1. Run the three agents in sequence, with each one passing its output to the next
  2. Run the three specialist agents in parallel, then have an aggregator merge their outputs
  3. Have one agent write all three analyses in turn within a single long conversation
  4. Have the three agents debate in rounds until they agree on one shared analysis

Correct answer: B. B is correct because because the analyses are independent, running them at the same time and merging the results finishes faster than running them one after another. A is incorrect because a sequential chain suits steps that depend on earlier results; here it only adds waiting time for independent work. C is incorrect because one agent working through all three in turn is slower and mixes three specialties in one context. D is incorrect because the task needs three distinct analyses combined, not agreement on a single view, and debate rounds add delay. Exam guide: Domain 2 (Fundamentals of GenAI), Task Statement 2.1.

Question 3: Fundamentals of GenAI

A travel company's agent runs on Amazon Bedrock AgentCore. Customers complain that each new conversation starts from scratch, so the agent forgets their seat and meal preferences from earlier trips. Which AgentCore capability addresses this?

  1. AgentCore Memory, which stores short-term and long-term context across sessions
  2. AgentCore Code Interpreter, which runs code securely for the agent
  3. AgentCore Observability, which traces and monitors each step of the agent
  4. AgentCore Gateway, which turns existing APIs into tools that agents can call

Correct answer: A. A is correct because AgentCore Memory is a managed capability that keeps conversation context and learned facts such as preferences, so an agent can recall them in later sessions. B is incorrect because a code sandbox lets the agent run calculations or scripts, not remember customers between sessions. C is incorrect because observability shows how the agent behaved, but it does not give the agent recall of past preferences. D is incorrect because Gateway connects agents to tools and APIs; it does not keep what a customer said in earlier sessions. Exam guide: Domain 2 (Fundamentals of GenAI), Task Statement 2.3.

Question 4: Applications of Foundation Models

A travel agency's planning agent holds long conversations. After many turns, the context window fills with old messages and raw tool outputs, and the agent starts forgetting the customer's stated budget. Which context engineering practice best addresses this?

  1. Summarize older turns and keep key facts such as the budget in a compact memory
  2. Raise the maximum output length so the agent can write longer replies
  3. Repeat the full conversation history twice in every request to reinforce it
  4. Raise the temperature so the agent considers a wider range of earlier details

Correct answer: A. A is correct because context engineering decides what enters the context window, and condensing old turns while keeping key facts in memory leaves room for what matters. B is incorrect because the problem is a crowded input context, and longer replies would add even more text to later turns. C is incorrect because duplicating history fills the context window faster and raises token costs, making the forgetting problem worse. D is incorrect because temperature affects how varied the generated text is; it does not decide which earlier details stay in the context. Exam guide: Domain 3 (Applications of Foundation Models), Task Statement 3.1.

Question 5: Security, Compliance, and Governance for AI Solutions

A utility company's billing agent calls a refund tool through Amazon Bedrock AgentCore Gateway. The business rule is that refunds above 500 USD are never issued without a supervisor, whatever the model decides. The team wants this enforced outside the agent's code and prompt. What should it use?

  1. An Amazon Bedrock Guardrails denied topic that blocks any discussion of refunds
  2. Amazon Bedrock AgentCore Identity, with a separate login for the refund tool
  3. A system prompt instruction telling the agent never to refund more than 500 USD
  4. Policy in AgentCore, with a rule that denies refund tool calls above 500 USD

Correct answer: D. D is correct because Policy in AgentCore evaluates each tool call at the gateway against deterministic rules, so the limit holds even if the model is manipulated or misreads its instructions. A is incorrect because a denied topic filters conversation content and would block all refund talk rather than enforcing a limit on tool calls. B is incorrect because Identity controls who the agent is and which credentials it uses, not which tool inputs a business rule permits. C is incorrect because a prompt instruction is guidance the model can ignore or be talked out of, so it is not a deterministic control. Exam guide: Domain 5 (Security, Compliance, and Governance for AI Solutions), Task Statement 5.1.

Questions in Our Bank for Each Exam Guide Task Statement

TaskExam guide task statementDomainQuestions
1.1Explain basic AI concepts and terminologiesFundamentals of AI and ML28
1.2Identify practical use cases for AIFundamentals of AI and ML28
1.3Describe the AI/ML development lifecycleFundamentals of AI and ML22
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
3.1Describe design considerations for applications that use foundation modelsApplications of Foundation Models34
3.2Choose effective prompt engineering techniquesApplications of Foundation Models28
3.3Describe the training and fine-tuning process for foundation modelsApplications of Foundation Models22
3.4Describe methods to evaluate foundation model performanceApplications of Foundation Models25
4.1Explain the development of AI systems that are responsibleGuidelines for Responsible AI33
4.2Recognize the importance of transparent and explainable modelsGuidelines for Responsible AI22
5.1Explain methods to secure AI systemsSecurity, Compliance, and Governance for AI Solutions32
5.2Recognize governance and compliance regulations for AI systemsSecurity, Compliance, and Governance for AI Solutions22

How These Questions Were Built

The bank has 390 questions, six full mock exams' worth, split to follow AWS's weights: 78 for Domain 1, 94 for Domain 2, 109 for Domain 3, 55 for Domain 4 and 54 for Domain 5. Each question is mapped to one of the exam guide's 14 task statements, as the table below shows.

Our platform shows four answer choices with one correct answer. Multiple-response questions become "which combination" choices, and matching and ordering questions show the full pairing or sequence as each choice.

Questions were drafted with AI assistance from the v1.2 exam guide, then fact-checked in a separate review pass against current AWS documentation. Products closed to new customers are left out, and the correct answers are spread evenly across A to D.

Source: AWS Certified AI Practitioner (AIF-C01) exam guide. Review the AWS AI Practitioner cheat sheet.

AWS AI Practitioner Practice Exam FAQ

What are the AWS AI Practitioner (AIF-C01) exam domains?

Exam guide v1.2 has five: Fundamentals of AI and ML (20%), Fundamentals of GenAI (24%), Applications of Foundation Models (28%), Guidelines for Responsible AI (14%) and Security, Compliance, and Governance for AI Solutions (14%). The weights are shares of scored content.

How many questions are on AIF-C01 and what score do I need?

65 questions in 90 minutes. 50 are scored and 15 are unscored questions AWS is trying out, which aren't marked. Results are a scaled score from 100 to 1,000, and 700 passes. You need to pass the exam overall, not each domain.

What question types are on the AWS AI Practitioner exam?

Multiple choice, multiple response (two or more correct answers out of five or more), ordering (put 3–5 steps in order) and matching (pair 3–7 prompts with answers). Unanswered questions count as wrong and there is no penalty for guessing.

What changed in the October 2026 AIF-C01 exam guide?

AWS published version 1.2 on October 29, 2026, and says guide changes reach the live exam about a month later. The current guide asks about agentic AI (MCP, multi-agent patterns, memory and tool use), context engineering, and AWS services such as Amazon Bedrock AgentCore, Strands Agents, Kiro and Amazon Quick. The exam code is still AIF-C01.

Do I need to know SageMaker Clarify or Amazon Kendra for AIF-C01?

AWS closed SageMaker Clarify, SageMaker Model Monitor, SageMaker Ground Truth, Amazon Kendra and Amazon Q Business to new customers on July 30, 2026, and the current exam guide doesn't list them. Learn what they did as concepts, such as checking a model for bias across groups, rather than the product names.

Do I need coding experience for AWS AI Practitioner?

No. AWS's guide says the target candidate uses AI on AWS but doesn't necessarily build it, and lists coding models, tuning hyperparameters and building pipelines as out of scope. You should know core AWS services, IAM, the shared responsibility model and AWS pricing models.

AWS AI Practitioner or AWS Cloud Practitioner first?

Cloud Practitioner (CLF-C02) covers AWS cloud basics: services, pricing, security and support. AI Practitioner (AIF-C01) is about AI, ML and generative AI concepts and the AWS services for them. If AWS itself is new to you, Cloud Practitioner first makes AIF-C01 easier; if you know AWS basics, start with AIF-C01.

Are these real AWS AI Practitioner exam questions?

No. Real AWS exam questions are confidential. Every question here is original, written to AWS's published exam guide v1.2, with an explanation for every answer and a tag for the guide task statement it tests.

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.