Signing in
5 practice tests, a 65-question timed simulation, and a free assessment.
AWS Certified AI Practitioner
Find your AWS AIF-C01 gaps in 15 minutes.
Learner actions will appear after the current account is confirmed.
Take a free assessment, see the decisions costing you points, or confirm where to validate breadth next instead of browsing a generic study list.
Know where you stand before another full practice test.
Targeted repair: Fix a focused gap around choosing prompt engineering, RAG, fine-tuning, and agent workflows.
If you miss it, CramHQ identifies the decision you misunderstood, helps you fix it, then checks whether you can apply it in a different scenario.
A team prepares company manuals so a generative application can find passages related to a user's question before forming an answer. Which component is used to organize the numeric representations so similar passages can be found efficiently?
This is the best fit because it organizes numeric representations so semantically close passages can be retrieved for grounding the response.
Ten practice questions from the CramHQ AWS AIF-C01 bank, with answers and explanations. No signup needed. When you are ready, start the free assessment. CramHQ ranks your weak spots across the whole exam by the points they cost, and your Pass Plan starts with the costliest.
A team reviews past ride requests to estimate whether a new request is likely to be canceled. The result is a score used by an existing dispatch workflow, and no new content is composed for the user. Which concept best describes this capability?
Answer: C
This fits because historical examples are used to estimate the likelihood of a future outcome, producing a score for another workflow.
“no new content is composed for the user”
This rules out generative AI content creation because the scenario explicitly denies composed output.
Use this as an elimination cue. Remove content-generation answers unless the stem requires original artifacts.
Why it is tempting: Generative AI is a prominent AI category, so it can look plausible whenever a scenario mentions an intelligent capability.
Why it is wrong: Generative AI is the right label when the system creates original content such as text, images, audio, code, or other artifacts.
Why it is tempting: Ride requests can contain text, so natural language processing may seem relevant if you focus on possible input format instead of the stated task.
Why it is wrong: Natural language processing fits when the main requirement is to understand, extract from, classify, translate, or transform human language.
Why it is tempting: The scenario mentions dispatch workflow, which can make orchestration sound relevant if you treat workflow involvement as autonomy.
Why it is wrong: Agentic AI task orchestration fits when an agent plans steps, invokes tools or APIs, and coordinates actions to complete a goal with autonomy.
A fraud-scoring model has been producing scores for several months. Recent reviewed cases show its scores no longer separate suspicious and normal activity as well as before. The team wants to choose the next lifecycle action. What should they do?
Answer: A
This uses recent verified cases to check whether the current model still meets expected performance and updates it only when the evidence supports that decision.
“Recent reviewed cases show its scores no longer separate suspicious and normal activity as well as before.”
This quote establishes current production feedback showing degraded model behavior, so the answer should use recent verified outcomes to evaluate performance before deciding on retraining.
Treat this as a feedback-loop cue. Compare options by whether they use current reviewed outcomes instead of old tests, synthetic substitutes, or more inference volume.
Why it is tempting: Synthetic examples can look controlled and easier to judge, especially when reviewed cases are messy.
Why it is wrong: When production performance changes, use trusted recent real-world outcomes to assess the running model's current quality.
Why it is tempting: A passed test set can feel authoritative because it supported the original deployment decision.
Why it is wrong: Validation at launch does not replace ongoing monitoring and evaluation against recent outcomes after the model is in use.
Why it is tempting: More requests can sound like more opportunities for correct scores, but each request still uses the same current model.
Why it is wrong: If model quality has degraded, the lifecycle action is to measure performance with recent outcomes and update the model when evidence supports it.
A review team needs an AI capability that can read a short incident note and compare it with a related scene photo before recommending whether the note is consistent with the evidence. Which model capability should the team choose?
Answer: A
This matches the need to compare the note with the photo as one evidence-checking task, rather than processing only one input type.
“read a short incident note and compare it with a related scene photo”
This quote states that the same task must use both written text and visual evidence, which activates multimodal capability selection.
Separate the input types first, then eliminate choices that handle only text, only audio, or content generation.
Why it is tempting: Summarizing and classifying the note sounds useful because the note is part of the evidence.
Why it is wrong: A text-focused capability is insufficient when the required judgment depends on visual evidence as well as text.
Why it is tempting: Witness statements and incident reviews can involve speech, so transcription can feel related to evidence handling.
Why it is wrong: Match the model capability to the actual inputs in the scenario, not to a different possible evidence source.
Why it is tempting: An image-generation model mentions the scene and the note, so it can look connected to both inputs.
Why it is wrong: Use generation capabilities when the task is to create new content; use understanding capabilities when the task is to evaluate existing evidence.
An analyst sends a long instruction plus a full set of reference pages to a foundation model. The request fails before meaningful output starts because the submitted text is too large for one call. What should the analyst choose first?
Answer: D
This addresses the documented per-interaction size boundary by reducing how much the model must consider at once while preserving the task intent.
“submitted text is too large for one call”
This quote identifies a per-call context-window constraint, so the fix must reduce or divide the submitted material rather than change generation behavior.
Treat this as a context-size cue. First eliminate options that change model behavior or settings, then choose the option that reduces what is sent at once.
Why it is tempting: Randomness settings can affect how a model chooses among possible outputs, so they can look like a way to make the model be selective.
Why it is wrong: Randomness settings do not increase the context window; oversized submitted text must be reduced, summarized, or split.
Why it is tempting: Retraining sounds powerful because it can change model behavior over time, so it can look like a way to make the model handle more reference material.
Why it is wrong: Model customization affects learned behavior or task adaptation; it does not remove the fixed amount of text a model can process in one call.
Why it is tempting: Instructions can guide what the model pays attention to when the request is accepted, so this can seem like a smart prioritization strategy.
Why it is wrong: System instructions consume context too; they can guide behavior but cannot expand the per-call context window.
An analyst asks a generative AI assistant whether an internal policy allows sharing a quarterly market report. The answer varies and cites assumptions; the policy document is not part of the request. Which prompt change best addresses the issue?
Answer: D
This supplies the missing policy details and directs the assistant to use them, reducing guesswork when the answer depends on private or provided information.
“the policy document is not part of the request”
Treat this as a context-supply cue. First eliminate changes that only affect tone or layout, then choose the option that gives the model the needed source material.
Why it is wrong: Use examples to shape style or pattern; use source context when the model needs facts from a specific document.
Why it is tempting: Role framing can focus the assistant's tone and perspective, especially for compliance-related tasks.
Why it is wrong: Use role framing to set perspective; provide source context when the answer depends on private or task-specific facts.
Why it is tempting: A table can make allowed and not-allowed outcomes easier to compare and can make the answer look more rigorous.
Why it is wrong: Output constraints control how the answer is presented; source context supplies the facts the model must reason from.
An application receives a user-submitted support transcript and must return a short case summary with a severity label. The needed details are all inside the transcript, and no company article or changing reference set must be searched. Which approach should be selected?
Answer: B
This fits because the model can summarize and label the content from the supplied input. No separate source collection is needed to provide missing facts.
“The needed details are all inside the transcript”
This tells you the model does not need outside factual grounding; it only needs to transform the provided input.
Why it is tempting: Knowledge Bases sound appropriate for generated answers that should be grounded in company information.
Why it is wrong: Use RAG when the model must retrieve external enterprise or domain content that is not already in the request.
Why it is tempting: Embeddings and vector stores are often mentioned with knowledge retrieval for generative AI applications.
Why it is wrong: This treats a single supplied input as if it were a reusable searchable corpus. Embeddings help compare and retrieve stored content, not replace direct use of provided text.
Why it is tempting: Labeled support transcripts sound relevant because the output includes a summary and severity label.
Why it is wrong: Use model customization when behavior or task style must be adapted through examples, not when the needed case facts are already supplied.
A team is choosing a model to support eligibility recommendations that reviewers must justify to customers. One candidate has slightly higher test accuracy but produces reasoning that is hard to trace. Which choice best reflects the responsible AI consideration?
Answer: C
This recognizes that human review and customer justification can make an inspectable decision basis more important than a small score advantage.
“reviewers must justify to customers”
A justification requirement activates explainability and transparency, so the best answer should support human review and defensible decisions.
Treat this as a reviewability cue. First eliminate options that optimize only accuracy, complexity, or speed, then compare which choice supports explanation.
Why it is tempting: Complex models can capture subtle patterns, so they may sound more reliable for sensitive workflows.
Why it is wrong: When decisions need justification, opaque or hard-to-trace model behavior is a risk even if the model seems powerful.
Why it is tempting: Higher test accuracy is an important evaluation metric and can look like the safest objective choice.
Why it is wrong: Accuracy is only one selection factor; reviewable decisions also need outputs that people can inspect, explain, and defend.
Why it is tempting: Lower latency improves user experience and can matter in production model selection.
Why it is wrong: Speed does not replace explainability when reviewers must understand and defend the basis for recommendations.
A team notices an AI feature gives less reliable recommendations for some customer groups than others. They want to choose the most appropriate responsible AI control before wider release. Which action should they take?
Answer: B
This directly checks whether the feature performs differently for groups that may be affected differently, using evidence tied to the expected user population.
“less reliable recommendations for some customer groups than others”
This phrase makes the issue about uneven model quality across groups, so the best control should measure performance separately for those groups.
Treat this as a subgroup-evaluation cue. Compare options by whether they produce evidence about group-specific model quality.
Why it is tempting: Guardrails are a common responsible AI control and can reduce harmful interactions.
Why it is wrong: Use content guardrails for unsafe or policy-violating inputs and outputs; use subgroup evaluation when the risk is unequal model quality across groups.
Why it is tempting: Least privilege is an important AWS governance practice and is often required around production systems.
Why it is wrong: Use IAM permissions to control access; use model evaluation to assess output quality and fairness across affected populations.
Why it is tempting: Model cards are useful for communicating intended use, limitations, and evaluation results.
Why it is wrong: Document known behavior after evaluation; when the concern is uneven reliability, first gather or verify subgroup performance evidence.
A team keeps AI interaction records for audit and improvement. Compliance asks them to control how long records remain usable, what happens as they age, and whether viewer permissions are still appropriate. Which recommendation best addresses this need?
Answer: D
This matches the documented governance need to manage records across their usable life and confirm that access remains appropriate as business and compliance conditions change.
“control how long records remain usable, what happens as they age”
This quote activates a lifecycle-governance rule: the best answer must manage retention or aging-based handling of AI records, not only protect or relocate them.
Scope the answer to lifecycle behavior first, then eliminate controls that only change privacy, location, or encryption posture.
Why it is tempting: De-identification is a real privacy safeguard for later use of AI records, so it feels governance-related.
Why it is wrong: Use de-identification for privacy-risk reduction; use lifecycle rules and access reviews for retention, aging, disposition, and continued permission validation.
Why it is tempting: Approved Regions and replication restrictions are familiar compliance controls for where data is stored and processed.
Why it is wrong: Use Region and replication controls for residency requirements; use lifecycle handling and access reviews when the requirement is record aging, retention, disposition, and continued access need.
Why it is tempting: Encryption with managed keys and rotation is a strong security practice for stored records.
Why it is wrong: Use encryption and key rotation to protect stored data; use lifecycle policies and access reviews to govern availability over time and continued permission justification.
A research team is reviewing controls for an AI workload that handles sensitive input records. A policy owner asks which requirement addresses the geographic boundary for the workload's information handling, rather than safeguards or lifecycle rules. Which requirement should the team identify?
Answer: B
This addresses the geographic boundary for handling workload information. Region selection is the documented AWS control point for keeping service resources and data handling within a selected location.
“geographic boundary”
This quote activates the data residency decision rule: identify the requirement that constrains where information is stored or processed.
Scope the question to location first, then eliminate answers that only address retention, encryption, or access permissions.
Why it is tempting: S3 Lifecycle rules are real governance controls for data disposal, so they can look relevant when sensitive AI outputs are mentioned.
Why it is wrong: Use lifecycle rules for retention, archival, and deletion timing; use residency requirements to define the geographic handling boundary.
Why it is tempting: KMS encryption is a strong safeguard for sensitive input records and generated outputs, so it feels aligned with security governance.
Why it is wrong: Encryption protects data confidentiality; data residency defines the permitted geographic location for storage or processing.
Why it is tempting: IAM permissions are essential for least privilege and sensitive-data access, so they can seem like the best governance control.
Why it is wrong: IAM policies control principals and permissions; residency requirements control the geographic boundary for data storage or processing.
CramHQ links wrong answers to the mistake behind them. Start with the free assessment, and CramHQ keeps finding your gaps as you practice and adds a repair for each to your Pass Plan.
Why study with CramHQ
Start with the free AWS AI Practitioner assessment. CramHQ uses your result to rank the next step, explain the decisions that matter, and check you again.
Learner actions will appear after the current account is confirmed.
The Pass Plan ranks gaps by the points they cost and opens the top repair.
The reasoning behind the right answer and where the alternatives break.
A fresh question on the same decision verifies the gap is closed.
Strengthened areas change colour; pick the next priority from what is left.
Coverage follows the published AWS AIF-C01 skill areas measured as of Apr 30, 2026, so your practice stays aligned.
Changes recorded from official AWS exam sources.
Compared with Exam guide version 1.0
CramHQ completed the AIF-C01 update for AWS exam guide version 1.1, including the expanded agentic AI and service scope.
Use separate exam-style practice quizzes to validate breadth, review explanations and your answers, use hints when stuck, and build confidence before the timed simulation, with an optional timer.
Use this after practice when you want an exam-style pacing, stamina, and breadth check.
Start with the free assessment; full Pass Plan access unlocks the timed exam simulation.
CramHQ shows the interaction patterns available for AWS AIF-C01 from the active exam path inventory.
Choose the best answer from exam-style options.
Select every correct option when more than one applies.
CramHQ helps you practice the decisions behind AWS AIF-C01 without using copied or confidential exam content.
Original practice questions aligned to public exam objectives.
Coverage is checked against the published exam objectives.
CramHQ is independent and is not affiliated with AWS.
No. You can start the assessment and review your report before creating an account. Sign up when you want to save the report and continue the Pass Plan.
Yes. The assessment, readiness report, and your first targeted repair are free.
No. CramHQ uses original practice items designed to test the decisions and concepts behind the exam.
No. CramHQ is independent and is not endorsed by AWS.
Save your report to keep the Pass Plan, fix your first gap free, and unlock the full Pass Plan when you want every repair, practice quizzes, and timed simulation.
That is exactly who it is built for. CramHQ keeps tracking the last few gaps between you and a comfortable pass, the ones a raw score hides, and puts the one costing you the most points at the top of your Pass Plan.
12 months of full access from purchase, with room to come back after a break.
Paid exam paths come with a 7-day money-back guarantee. Conditions apply; see the Refund Policy. You can also try the assessment and your first repair free before you buy. Refund Policy
Cloud Practitioner readiness is a strong starting point before deeper AWS architecture practice.
Solutions Architect prep pairs service-selection judgment with secure, resilient, high-performing, and cost-optimized architecture practice.
