Free AWS AIF-C01 Practice Exam: AI Practitioner

Try this 65-question AWS AIF-C01 practice exam from the current IT Mastery bank. It includes 45 single-answer questions and 20 Select TWO questions, with five workflow diagrams, evidence tables and explanations for every answer.

These are original IT Mastery practice questions, not official AWS questions, copied live-exam content or exam dumps. Mastery Exam Prep is independent of AWS.

How to take this practice exam

• Set aside 90 minutes if you want a timed attempt. Record your answers before revealing the explanations.
• Choose one answer unless the question says Select TWO. For practice scoring, award one point only when the complete correct set is selected.
• Treat NOT or INCORRECT as part of the task. Follow the stated rule even when an alternative sounds reasonable in another situation.
• For diagrams, follow the arrows and decision branches. Use the text description or full-size view when helpful.
• Afterward, review missed questions and correct guesses by domain, then practise those topics in the app.

AWS’s guide lists 65 items, including 15 unscored items, and possible multiple-choice, multiple-response, ordering and matching formats. This page matches the total length but uses an editorial mix of single-answer and Select TWO exercises. It does not reproduce ordering or matching interfaces, hidden unscored items or AWS scaled scoring. A raw score out of 65 cannot establish that you will pass.

Practice-set coverage

Practice questions

Questions 1-25

Question 1

Topic: Security, compliance and governance

An insurer is rolling out an approved generative AI assistant for claims communications. Public AI services are prohibited, and DLP, least-privilege access, and review controls will remain.

Operating model:

• Agents may send routine drafts only after checking coverage and amounts against the claim record.
• Supervisors review flagged high-impact drafts and lead incident investigation and escalation.

A pilot found agents using public AI, exposing claim details, sending unchecked drafts, and not reporting inaccurate output promptly.

Which training and acceptable-use approach best addresses these findings?

Options:

• A. Teach agents data handling, draft verification, and prompt reporting for any AI service; teach supervisors review, investigation, and escalation duties.

• B. Teach agents approved-tool and data rules, draft verification, and prompt reporting; teach supervisors review, investigation, and escalation duties.

• C. Teach agents approved-tool and data rules, draft verification, and quarterly reporting; teach supervisors review, investigation, and escalation duties.

• D. Teach agents approved-tool and data rules plus prompt reporting; assign draft verification, investigation, and escalation entirely to supervisors.

Best answer: B

Explanation: Effective AI governance training should reflect each role’s authority and actual workflow. Agents need to know which tools are approved, how claim information may be handled, how to verify generated coverage and payment statements, and when to report an incident. Supervisors need training for high-impact review, investigation, and escalation. These expectations directly address the pilot failures while preserving DLP, access restrictions, and review controls.

Training reinforces acceptable behavior but does not replace enforceable controls. Likewise, knowing how to protect data does not make a prohibited public AI service acceptable. The key is aligning tool boundaries, verification, and reporting duties with each employee’s responsibilities.

Why each option fits or fails:

A. Information-handling training does not authorize public AI services when the insurer has explicitly prohibited their use.

B. This approach addresses each observed failure, assigns responsibilities according to the operating model, and complements the existing technical controls.

C. Quarterly reporting delays investigation of inaccurate outputs or sensitive-data exposure when prompt incident reporting is needed.

D. Agents send routine communications themselves, so transferring all verification to supervisors conflicts with the required workflow and leaves agents insufficiently accountable.

Question 2

Topic: Foundation-model applications

A support application must use approved policy updates within one day, preserve a tone that changes only quarterly, and avoid model retraining for each document update. The team uses the workflow below, but handoff H is not implemented and answers remain stale.

Which operating change should fill handoff H?

Options:

• A. Insert all approved changes as few-shot demonstrations in every request.

• B. Ingest approved changes and refresh the vector index within one day.

• C. Fine-tune the model on approved changes and redeploy it weekly.

• D. Continue pre-training on approved changes and redeploy the model weekly.

Best answer: B

Explanation: RAG separates frequently changing knowledge from model parameters. Handoff H should ingest approved documents, split and embed their content, and update the vector index so each change becomes retrievable within one day of approval. The owners publish policy changes weekly, but that publication frequency is separate from the permitted update delay. Fine-tuning can serve the comparatively stable tone requirement on its quarterly change cycle. Fine-tuning or continued pre-training for each policy change would add recurring training and deployment work and would not implement the missing retrieval-index handoff. Supplying all changes as few-shot demonstrations instead increases context demands and does not complete the displayed ingestion path. Match each component’s update process to both the purpose of its data and the stated freshness requirement.

Why each option fits or fails:

A. Few-shot demonstrations guide response behavior but are an inefficient knowledge-update mechanism and increase recurring input-token costs.

B. Completing ingestion and indexing within one day makes each approved change retrievable on time without changing the model’s parameters.

C. Weekly fine-tuning adds recurring training, evaluation, and deployment work while leaving the depicted retrieval index unchanged.

D. Continued pre-training is costly customization intended for broader domain adaptation, not the missing source-to-index update handoff.

Question 3

Topic: Foundation-model applications

A customer-support team is adapting a foundation model. It needs to:

• answer recurring refund requests in an approved format and tone
• cite current refund rules, which change weekly

The team has thousands of reviewed request-response pairs demonstrating the desired behavior. Current policy documents are maintained in an authorized knowledge repository.

Which TWO conclusions are supported by this situation? Select TWO.

Options:

• A. Instruction tuning automatically stays current when source policy documents change.

• B. Instruction tuning can reinforce the required response format and tone.

• C. Retrieval at inference can supply policy details that change between updates.

• D. Retrieval permanently encodes approved response behavior into the model parameters.

• E. Continued pretraining on raw policies best teaches the required format and tone.

Correct answers: B and C

Explanation: Instruction tuning uses suitable request-response pairs to update a model’s parameters toward a consistent task, format, or tone. The reviewed examples therefore fit the team’s stable behavioral requirement. Weekly policy changes are a different problem: encoding them through repeated tuning would be slow and could quickly become outdated. Retrieval-augmented generation can instead obtain authorized, current policy content at inference time, assuming the repository is properly updated and searched. These techniques are complementary: instruction tuning shapes how the model responds, while retrieval supplies changing facts. Continued pretraining on raw documents is less direct for teaching a specific response pattern, and retrieval does not itself alter model parameters.

Why each option fits or fails:

A. Instruction tuning changes model parameters during adaptation; later document changes do not automatically update those learned parameters.

B. The reviewed request-response pairs directly demonstrate the consistent behavior that instruction tuning should incorporate into the model’s parameters.

C. Retrieving from the maintained policy repository provides current facts without requiring repeated parameter updates whenever refund rules change.

D. Retrieval supplies context during inference and does not permanently modify the model’s parameters or teach lasting response behavior.

E. Raw policy text provides domain content but does not demonstrate the request-to-response behavior as directly as paired examples do.

Question 4

Topic: Responsible AI

An AI support agent receives customer text, drafts responses using order records, and can call a refund API. The application must screen disallowed input, detect unsupported or malformed responses before display, and require manager approval for refunds over $500. Which layered safeguard design correctly assigns a control to each failure path?

Options:

• A. Apply input content checks, validate output against records and schema, and enforce manager approval at the refund API.

• B. Apply input content checks, validate output against records and schema, and state the approval limit in the agent instructions.

• C. Authenticate customers before processing, validate output against records and schema, and enforce manager approval at the refund API.

• D. Apply input content checks, filter generated content for safety, and enforce manager approval at the refund API.

Best answer: A

Explanation: Layered safeguards provide defense in depth by controlling different stages of an AI workflow. Input checks assess content before it reaches the model. Output validation checks generated claims against trusted records and confirms that required structure is present. Action controls enforce authorization or human approval where the refund is actually executed. These controls are not interchangeable: authentication does not inspect input content, safety filtering does not establish factual accuracy, and natural-language instructions do not enforce business permissions.

Consequential actions should be governed at the tool or resource boundary, not merely requested in the model’s instructions.

Why each option fits or fails:

A. Each safeguard operates at the relevant boundary: incoming content, generated claims and structure, and the consequential action.

B. Agent instructions can guide behavior, but they do not enforce the required approval at the API or resource boundary.

C. Authentication establishes identity, but it does not screen submitted text for disallowed or malicious content.

D. A safety filter can detect configured content risks, but it does not verify factual support or required response structure.

Question 5

Topic: AI and ML fundamentals

A fraud model flags 200 transactions for manual review. Reviewers confirm fraud in 150 and find 50 legitimate. Each review costs $4. No information is available about fraudulent transactions that were not flagged.

Which interpretation is correct?

Options:

• A. Recall is 75%; 50 unnecessary reviews cost $200; missed fraud remains unknown.

• B. Precision is 75%; 200 unnecessary reviews cost $800; missed fraud remains unknown.

• C. Precision is 75%; 50 unnecessary reviews cost $200; missed fraud remains unknown.

• D. Precision is 25%; 150 unnecessary reviews cost $600; missed fraud remains unknown.

Best answer: C

Explanation: Precision measures how many positive predictions are confirmed: 150 of 200 flagged transactions are fraud, so precision is 75%. The remaining 50 are false positives that consume review capacity without confirming fraud. At $4 each, those unnecessary reviews cost $200. Precision does not indicate how many fraudulent transactions the model failed to flag because the total number of actual fraud cases, including false negatives, is not provided. That missing information is required to calculate recall. The key distinction is that precision describes the quality and workload efficiency of positive predictions, not missed-positive coverage.

Why each option fits or fails:

A. The 75% calculation uses all flagged cases as its denominator, which measures precision rather than recall.

B. The $800 amount covers every review, but only the 50 legitimate transactions represent unnecessary reviews caused by false-positive predictions.

C. Precision is 150 confirmed positives divided by 200 flagged cases, while the 50 false positives create $200 in unnecessary review cost.

D. The 25% value is the false-positive share, and the 150 confirmed fraud cases are useful positive predictions rather than unnecessary reviews.

Question 6

Topic: AI and ML fundamentals

A contact center uses a managed AI API to summarize call recordings.

Pilot and workflow evidence:

Which TWO customer actions are supported by this evidence? Select TWO.

Options:

• A. Validate required deadline preservation before summaries enter operational use.

• B. Approve operational use based on the pilot’s 96% fluency result.

• C. Delegate recording-region authorization decisions to the managed API provider.

• D. Schedule operating-system patching for the API’s underlying serving hosts.

• E. Restrict source access to recordings each requesting analyst may process.

Correct answers: A and E

Explanation: A managed AI API transfers operation of the service infrastructure, such as host patching and scaling, to the provider. The customer still controls the surrounding business workflow, including which source records the application may access and whether outputs are acceptable for their intended use. Here, broad recording access conflicts with the assigned-region policy, so the customer must enforce the proper access boundary. The representative pilot also shows that fluent summaries can omit required deadlines. The customer must therefore validate this business-critical requirement and introduce suitable review or handling before summaries become customer notices. Managed operation does not make the provider responsible for customer data selection or output fitness.

Why each option fits or fails:

A. The pilot found deadline omissions in 28 representative calls, requiring customer evaluation and an appropriate review or validation control.

B. Fluency does not establish that summaries preserve required deadlines, which directly affects the intended workflow.

C. The customer defines which recordings its analysts may process; consuming a managed API does not transfer that business authorization decision.

D. The evidence assigns host patching and scaling to the managed API provider, not the customer.

E. The application can access recordings beyond each analyst’s permitted region, so the customer must enforce the business access boundary.

Question 7

Topic: Responsible AI

A company will generate 1,000,000 summaries each month. Required quality is >= 90, and required p95 latency is <= 2.0 seconds. Tests used representative requests and planned production configurations. Energy includes serving overhead; training and hardware-manufacturing impacts were not measured.

Which TWO conclusions does the evidence support? Select TWO.

Options:

• A. Model A has lower total impact than Model B because it reuses hardware.

• B. The evidence cannot establish the lowest total lifecycle impact among the models.

• C. Model B uses the least measured inference energy for the qualifying workload.

• D. Model B has the lowest total impact because its inference energy is lowest.

• E. Model C’s higher quality will reduce retries enough to offset its energy use.

Correct answers: B and C

Explanation: Environmental comparisons require equivalent workload, quality, latency, and measurement boundaries. Each model satisfies the business thresholds, and the energy figures cover the same monthly volume and include serving overhead. Model B therefore has the lowest measured operational inference energy at 540 kWh per month. However, operational energy is only one part of lifecycle impact. Model B requires new hardware, while Models A and C reuse existing accelerators, and no manufacturing or training impacts were quantified. Consequently, the evidence supports an inference-energy ranking but not an overall lifecycle-impact ranking. Higher quality beyond the requirement also does not establish environmental savings unless reduced retries or other resource effects are measured.

Why each option fits or fails:

A. Hardware reuse is relevant, but its benefit was not quantified against Model B’s lower operating energy.

B. Training and hardware-manufacturing impacts are unmeasured, so inference energy and capacity plans alone cannot determine total lifecycle impact.

C. All three models meet the stated thresholds, and Model B consumes the least inference energy at the common monthly workload.

D. Lower measured inference energy does not account for the environmental impacts of training or acquiring the new accelerator.

E. The quality results do not provide retry rates or evidence that avoided retries would offset Model C’s greater inference energy.

Question 8

Topic: Foundation-model applications

A company evaluates foundation models using the same representative tasks and expected traffic. Requirements are:

• Task success rate >= 88%
• p95 latency <= 800 ms
• Available in the approved deployment Region
• Lowest monthly cost among models meeting all requirements

Which foundation model should the company select?

Options:

• A. Select Model Atlas.

• B. Select Model Cedar.

• C. Select Model Birch.

• D. Select Model Dune.

Best answer: B

Explanation: Foundation model selection first applies mandatory requirements to determine feasibility. Birch is excluded because its p95 latency exceeds 800 ms, while Dune is excluded because it is unavailable in the approved Region. Atlas and Cedar both satisfy the minimum task success rate, latency limit, and deployment requirement. The stated business priority then resolves the trade-off: Cedar costs $9,000 per month compared with Atlas at $12,000 per month. Atlas’s higher quality score does not justify selection because the requirement asks for the lowest-cost model after all mandatory thresholds are met. Hard constraints should be applied before optimizing a preferred metric such as cost.

Why each option fits or fails:

A. Atlas meets every mandatory requirement, but its monthly cost exceeds Cedar’s, so it is not preferred under the stated cost priority.

B. Cedar satisfies the quality, latency, and deployment requirements and has the lowest cost among the feasible models.

C. Birch has the highest task success rate and lower cost, but its p95 latency exceeds the mandatory 800 ms limit.

D. Dune has the lowest cost and acceptable performance, but it is unavailable in the required approved deployment Region.

Question 9

Topic: Foundation-model applications

A customer asks an AI assistant to cancel an order. The following trace is captured before the assistant responds:

The assistant drafts:

Your order has been canceled.

Which response approach should replace the draft?

Options:

• A. Confirm cancellation because the unshipped status proves the order record changed.

• B. Explain eligibility, then use an authorized action path and verify its result.

• C. Confirm cancellation because policy retrieval establishes authority for the business action.

• D. Retrieve the policy again, then treat stronger grounding as permission to cancel.

Best answer: B

Explanation: Grounding and action authority are separate capabilities. Retrieval lets the assistant read the cancellation policy and current order status, so it can explain that the order appears eligible. However, the denied permission means no authorized transaction occurred, and the missing receipt confirms there is no evidence of cancellation. The assistant should send the request through an authorized tool or workflow and confirm success only from the system of record. Better retrieval cannot grant permissions or change business data, while even an authorized action must return a successful outcome before the assistant reports completion.

Why each option fits or fails:

A. The unshipped status shows eligibility before cancellation; it does not prove that the requested change occurred.

B. The trace supports an eligibility explanation, but the denied permission and missing receipt require authorized execution and returned outcome evidence.

C. Retrieved policy content explains cancellation rules, but it does not grant permission to modify an order.

D. Additional retrieval may improve the explanation, but knowledge context cannot provide action authorization or transaction evidence.

Question 10

Topic: Foundation-model applications

A support assistant pilot used representative, equally weighted English and Spanish cases.

Mandatory release gates:

• Grounded-answer rate >=92% in each language
• Zero unconfirmed refund executions

Tradeable objectives: Maximize task completion; target p95 latency <=3.0 seconds and cost <=$0.25 per case.

Operations column: p95 latency / cost per case / unconfirmed refunds

Which proposal does NOT satisfy the release policy?

Options:

• A. Disable refund execution, validate the safeguard, then release safeguarded.

• B. Keep compact-fast in pilot and improve Spanish grounding before release.

• C. Release retrieval-plus, accepting its cost and latency tradeoffs.

• D. Release compact-fast gradually while monitoring Spanish grounding after deployment.

Best answer: D

Explanation: Mandatory release gates must be satisfied before deployment; favorable business outcomes or operational metrics cannot compensate for a failed gate. Compact-fast has the highest task completion and meets the latency and cost targets, but its 88% Spanish grounded-answer rate fails the required subgroup threshold. Monitoring after a limited release does not replace a pre-release requirement.

Retrieval-plus may trade higher latency and cost because those measures are objectives rather than gates. Safeguarded can be considered after refund execution is disabled and the control is validated, because its grounded-answer rates already pass. The key distinction is between requirements that block release and objectives that can be balanced against user impact and business value.

Why each option fits or fails:

A. Both language-quality rates pass, and disabling refund execution removes the consequential action path while the control is validated before release.

B. Holding compact-fast prevents release below the mandatory Spanish grounded-answer threshold while preserving its favorable cost, latency, and completion results for later evaluation.

C. Retrieval-plus passes both language-quality gates and has no unconfirmed refunds; its cost and latency misses concern explicitly tradeable objectives.

D. A gradual rollout still releases compact-fast when its 88% Spanish grounded-answer rate is below the mandatory 92% threshold.

Question 11

Topic: Generative AI fundamentals

A team wants each layer of an agentic application to remain independently replaceable.

Architecture review:

Which technology best fits the agent-logic layer?

Options:

• A. Use SageMaker AI as the developer framework.

• B. Use AgentCore Runtime as the developer framework.

• C. Use Agents for Amazon Bedrock as the developer framework.

• D. Use Strands Agents as the developer framework.

Best answer: D

Explanation: Strands Agents fits the developer-owned agent-logic layer. It is an open-source SDK used to define agent behavior, connect models, and expose tools. These responsibilities are separate from foundation-model inference and application operations. Amazon Bedrock can provide managed access to foundation models, while Amazon Bedrock AgentCore can supply modular operational capabilities such as runtime, identity, memory, tool connectivity, and observability. Selecting Strands does not require those other layers to use a particular model or hosting environment. The key distinction is framework code versus model inference and managed infrastructure.

Why each option fits or fails:

A. SageMaker AI supports managed ML development and operations rather than serving as the specified open-source agent SDK.

B. AgentCore Runtime operates agent applications but does not define their behavior and tools as the required developer framework.

C. Agents for Amazon Bedrock provides managed agent orchestration, but it does not fulfill the requirement for an independently hosted open-source SDK.

D. Strands Agents is an open-source SDK for defining agent behavior and tools while keeping model and hosting choices separable.

Question 12

Topic: AI and ML fundamentals

A neural network is trained using labeled images of defective and nondefective products. Which statement best explains how the network learns useful representations for classification?

Options:

• A. Layers apply a complete set of visual rules explicitly written by engineers.

• B. Layers adjust learned weights from errors and progressively combine image patterns.

• C. Layers store every training image and compare new images with exact copies.

• D. Layers group similar pixels without labels and assign each group a defect category.

Best answer: B

Explanation: A neural network learns representations by adjusting weights during training to reduce prediction errors on examples. For image classification, early layers may respond to simple visual patterns such as edges or textures. Later layers combine those patterns into representations associated with product features and defects. The model is not given an exhaustive list of business rules, nor does it need to preserve every training image as an exact template. During inference, the trained layers apply these learned transformations to a new image and produce a classification. The key distinction is that useful patterns emerge from weight updates guided by examples and training feedback.

Why each option fits or fails:

A. Traditional rule-based systems use explicitly written conditions, whereas neural networks learn weighted patterns from training examples.

B. Training feedback adjusts weights so successive layers can transform pixels into patterns useful for predicting the provided labels.

C. A neural network encodes recurring patterns in learned weights rather than relying on exact copies of all training examples.

D. Grouping inputs without labels describes clustering, while supervised neural network training uses labeled errors to adjust weights.

Question 13

Topic: Responsible AI

A public benefits agency pilots an AI assistant intended for English- and Spanish-speaking applicants, including people using screen readers over intermittent mobile connections. A sufficiently sized, representative pilot finds that every group can open the assistant.

Successful application completion:

Which action is most supported before broad rollout?

Options:

• A. Aggregate all pilot groups, compare overall completion, and decide rollout.

• B. Investigate mobile performance, optimize it, and retest combined-context completion.

• C. Investigate intersecting access barriers, remediate them, and retest subgroup completion.

• D. Investigate Spanish localization, revise it, and retest Spanish-language completion.

Best answer: C

Explanation: Inclusive design is measured by whether intended users can successfully use a service in their real interaction contexts, not merely whether they can access it. The pilot shows nominal availability for every group, yet applicants using screen readers with intermittent mobile connections complete tasks less often, especially in Spanish. Because the evidence combines language, assistive technology, and connectivity, it does not establish one isolated cause. The agency should investigate the intersecting barriers with affected users, improve the experience, and retest task completion using representative subgroup data. Aggregate performance or a single-factor fix could leave important barriers undiscovered.

Why each option fits or fails:

A. An aggregate result could conceal the substantial completion gap experienced by an intended subgroup.

B. The results do not isolate network performance as the cause because screen-reader use and language also vary across the observed groups.

C. The evidence shows poorer outcomes in the combined interaction context, requiring affected-user research, remediation, and representative subgroup evaluation.

D. Spanish localization may contribute to the gap, but it does not explain or address the reduced English completion in the same access context.

Question 14

Topic: Foundation-model applications

A company investigates why its grounded support assistant still states a superseded return period. Select TWO conclusions supported by the evidence.

Source and retrieval observations:

• The connected repository contains revision 7, effective June 1, specifying 30 days. Revision 6 was removed on May 31.
• The June 2 knowledge base sync completed successfully.
• Retrieval uses top_k=1 with no metadata filter.

Options:

• A. Revision 7 was ingested successfully and is available for retrieval.

• B. Revision 7 reached model context but was ignored during generation.

• C. The foundation model requires retraining before revision 7 can be used.

• D. Revision 6 remains authoritative in the connected source repository.

• E. Revision 6 remains searchable and supplied the stale grounding context.

Correct answers: A and E

Explanation: A grounded assistant depends on the retrieval index, not merely the current source document. Revision 7 is searchable, proving that its ingestion succeeded. However, revision 6 was not retired from the index and lacks metadata that could distinguish it as superseded. It ranked first, and top_k=1 caused it alone to enter the model context. The corrective investigation should therefore focus on deletion synchronization, ingestion status, document metadata, effective-date filtering, and retrieval ranking. Foundation model retraining is separate from refreshing or cleaning the searchable representation and would not remove the stale indexed record.

Why each option fits or fails:

A. Revision 7 appears as a ranked search result after the successful sync, showing that its content reached the searchable representation.

B. Revision 7 ranked second and was excluded from context by top_k=1, so the model never received it for this response.

C. Retraining changes model parameters, whereas this failure concerns stale indexed content and retrieval selection.

D. The repository removed revision 6 and identifies revision 7 as the effective source. Revision 6 remains searchable, but that does not make it authoritative.

E. Revision 6 ranked first and was the sole record sent to the model because retrieval used top_k=1.

Question 15

Topic: Security, compliance and governance

An expense assistant uses a foundation model to propose structured tool arguments. The Finance API trusts the Tool Gateway’s service role, and the global transfer range is USD 1-10,000.

Which control should be inserted between the schema validator and Tool Gateway to prevent the displayed unauthorized or altered transaction from executing?

Options:

• A. Authorize the proposed fields against the employee’s request and transaction entitlements.

• B. Verify the referenced accounts and available balance before sending the transaction.

• C. Apply a content-safety guardrail to the request and proposed transaction record.

• D. Recheck the proposed fields against the JSON schema and global amount range.

Best answer: A

Explanation: Generated tool arguments must be treated as untrusted instructions. Schema validation proves that fields have acceptable names, types, and broad ranges; it does not prove that values match the user’s intent or permissions. Here, the valid record follows the execution path to a gateway using a service role even though Project B and USD 5,000 conflict with the request and entitlement data. Before invocation, the application must bind the proposed action to the authenticated employee and validate its source, recipient, amount, and required approval against trusted authorization data. Account checks, content filtering, and audit logging address other risks but cannot replace action-level authorization.

Why each option fits or fails:

A. This check would detect that both the source project and amount conflict with the employee’s request and permitted transaction scope.

B. Account existence and sufficient funds establish business validity but do not establish the employee’s authority to transfer those funds.

C. A content-safety guardrail can assess configured harmful-content risks, but it does not enforce transaction permissions or confirm intended values.

D. The record already satisfies those structural checks, which do not determine whether this employee may perform the proposed transaction.

Question 16

Topic: AI and ML fundamentals

A support team processes recorded Spanish-language calls. The customer case system requires English text. Company policy requires a reviewer to confirm that the translation preserves the source meaning before it enters the case record.

Which assignment of services and review timing correctly completes the workflow?

Options:

• A. Stage A: Amazon Translate; Stage B: Amazon Transcribe; approve before the case update.

• B. Stage A: Amazon Transcribe; Stage B: Amazon Comprehend; approve before the case update.

• C. Stage A: Amazon Transcribe; Stage B: Amazon Translate; review after the case update.

• D. Stage A: Amazon Transcribe; Stage B: Amazon Translate; approve before the case update.

Best answer: D

Explanation: Speech transcription and text translation are separate workflow stages. Amazon Transcribe consumes recorded speech and produces a source-language transcript. Amazon Translate then consumes that transcript and produces text in the target language. The reviewer must compare the English translation with the Spanish transcript and block the case update until the meaning check passes. Failed checks enter the correction queue. Reversing the services fails because Amazon Translate is not a speech-to-text service, while Amazon Comprehend performs language analysis rather than translation. The control must operate before the translated content reaches the customer case record.

Why each option fits or fails:

A. Amazon Translate expects text rather than recorded speech, so it cannot perform the first handoff shown.

B. Amazon Comprehend analyzes text but does not translate the Spanish transcript into the required English text.

C. The services match the data types, but post-update review allows unverified translated content to enter the case record.

D. Amazon Transcribe produces Spanish text from the audio, Amazon Translate converts that text to English, and approval precedes downstream use.

Question 17

Topic: Responsible AI

A company evaluates a loan-risk classifier using a validation dataset with equal representation from two demographic groups. The sample contains mostly web submissions, although production applications arrive through both web and mobile channels.

The company must assess fairness and data suitability before approval. Which proposed interpretation or action is NOT consistent with this requirement?

Options:

• A. Evaluate each group using samples representative of its production conditions.

• B. Interpret subgroup error rates alongside the consequences of false negatives.

• C. Investigate whether unequal missing-field rates contribute to subgroup errors.

• D. Interpret equal counts as resolving fairness concerns despite subgroup error differences.

Best answer: D

Explanation: Equal group counts establish balance by count, not fairness of model outcomes. A balanced validation dataset can make subgroup metrics easier to estimate, but it may still misrepresent production conditions within each group. Here, the mostly web-based sample may not reflect mobile submissions, while Group B has both more missing data and a higher false-negative rate. These findings justify examining data quality, sampling conditions, subgroup errors, and the consequences of those errors. The disparity does not by itself prove unlawful discrimination, but equal counts cannot be used to dismiss it. Fairness assessment requires representative evidence and outcome analysis, not demographic balance alone.

Why each option fits or fails:

A. Including relevant web and mobile conditions tests whether subgroup performance generalizes beyond the mostly web-based validation sample.

B. Examining both subgroup performance and error consequences is appropriate because the same false-negative rate can have different practical significance.

C. The data-quality difference could affect model inputs and help explain the observed disparity, so investigating it supports the fairness assessment.

D. Balanced counts improve numerical representation, but they do not negate unequal error rates, differing data quality, or unrepresentative submission conditions.

Question 18

Topic: Security, compliance and governance

A reviewer needs the actual input and output content, caller ARN, and invocation time for supported Amazon Bedrock model invocations. The log destination will have appropriate access controls and retention. Which logging capability directly provides this evidence?

Options:

• A. Enable VPC Flow Logs for Amazon Bedrock endpoint traffic

• B. Enable CloudTrail logging for Amazon Bedrock API activity

• C. Enable Amazon Bedrock model invocation logging

• D. Enable AWS Config recording for Amazon Bedrock resources

Best answer: C

Explanation: Amazon Bedrock model invocation logging is designed to capture supported model interaction records, including request and response content and metadata such as the caller ARN and invocation time. Because prompts and responses may contain sensitive information, the destination requires appropriate access restrictions, encryption, and retention controls. CloudTrail complements these records by auditing broader API activity, including calls that change logging or other service configurations, but it does not replace invocation logging when reviewers need actual interaction content. The key distinction is model payload evidence versus general API audit evidence.

Why each option fits or fails:

A. VPC Flow Logs capture network traffic metadata such as addresses, ports, and byte counts, not model prompts and responses.

B. CloudTrail records API audit activity and configuration changes, but it does not provide the full model input and output content needed here.

C. Model invocation logging can capture supported request and response content along with invocation metadata such as caller identity and time.

D. AWS Config records supported resource configuration and configuration history, not the content exchanged during model invocations.

Question 19

Topic: Foundation-model applications

A company will customize a foundation model to draft replies for customer cases newly opened after September 30. Records were split by exported filename. Versions of the same case retain most of the original conversation.

Which TWO conclusions does the evidence support? Select TWO.

Options:

• A. Cases C-305 and C-411 represent the intended future-use period.

• B. Unique exported filenames make the current evaluation partition independent.

• C. The case dates establish that evaluation topics match production topics.

• D. Case C-104 v2 measures generalization to an unseen customer case.

• E. Case C-104 causes leakage across training and evaluation partitions.

Correct answers: A and E

Explanation: A meaningful evaluation set must contain independent examples that represent intended future use. Splitting by filename is insufficient when related records share an underlying customer case. Placing C-104 v1 in training and C-104 v2 in evaluation can let the model benefit from repeated conversation content, inflating measured performance. Records should first be grouped by customer case so every version remains in one partition.

The intended workload covers newly opened cases after September 30, so the distinct October and November cases provide appropriate temporal holdout examples. However, their dates alone do not prove that their topics match the full production distribution; representativeness still requires separate review.

Why each option fits or fails:

A. These distinct cases were opened after September 30, matching the stated timing of the intended production workload.

B. Filename uniqueness does not provide independence when different files contain related versions of the same underlying case.

C. Dates identify temporal alignment but provide no evidence about case topics or their similarity to the production distribution.

D. C-104 is not unseen because an earlier version of that customer case is included in training.

E. Both partitions contain versions of the same underlying case with shared conversation content, allowing recollection rather than testing generalization.

Question 20

Topic: Responsible AI

A team evaluates a linear classifier used to route benefit applications for manual review. Labels were audited, and all datasets were sampled from the same period and applicant population. The target F1 score is 0.80.

Increasing training records from 20,000 to 100,000 changed each score by less than 0.01. Domain experts identify important feature interactions that the current representation does not capture.

Which TWO conclusions are supported?

Options:

• A. The feature representation or model capacity warrants review.

• B. The evidence indicates overfitting with high statistical variance.

• C. A training-to-evaluation distribution shift is the primary issue.

• D. More similarly represented training examples are the best first remedy.

• E. The evidence indicates underfitting with high statistical bias.

Correct answers: A and E

Explanation: Underfitting occurs when a model fails to learn important patterns even in its training data. Here, training, validation, and independent evaluation F1 scores are all similarly below the target. The learning plateau and missing feature interactions further indicate high statistical bias caused by unsuitable features, insufficient model capacity, or both. The team should first investigate better feature representation or a model capable of expressing the relevant relationships. In this context, statistical bias describes systematic prediction error and is distinct from demographic or societal bias. Overfitting would instead show a larger gap between strong training performance and weak evaluation performance.

Why each option fits or fails:

A. The uncaptured feature interactions and performance plateau suggest the classifier or its inputs cannot represent important task patterns.

B. Overfitting would normally produce strong training performance with substantially weaker validation or evaluation performance, which is not observed.

C. The datasets represent the same period and population, while poor training performance shows that an evaluation-only distribution change is not the main explanation.

D. Performance already plateaued after a large data increase, so more data with the same representation is unlikely to address the missing patterns.

E. Persistently poor and similar performance across training and evaluation data is characteristic of underfitting rather than poor generalization alone.

Question 21

Topic: Responsible AI

A company is building an Amazon Bedrock assistant to answer employees’ questions about current, company-specific health-benefit eligibility. The assistant will use retrieval-augmented generation, cite sources, and decline unsupported answers.

Candidate sources:

Which source is most suitable as the primary grounding corpus?

Options:

• A. Use the regulator guidance library as the primary source.

• B. Use the benefits-team policy portal as the primary source.

• C. Use the recent support-chat collection as the primary source.

• D. Use the archived employee handbook as the primary source.

Best answer: B

Explanation: Data-source suitability depends on alignment with the intended task, not simply reputation, volume, or public availability. The benefits policy portal contains the company’s governing rules, is refreshed daily, has documented quality controls, and is approved for assistant use. These properties make it the strongest primary source for current eligibility answers. Regulator guidance can provide a legal baseline, support chats can reveal common questions, and an older handbook can preserve historical context, but none is as authoritative and current for the stated task. Retrieval-augmented generation cannot correct omissions or inaccuracies already present in its source corpus.

Why each option fits or fails:

A. It is current and reviewed, but statutory minimums do not define the company’s complete eligibility rules.

B. It is company-specific, current, approved for this use, and maintained through versioning and benefits-owner review.

C. The chats are recent and permitted, but case-resolution conversations are anecdotal rather than an authoritative statement of policy.

D. It is company-specific and approved, but the 2022 edition may conflict with current eligibility rules.

Question 22

Topic: Responsible AI

A team uses Amazon Bedrock Model Evaluations to compare two foundation models. Human reviewers apply the same task-success rubric to 200 representative cases using identical prompts and inference settings.

Individual Model B record:

Recommendation: High risk; send for manual credit review. Generated rationale: A recent late payment caused this recommendation.

No attribution or counterfactual analysis was performed for the individual record.

Which conclusions does the evidence support? Select TWO.

Options:

• A. The faithfulness of the individual rationale remains unverified by this evaluation.

• B. The pass-rate difference identifies the mechanism behind Model B’s recommendation.

• C. The generated rationale establishes that late payment caused the recommendation.

• D. Model B had a higher pass rate on the specified evaluation set.

• E. Model B will retain its seven-point advantage across all production cases.

Correct answers: A and D

Explanation: Amazon Bedrock Model Evaluations characterize observed behavior under defined cases, rubrics, prompts, and inference settings. Model B passed 178 cases compared with Model A’s 164, supporting the limited conclusion that Model B performed better on this evaluation. The result does not guarantee future performance or reveal how an individual output was produced.

The rationale for the high-risk recommendation is model-generated content. Because no decision-level analysis validated it, the claim that a late payment caused the recommendation remains unverified. Aggregate evaluation evidence and faithful explanations of individual decisions answer different questions and should not be treated as interchangeable.

Why each option fits or fails:

A. No decision-level analysis validated the rationale, so the aggregate evaluation leaves its faithfulness unresolved.

B. Aggregate rubric results compare observed outputs; they do not expose the mechanism producing an individual recommendation.

C. A model-generated rationale is another output and does not prove that the stated factor causally determined the recommendation.

D. The shared cases and settings permit the observed 89% versus 82% comparison, limited to this evaluation.

E. Representative evaluation evidence can inform expectations, but it cannot guarantee the same performance gap for every future production case.

Question 23

Topic: Responsible AI

A customer-support assistant uses Amazon Bedrock Guardrails.

Policy: Every candidate response must be assessed against current harmful-language filters and denied topics immediately before display. Cache entries can outlive policy updates.

Current flow:

Testing shows that prohibited model-generated text triggers an intervention, but harmful cached text is displayed. Which change best closes the control gap?

Options:

• A. Call ApplyGuardrail on each user request immediately before cache routing.

• B. Call ApplyGuardrail on each model response immediately before cache insertion.

• C. Run ApplyGuardrail against all cache entries on a scheduled interval.

• D. Call ApplyGuardrail on each merged candidate response immediately before display.

Best answer: D

Explanation: Amazon Bedrock Guardrails assess only content that passes through a guardrail-enabled model invocation or an explicit ApplyGuardrail request. In the displayed flow, model-generated text is assessed, but a cache hit reaches the merge point without that safeguard. Applying the guardrail after both branches merge establishes a common control boundary immediately before display. The application must configure the relevant content filters and denied topics, inspect the intervention result, and prevent blocked text from reaching the user. Representative testing remains necessary because enabling a guardrail reduces defined risks but does not guarantee that every inappropriate response will be detected. Input-only, write-time, and scheduled checks do not satisfy the stated per-display policy.

Why each option fits or fails:

A. Assessing requests can block harmful inputs, but it does not assess harmful text returned from either response branch.

B. This protects newly generated responses, but cached entries can bypass assessment or remain inconsistent with later policy updates.

C. Scheduled scanning can detect stored harmful content, but unsafe entries may still be displayed between scans.

D. The merged checkpoint receives both cached and model-generated responses, so every user-visible candidate is assessed under the current policy.

Question 24

Topic: Foundation-model applications

A team evaluates policy summaries with ROUGE-1 recall: overlapping unigram occurrences divided by 30 reference-token occurrences. Text is lowercased, punctuation is removed, and scores are rounded to two decimals.

Evaluation exhibit:

Source:
A supplier must submit a remediation plan within 10 business days
after receiving written notice of a critical audit finding. The security
director must approve the plan before work begins.

Summary A (ROUGE-1 recall: 0.87):
A supplier must submit a remediation plan within 10 business days after
a critical audit finding. The security director must approve the plan
before work begins.

Summary B (ROUGE-1 recall: 0.70):
After the supplier receives written notice of a critical audit finding,
it has 10 business days to provide a corrective plan. The security
director must approve it before implementation starts.

Which TWO conclusions are supported by the exhibit?

Options:

• A. Summary B retains the written-notice trigger and the required director approval.

• B. Summary A has greater overlap, but its score does not establish practical correctness.

• C. Summary B should be rejected because paraphrasing reduces its reference-token overlap.

• D. Summary A preserves the deadline condition more faithfully because its score is higher.

• E. Summary B necessarily omits more source facts because its score is lower.

Correct answers: A and B

Explanation: ROUGE-1 recall measures how many reference unigram occurrences appear in a candidate summary. Under the stated convention, Summary A covers 26 of 30 reference-token occurrences, while Summary B covers 21 of 30. However, Summary A omits the crucial condition that the deadline begins after the supplier receives written notice, potentially changing when the deadline starts. Summary B uses more paraphrasing but preserves that trigger, the 10-business-day deadline, and director approval before implementation. ROUGE can indicate lexical coverage, but it does not establish factuality, practical usefulness, or preservation of critical qualifications.

Why each option fits or fails:

A. Summary B preserves receipt of written notice as the deadline trigger and requires the security director’s approval before implementation.

B. The values show greater reference-token overlap for Summary A, while ROUGE does not verify factual accuracy or preservation of qualifications.

C. Valid paraphrases can reduce lexical overlap while preserving meaning, so a lower ROUGE score alone does not justify rejection.

D. Summary A changes the trigger from receiving written notice to the audit finding itself, despite having greater lexical overlap.

E. ROUGE-1 recall measures token overlap rather than semantic fact coverage, so the lower score does not prove additional factual omissions.

Question 25

Topic: Generative AI fundamentals

A team compares three foundation models using documented capabilities and successful test results.

Which TWO conclusions does the evidence support? Select TWO.

Options:

• A. Lyra generates images from text input.

• B. Lyra understands image input and generates images.

• C. Echo generates audio from audio input.

• D. Orion understands image input but generates text output.

• E. Orion generates images from image input.

Correct answers: A and D

Explanation: Input and output modalities describe different capabilities. Orion accepts images and produces descriptions, so it can understand image content while generating only text in the demonstrated workflow. Lyra accepts text and can produce an image, establishing text-to-image generation even though image understanding is not shown. Echo accepts audio but returns a transcript, demonstrating audio-to-text processing rather than speech or audio generation.

A model described as multimodal does not necessarily support every direction between its modalities. Always verify both the supported input and the produced output.

Why each option fits or fails:

A. Lyra accepts text requests and has demonstrated image output, establishing a text-to-image capability.

B. Lyra generates images, but the exhibit lists only text as its supported input, so image understanding is not established.

C. Echo consumes audio and returns text, which demonstrates audio understanding or transcription rather than audio generation.

D. Orion successfully processes an uploaded image, but the demonstrated output is a text description rather than a generated image.

E. Image input demonstrates image understanding, while Orion’s documented result provides no evidence of image generation.

Questions 26-50

Question 26

Topic: Security, compliance and governance

Policy: SourcePrepRole may process SSNs. Among human users, only FraudInvestigator members may access SSNs, and the support assistant must not return them. The human groups below are disjoint. All stores are encrypted at rest.

Which interpretation is best supported by the exhibit?

Options:

• A. Only assistant retrieval violates; prepared objects are internal derived data.

• B. Only prepared-object reads violate; service-role retrieval prevents support-user SSN access.

• C. Both prepared-object reads and assistant retrieval violate the SSN policy.

• D. Neither path violates because the raw source is restricted and encrypted.

Best answer: C

Explanation: Access controls must follow sensitive information throughout its lifecycle, including prepared datasets, retrieval stores, and application outputs. The raw S3 objects are restricted, but unchanged SSNs were copied into prepared objects readable by unauthorized analysts. The retrieval collection also retains the chunk text, and the assistant returns that text to support users. The service role’s permission to read the collection does not authorize every end user to receive its contents. Encryption at rest protects stored data from certain infrastructure threats, but it does not prevent access through granted permissions. Sensitive fields should be removed or tokenized before creating derived artifacts, and each downstream access path must enforce appropriate authorization.

Why each option fits or fails:

A. Derived status does not remove sensitivity, so direct analyst access to prepared objects containing unchanged SSNs also violates the policy.

B. A service role mediating retrieval does not prevent disclosure when its retrieved content is returned to unauthorized support users.

C. Data analysts can read derived SSNs directly, while support users can receive the same SSNs through retrieval mediated by the service role.

D. Protecting and encrypting the source does not secure sensitive values copied into broadly accessible downstream artifacts.

Question 27

Topic: Generative AI fundamentals

A generative application reports these averages for the same requests. The stages occur sequentially:

• Retrieval: 2.1 seconds
• Queueing before the model: 3.0 seconds
• Model response latency: 0.8 seconds
• Model throughput: 20 requests/second

Which statement correctly interprets the user waiting time?

Options:

• A. Waiting is about 5.9 seconds despite the model’s 0.8-second latency.

• B. Waiting is about 3.0 seconds because the slowest stage determines it.

• C. Waiting is about 0.8 seconds because model latency covers the workflow.

• D. Waiting is about 0.05 seconds because throughput determines request duration.

Best answer: A

Explanation: Response latency measures how long a particular operation takes, while throughput measures the number of requests processed per unit of time. End-to-end user waiting includes all sequential stages and delays in the request path. Here, the average wait is 2.1 + 3.0 + 0.8 = 5.9 seconds. The throughput of 20 requests/second describes aggregate processing capacity, not how quickly an individual request completes. Consequently, a model can have low response latency and adequate throughput while users still experience substantial delays from retrieval or queueing.

Why each option fits or fails:

A. Sequential retrieval, queueing, and model latency add to an average end-to-end wait of 5.9 seconds.

B. The longest stage determines total time only when stages fully overlap; these stages are sequential.

C. Model latency measures processing after the request reaches the model, excluding retrieval and queueing.

D. The reciprocal of throughput is not per-request waiting time because multiple requests can be processed concurrently.

Question 28

Topic: Foundation-model applications

A travel-booking agent must follow this policy:

• Clarify missing preferences that could change which flight qualifies.
• Before purchase, show the exact flight, seat, refund terms, and total, then obtain approval.
• If a purchase result is unknown, query its status and do not submit another purchase. Escalate if the status remains unknown.

The user says:

Book the cheapest refundable Monday morning flight. I prefer an aisle seat, but any seat is acceptable.

The agent proposes an exact itinerary with seat 12C and a total price. The user approves it. The purchase tool then times out and returns an unknown result.

Which TWO conclusions are supported? Select TWO.

Options:

• A. The agent should retry because the approved purchase details are unchanged.

• B. The agent could settle the exact seat during the confirmation step.

• C. The agent should check purchase status and escalate if uncertainty remains.

• D. The agent may report failure because the tool returned no confirmation.

• E. The agent had to clarify the aisle preference before selecting a flight.

Correct answers: B and C

Explanation: Consequential agent actions require confirmation before creating an external commitment and careful handling when the outcome is uncertain. The user made the aisle preference optional, allowing the agent to select a flight and present the exact seat with the other details for approval. After the approved purchase was submitted, the timeout did not prove success or failure. The agent should query the transaction status and escalate if it cannot determine the outcome. Retrying or reporting failure prematurely could respectively create a duplicate booking or misrepresent a completed transaction.

Why each option fits or fails:

A. A second submission could create a duplicate external commitment even when its details match the approved purchase.

B. The seat preference was nonbinding, so the agent could select a qualifying flight and include the exact seat in the approval request.

C. A timeout does not establish failure, so checking status and escalating an unresolved result avoids creating a duplicate booking.

D. An unknown tool result is not evidence that the purchase failed; the original transaction may have completed.

E. The user explicitly accepted any seat, so the aisle preference did not create ambiguity about which flights qualified.

Question 29

Topic: Generative AI fundamentals

A company evaluates whether a generative AI assistant’s explanations show that its policy recommendations are trustworthy.

Evaluation exhibit:

Which interpretation is best supported by the evidence?

Options:

• A. The persuasiveness rating verifies that the explanations reliably represent model computation, while the benchmark measures only recommendation consistency.

• B. The benchmark supports 89% recommendation accuracy and the audit supports 93% claim grounding, but neither verifies faithful internal reasoning.

• C. The citation audit establishes 93% recommendation accuracy, while the held-out benchmark primarily evaluates the quality of the explanations.

• D. The rerun establishes overall reliability because most recommendations matched, making separate accuracy and source-support checks unnecessary.

Best answer: B

Explanation: Trustworthiness requires separating fluency, task performance, grounding, and reproducibility. The held-out result gives a recommendation accuracy of 356 / 400, or 89%. The citation audit shows that 93% of sampled factual claims were supported by their cited passages; it does not mean that 93% of recommendations were correct. The 4.8 persuasiveness score measures reader perception, not factuality or faithful reasoning. Similarly, 47 matching reruns indicate output stability, not correctness. Generated explanations can be persuasive post-hoc narratives and may not represent the model’s internal computation. Reliable evaluation therefore depends on appropriate benchmarks, source checks, and reproducible measurements rather than explanation length or confidence.

Why each option fits or fails:

A. Persuasiveness measures how convincing users found the text, not whether the explanation is accurate or faithful to the model’s computation.

B. The held-out benchmark measures recommendation correctness, while the citation audit measures sampled claim support; generated explanations do not necessarily reveal internal computation.

C. The citation audit measures support for sampled claims, whereas comparison with reference recommendations measures task accuracy rather than explanation quality.

D. Repeatability measures consistency rather than correctness or grounding, so matching recommendations cannot replace independent performance and citation checks.

Question 30

Topic: Foundation-model applications

A company will compare two foundation models that summarize policy notices. The decisive criterion is whether each summary preserves triggering conditions, deadlines, and required actions. The company has representative prompts and reference material and wants repeatable scoring at scale without assigning every response to human raters. Which Amazon Bedrock evaluation approach best fits?

Options:

• A. Use human scoring with an explicit preservation rubric across the prepared dataset.

• B. Use evaluator-model scoring with an explicit preservation rubric across the prepared dataset.

• C. Use semantic-similarity scoring against the references and rank models by aggregate score.

• D. Use evaluator-model scoring with a general quality rubric across the prepared dataset.

Best answer: B

Explanation: Amazon Bedrock can support foundation-model comparisons through computed metrics, an evaluator model such as LLM-as-a-judge, or human ratings. The evaluation method must match the business criterion. Preserving conditions, deadlines, and required actions is a qualitative, domain-specific requirement that semantic similarity alone does not establish. An evaluator model can apply an explicit rubric consistently across representative prompts and relevant reference material, providing scalable comparative evidence. Human judgments can still help calibrate or validate the evaluator. Bedrock supplies the evaluation mechanism, but the team must define the dataset, criterion, and success threshold.

Why each option fits or fails:

A. Human ratings can assess the rubric effectively, but rating every response conflicts with the requirement to avoid per-response human review.

B. An evaluator model can apply a defined qualitative criterion repeatedly at scale, while the representative dataset and references ground the comparison.

C. Semantic similarity measures overall correspondence, but a high score does not establish that every required condition, deadline, and action was preserved.

D. A general quality rubric may reward fluency or relevance while missing the specific preservation failures that determine success.

Question 31

Topic: Generative AI fundamentals

A retailer’s support agent must use inventory and refund tools owned by different teams. The company wants a common integration method. Its identity service provides per-user permissions, and the refund service must reject unauthorized calls at the action boundary. The application will validate tool responses.

Which design appropriately uses Model Context Protocol (MCP)?

Options:

• A. Expose tools through MCP servers; authenticate each server connection once and grant identical refund permissions to every session.

• B. Expose tools through MCP servers; let MCP capability discovery select permitted tools and replace action-level authorization.

• C. Expose tools through MCP servers; let the agent application select tools and preserve identity-based authorization at each action boundary.

• D. Expose tools through MCP servers; treat validated responses as evidence that each requested refund action was authorized.

Best answer: C

Explanation: MCP provides a common way for an agent application to discover and interact with external tools or context sources. It reduces connector-specific integration work, but protocol compatibility does not decide which tool the agent should use or whether an action is safe and permitted. The application must retain appropriate tool-selection logic, propagate the relevant identity, and enforce authorization at the actual service or resource boundary. Tool responses also require separate validation because a compatible tool can still return incorrect or unsafe data. MCP enables communication; it does not replace authentication, authorization, or output validation.

Why each option fits or fails:

A. A shared authenticated connection does not satisfy the requirement to enforce different permissions for individual users.

B. Capability discovery describes available tools but does not determine whether a particular user is authorized to perform an action.

C. MCP standardizes access to external capabilities, while the application and services remain responsible for tool selection and permission enforcement.

D. Response validation can assess returned data, but it does not prove that the underlying action was permitted for the user.

Question 32

Topic: AI and ML fundamentals

A logistics company trains a delivery-scheduling agent in a simulator. At each step, the agent selects a delivery, changing the vehicle’s location, remaining packages, and later delivery choices.

• Rewards include +10 for an on-time delivery and -4 for a late delivery.
• The policy is updated to increase cumulative reward across the full route.
• Dispatchers compare completed routes; their preferences are converted into scores used during policy updates.
• Route logs are also stored in a dashboard but are not used for learning.

Which TWO conclusions are supported?

Options:

• A. Dispatcher preferences contribute to the learning reward signal.

• B. On-time results are supervised labels for route classification.

• C. Dashboard logs provide a learning signal to the policy.

• D. The policy is being trained through reinforcement learning.

• E. The agent should maximize only the next immediate reward.

Correct answers: A and D

Explanation: Reinforcement learning involves an agent taking actions in an environment, observing resulting states and rewards, and improving a policy toward a goal. Here, each delivery choice changes the agent’s subsequent experience, while on-time and late-delivery rewards measure progress toward effective routes. Because training maximizes cumulative reward, delayed outcomes can affect how earlier actions are learned. Human preferences can also become part of reinforcement learning when they are converted into scores that influence policy updates. In contrast, operational feedback or logs that are only stored do not become learning signals. The key distinction is whether feedback actually guides policy learning, not whether the feedback comes from a human or an automated system.

Why each option fits or fails:

A. The preference scores influence policy updates, so they function as reward information rather than merely stored application feedback.

B. The on-time results provide rewards for actions in a sequential learning process, not labeled categories for training a classifier.

C. The logs are stored for reporting but do not enter policy updates, so they do not influence the learning process.

D. The agent takes actions that affect subsequent states, receives goal-related rewards, and updates its policy to improve cumulative reward.

E. The policy is updated using cumulative route reward, which allows future outcomes to influence the value of earlier actions.

Question 33

Topic: Security, compliance and governance

A company stores documents for an Amazon Bedrock knowledge base in an Amazon S3 bucket. Its rule requires versioning and default encryption with a specified customer-managed KMS key. During the next quarter, a reviewer must identify configuration changes and determine whether the bucket met the rule after each change. Which approach best provides this evidence?

Options:

• A. Assess the bucket control with Audit Manager and review collected assessment evidence.

• B. Log bucket API calls with CloudTrail and analyze management event history.

• C. Scan the bucket with Amazon Inspector and review vulnerability findings after changes.

• D. Record the bucket with AWS Config and apply change-triggered Config rules.

Best answer: D

Explanation: AWS Config records configuration changes for supported AWS resources and maintains configuration history. Change-triggered Config rules can evaluate the Amazon S3 bucket after a recorded change, showing whether versioning and the required KMS encryption configuration satisfy the organizational rule. CloudTrail records API activity and can help identify who requested a change, but an API event is not a configuration snapshot or compliance evaluation. Audit Manager can organize collected evidence for an assessment, while Amazon Inspector focuses on vulnerability findings for supported workloads. AWS Config evidence supports this specific resource control; it does not prove that the entire AI solution complies with every security or regulatory obligation.

Why each option fits or fails:

A. Audit Manager organizes evidence for assessments, but it does not itself continuously record bucket configurations or evaluate each configuration change.

B. CloudTrail identifies API activity, but its event history does not directly provide configuration snapshots and rule evaluations for each resulting state.

C. Amazon Inspector produces vulnerability findings for supported workloads rather than evaluating Amazon S3 configuration states against organizational rules.

D. AWS Config records resource configuration history and can evaluate the bucket against organizational requirements when its configuration changes.

Question 34

Topic: Generative AI fundamentals

A marketing team selects an image-capable foundation model to create new product scenes and edit existing photos. During generation, the model starts with a noisy representation and repeatedly predicts and removes noise while following text and image guidance. Which approach best matches this behavior?

Options:

• A. Use an autoregressive model with sequential image tokens.

• B. Use vector search to retrieve similar stored images.

• C. Use a diffusion model with iterative denoising.

• D. Use a GAN with adversarial generator training.

Best answer: C

Explanation: Diffusion models generate media through an iterative denoising process. They begin with a noisy representation and repeatedly estimate and remove noise, often conditioned on a text description, source image, or other supported input. This process can produce a new image or modify an existing one. By contrast, GANs use adversarial training between generator and discriminator networks, while autoregressive models generate a sequence of elements. Vector search retrieves stored content, so it does not explain the creation of a new scene. The repeated refinement of noise is the decisive evidence for diffusion-based generation.

Why each option fits or fails:

A. Autoregressive generation predicts successive tokens or elements based on earlier ones rather than iteratively denoising a representation.

B. Vector search finds existing images by similarity; it does not create or edit media by progressively removing noise.

C. Diffusion generation progressively removes noise from a representation, using conditioning to guide the result toward the requested image.

D. A GAN learns through competition between a generator and discriminator, not through the described repeated removal of noise.

Question 35

Topic: Foundation-model applications

A manufacturer is building a grounded investigation assistant in Amazon Neptune Analytics.

Data:

• Suppliers, components, products, and incidents are connected by graph edges.
• Incident reports are free text and often describe similar failures differently.
• Supplier and product relationships change frequently.

The assistant must find incidents similar to a new report and identify affected products through current supplier-component relationships. Which retrieval approach best supports this goal?

Options:

• A. Use vector similarity, then traverse connected supplier-component-product paths.

• B. Use graph traversal over exact entity and incident properties.

• C. Use vector similarity over flattened supplier and incident documents.

• D. Use keyword filtering, then vector reranking of incident reports.

Best answer: A

Explanation: Graph-aware retrieval adds value when an answer depends on paths among connected entities rather than one isolated passage. Here, vector similarity can identify past incident narratives that are semantically related despite different wording. Graph traversal can then follow the current supplier-to-component-to-product relationships associated with those incidents. Amazon Neptune Analytics supports both graph and vector capabilities, provided the data model and retrieval integration represent the required edges and embeddings. Vector search alone may miss or stale-copy relationships, while graph traversal alone may not recognize semantically similar narratives. The combined workflow grounds the response in both textual similarity and connected facts.

Why each option fits or fails:

A. Vector search finds semantically similar reports, while graph traversal follows current entity relationships needed to identify affected products.

B. Graph traversal resolves connected entities, but exact properties may miss incident narratives that describe similar failures using different terminology.

C. Flattened documents support semantic matching but can duplicate or retain stale relationship data when supplier and product connections change.

D. This can improve report retrieval but does not traverse graph edges to establish which products are connected through current supplier-component relationships.

Question 36

Topic: AI and ML fundamentals

A company describes its invoice workflow as AI-enabled. An auditor compares two releases.

Policy: Amount > $10,000 routes to manager review; blocked suppliers are rejected. No policy edits occurred between releases.

Which interpretation is best supported by the evidence?

Options:

• A. Rule-based categorization followed by rule-based routing and rejection

• B. Learned categorization followed by learned routing and rejection

• C. Learned categorization followed by rule-based routing and rejection

• D. Rule-based categorization followed by learned routing and rejection

Best answer: C

Explanation: The workflow combines machine learning inference with deterministic business logic. The description mapper learns relationships from labeled invoices, and its changed output for identical text after retraining supports this interpretation. In contrast, manager review and rejection continue to follow explicit, unchanged policy conditions based on amount and supplier status.

An AI-enabled system can therefore use learned patterns in one stage and conventional rules in another. Natural-language input or an AI label does not mean every decision is produced by machine learning; determine which components are trained from examples and which apply policies written by people.

Why each option fits or fails:

A. Routing and rejection are rule-based, but retraining and the changed category for identical text indicate a learned description mapping.

B. The labeled examples support learned categorization, but routing and rejection directly follow the unchanged human-written policy conditions.

C. Retraining changed the text category, while the unchanged amount and supplier policies continued to determine review and rejection outcomes.

D. The category changed after retraining on labeled examples, while no evidence indicates that routing or rejection is predicted from learned relationships.

Question 37

Topic: Generative AI fundamentals

An order-support application receives this context for the customer request, “Can I return order 184, and was my cancellation completed?”

Which TWO conclusions does this evidence support? Select TWO.

Options:

• A. The tool result shows cancellation was rejected because the order had already shipped.

• B. The policy excerpt’s quoted directive remains source content, not an application instruction.

• C. The retrieved note authorizes every request despite the application’s system-level instruction.

• D. The example establishes that order 184 is return-eligible under the same demonstrated conditions.

• E. The tool result becomes a reusable demonstration governing handling of later cancellation requests.

Correct answers: A and B

Explanation: A model interaction can combine several context roles. Instructions define desired behavior, examples demonstrate a response pattern, reference material supplies factual evidence, and tool results report observed action outcomes. Here, the worked exchange concerns another item, so it cannot establish order 184’s eligibility. The retrieved policy provides reference material, but command-like wording quoted inside that source remains content rather than becoming a trusted application instruction. The tool output directly establishes that the cancellation was rejected because the order had already shipped. Supplying examples, retrieved text, or tool output during inference does not retrain the model or automatically convert those materials into instructions.

Why each option fits or fails:

A. The tool directly reports a rejected status and identifies shipment as the reason, establishing the observed action outcome.

B. Imperative language inside retrieved external material remains reference content unless trusted application logic explicitly promotes it to an instruction.

C. Retrieved text does not gain instruction priority merely because it contains command-like language that conflicts with trusted application instructions.

D. The example concerns a different unopened item and provides no facts establishing that order 184 has the same conditions.

E. A tool response records one workflow outcome; it does not automatically become a demonstration or govern future requests.

Question 38

Topic: Foundation-model applications

A team evaluates 1,000 representative requests. Each stage receives only requests that passed the preceding stage. The business target is correct completion within 30 seconds.

Only 522 requests achieved the end-to-end target. Which improvement should the team prioritize?

Options:

• A. Improve external-action reliability and latency.

• B. Improve retrieval relevance before model evaluation.

• C. Improve extraction accuracy before downstream tuning.

• D. Improve model-decision accuracy before tool changes.

Best answer: A

Explanation: Stage-level yield and latency reveal where an end-to-end workflow loses business value. Extraction, retrieval, and model decision each retain at least 96% of the requests entering that stage, with p95 durations of 5 seconds or less. The external action completes only 648 of 913 requests, a loss of 265 requests, and its 42-second p95 duration exceeds the entire 30-second target. Therefore, action-tool reliability and latency are the strongest supported priorities for improving correct, timely completion. High-performing upstream components cannot compensate for a failing or slow final action stage.

Why each option fits or fails:

A. The action stage has the largest conditional loss, completing only 648 of 913 requests, and its p95 duration alone exceeds the end-to-end target.

B. Retrieval loses 39 of 980 requests and remains fast, making it a smaller contributor to unsuccessful or late outcomes.

C. Extraction succeeds for 98% of requests and has low latency, so its propagated loss is much smaller than the action-stage loss.

D. The model decision succeeds for 913 of 941 requests with acceptable latency, while many more requests fail or run late during the action.

Question 39

Topic: Security, compliance and governance

A retailer analyzes repeat-purchase patterns by age band and the first three characters of customers’ postal codes. Purchases must remain linkable to the same customer over time, but the workflow does not need customer identities.

The source currently sends names, emails, full birth dates, street addresses, account IDs, and purchase records. Which TWO actions best protect privacy while preserving the required analysis? Select TWO.

Options:

• A. Aggregate purchases by week, retaining age-band and postal-prefix totals for analysis.

• B. Send complete profiles to the model, masking names and emails in responses and dashboards.

• C. Retain complete encrypted profiles, using IAM permissions as the primary minimization control.

• D. Derive age bands and postal prefixes at source, omitting unnecessary direct identifiers.

• E. Tokenize account IDs before transfer, separate the mapping, and restrict record access.

Correct answers: D and E

Explanation: Data minimization limits collection and processing to attributes required for the business task. Age bands and postal prefixes should be derived before transfer so full birth dates and addresses never enter the AI workflow. Names and emails are unnecessary. Because repeat purchases must remain linked, a random customer token can replace the account ID while its mapping stays in a separate controlled system.

Tokenization and generalization reduce privacy risk but do not make data anonymous. Purchase histories, demographics, and location prefixes may still enable re-identification when combined, so tokenized records require access controls and appropriate retention. Encryption or output masking alone cannot replace minimizing data at ingestion.

Why each option fits or fails:

A. Weekly aggregates remove the customer-level continuity required to identify repeat-purchase patterns over time.

B. Output masking does not protect unnecessary identifiers already exposed to the model, processing path, or possible invocation logs.

C. Encryption and IAM protect access but do not minimize collection because the workflow still receives attributes it does not need.

D. Source-side generalization supplies the required segments while preventing names, emails, full birth dates, and street addresses from entering the workflow.

E. Tokens preserve longitudinal linking, while separation and access controls recognize that tokenized records can remain sensitive or re-identifiable.

Question 40

Topic: Generative AI fundamentals

An insurer’s six-week GenAI claims-summary demonstration produced promising user ratings with 40 employees. Production demand could range from 500 to 8,000 summaries weekly, token and retrieval costs vary by document, and reviewers can assess only 250 summaries weekly.

The next step must estimate cost per acceptable summary and quality at increasing load while controlling review volume and spending. Which proposed action does NOT satisfy this requirement?

Options:

• A. Add user cohorts after each tier meets defined cost and quality thresholds.

• B. Continue capped access and measure reviewed quality and cost per acceptable summary.

• C. Run capped demand tiers and pause between tiers for reviewer assessment.

• D. Open access companywide with alerts, adding quotas only if spending accelerates.

Best answer: D

Explanation: A successful small demonstration establishes initial feasibility, not production readiness across uncertain demand levels. A controlled rollout should increase usage in measured tiers, apply spending limits, and collect representative quality evidence within the reviewers’ capacity. Tracking cost per acceptable summary connects variable token and retrieval expenses with usable outcomes, providing meaningful unit economics. Budget alerts provide visibility, but they do not prevent unexpected spending; quotas or equivalent controls must be applied before broad expansion. Immediate companywide access therefore conflicts with the requirement to learn cost, quality, and operational limits before wider adoption.

Why each option fits or fails:

A. Threshold-based expansion uses evidence from each cohort before increasing demand, supporting a controlled and measurable rollout.

B. Capped access supports reliable unit-cost and quality measurement while keeping the evaluation workload within available review capacity.

C. Capped tiers and assessment pauses reveal operational limits while preserving sufficient reviewer capacity for meaningful quality checks.

D. Alerts report spending but do not limit it, and immediate companywide access bypasses controlled evidence gathering before broad adoption.

Question 41

Topic: AI and ML fundamentals

A company is comparing a publicly downloadable pretrained model with developing a model for a document-classification task.

Requirements: F1 score >= 0.90 and launch within 8 weeks.

The supplier publishes pretrained base-model updates. Other artifact rights and subgroup results were not assessed. Which TWO conclusions does the evidence support?

Options:

• A. The custom model’s aggregate F1 establishes equal quality across document subgroups.

• B. Task-specific development meets the quality target and requires company-led lifecycle operations.

• C. The pretrained source meets the launch target but misses the measured quality target.

• D. Provider base-model updates remove the company’s need to monitor the pretrained application.

• E. The pretrained license permits fine-tuning and redistribution because internal inference is approved.

Correct answers: B and C

Explanation: Model-source selection requires separate evaluation of task quality, delivery time, permissions, adaptation, and ongoing ownership. On the same held-out test, task-specific development clears the 0.90 F1 requirement, whereas the pretrained source does not. However, the pretrained source meets the 8-week schedule and task-specific development does not. Developing the model also leaves hosting, monitoring, and retraining with the company. Public availability and approval for internal inference do not imply rights to fine-tune or redistribute weights. Likewise, supplier updates do not remove application-level monitoring, and an aggregate metric cannot establish subgroup performance.

A source should be selected from demonstrated capabilities and explicit permissions, not assumptions based on download availability.

Why each option fits or fails:

A. An aggregate F1 score does not show whether performance is consistent across subgroups that were not separately evaluated.

B. Its 0.92 F1 score clears the threshold, while hosting, monitoring, and retraining remain company responsibilities.

C. Its 4-week estimate is within the deadline, but its 0.86 F1 score is below the required 0.90.

D. Supplier updates do not replace the company’s stated responsibility to validate and monitor its complete application.

E. Approval for internal inference does not establish permission to modify or redistribute the downloaded model artifact.

Question 42

Topic: Security, compliance and governance

An AI team uses Amazon Bedrock with an Amazon S3 knowledge source. AWS Trusted Advisor provides the following review artifact.

Which TWO responses are supported by the evidence? Select TWO.

Options:

• A. Treat the missing application findings as evidence of legal compliance.

• B. Conclude that the over-age IAM access key was compromised.

• C. Assess AI content, application authorization, and compliance obligations separately.

• D. Review the public S3 access and rotate the over-age IAM key.

• E. Conclude that unauthorized actors downloaded the public RAG documents.

Correct answers: C and D

Explanation: AWS Trusted Advisor provides best-practice findings for supported checks within the applicable service and support-plan scope. Here, it supplies actionable infrastructure evidence: public access conflicts with the bucket’s employee-only purpose, and the access key exceeds the organization’s age limit. These findings justify investigation and remediation, but they do not prove that data was accessed or credentials were compromised.

Trusted Advisor is only one input to a wider AI workload review. Separate evidence is needed to assess prompts and outputs, retrieval authorization, model behavior, application security, data handling, and legal obligations. Clearing the listed recommendations would therefore not certify the complete AI application as secure or compliant.

Why each option fits or fails:

A. Those areas were outside the report’s scope, so their absence does not provide affirmative evidence of compliance.

B. Exceeding the rotation policy raises credential risk but does not demonstrate that the key was used or compromised.

C. The stated coverage excludes model content, RAG authorization behavior, and legal duties, so those areas require separate evidence and review.

D. The employee-only intent conflicts with public read, and the key exceeds the stated rotation limit, so both findings warrant corrective action.

E. A public-read configuration establishes exposure, but it does not prove that an unauthorized download occurred.

Question 43

Topic: Security, compliance and governance

A security operations center triages account-takeover alerts containing attacker-controlled text.

• A classifier score estimates compromise probability. In independent representative testing, 92% of alerts scoring >= 0.90 were confirmed compromises.
• A generative AI summary may state “high confidence,” but that wording has not been independently validated.
• Scores >= 0.90 receive expedited human review; other alerts receive standard review. An analyst must approve account suspension.

Which proposed action is NOT consistent with this workflow?

Options:

• A. Use classifier scores >= 0.90 to prioritize expedited analyst review.

• B. Use lower classifier scores to retain alerts in standard analyst review.

• C. Use recent representative labels to monitor classifier calibration and drift.

• D. Use the assistant’s “high confidence” statement as an equivalent >= 0.90 score.

Best answer: D

Explanation: Confidence indicators are useful only when their meaning and relationship to correctness have been established for the relevant workflow. Here, the classifier produces a probability estimate, and independent representative testing supports using >= 0.90 for expedited review. The 92% result still means some high-scoring alerts are incorrect, so it does not justify treating compromise as proven or bypassing analyst approval.

The generative assistant’s phrase “high confidence” is a different kind of indicator. Because it has not been validated against labeled outcomes, especially for attacker-controlled inputs, it cannot be converted into or treated as the classifier threshold. Continued validation is also appropriate because calibration may change as input patterns drift.

Why each option fits or fails:

A. The independently tested threshold supports prioritization, while human review addresses the remaining possibility of an incorrect classification.

B. A score below the expedited threshold does not establish safety, so retaining the alert for standard review is appropriate.

C. Ongoing validation can detect whether the score’s relationship to actual correctness changes as production inputs evolve.

D. Verbal self-assurance has no validated mapping to the classifier’s probability estimate and cannot substitute for the tested threshold.

Question 44

Topic: AI and ML fundamentals

A support team trains a multilayer transformer neural network by adjusting its weights on historical customer chats and approved agent replies. During inference, it receives a new issue and composes a newly worded reply instead of selecting a stored template.

Which conclusion is best supported by the system’s learning process and output?

Options:

• A. It demonstrates machine learning and generative AI, but not deep learning because its output is natural language.

• B. It demonstrates machine learning, but neither deep learning nor generative AI because it learned from labeled conversations.

• C. It demonstrates machine learning and deep learning, but not generative AI because its training replies were approved.

• D. It demonstrates machine learning, deep learning, and generative AI through neural training and novel text output.

Best answer: D

Explanation: Machine learning describes systems that learn patterns or parameters from data. Deep learning is a family of machine learning approaches based on multilayer neural networks, including transformers. Generative AI describes the capability to create new content such as text, images, or audio. These categories overlap rather than compete: this system learns by adjusting neural-network weights, uses a deep architecture, and generates a new textual response during inference. Human approval of training examples would not change these classifications. The key distinction is that learning method identifies machine learning and deep learning, while the newly composed output identifies generative AI.

Why each option fits or fails:

A. Output modality does not determine whether deep learning is used; the multilayer transformer neural network establishes the deep learning approach.

B. Labeled examples can train deep neural networks, and producing newly composed replies demonstrates generation rather than classification alone.

C. Using approved training replies does not prevent generation; the system creates a newly worded reply rather than selecting stored content.

D. Weight adjustment from examples is machine learning, the multilayer neural architecture is deep learning, and composing new text is generative AI.

Question 45

Topic: Foundation-model applications

A customer-support team fine-tunes a foundation model to classify tickets. Its goal is an overall F1 score of at least 0.82 and an escalation-ticket F1 score of at least 0.65.

All runs use identical settings and the same untouched, representative validation set.

Which action is best supported for the next data-preparation experiment?

Options:

• A. Collect more routine tickets from the existing sources.

• B. Collect more diverse escalation examples from underrepresented sources.

• C. Increase fine-tuning epochs on the 10,000-ticket collection.

• D. Replace the foundation model before changing the dataset.

Best answer: B

Explanation: Learning results should be evaluated by both dataset size and coverage. Increasing the routine-heavy collection from 6,000 to 10,000 tickets did not improve either metric, so merely adding more similar data is unlikely to help. In contrast, adding diverse escalation examples increased escalation F1 from about 0.44 to 0.68 while maintaining the required overall score. This indicates that the main limitation is underrepresentation of relevant escalation cases rather than total corpus size. The next experiment should therefore expand coverage of those cases and continue evaluation on an untouched, representative validation set. More training data is useful when it addresses an observed coverage gap, not simply because the corpus becomes larger.

Why each option fits or fails:

A. Adding same-source routine tickets produced a performance plateau and did not improve escalation-ticket performance.

B. Targeted escalation data substantially improved the weak slice while preserving the required overall performance, showing that relevant coverage is useful.

C. The evidence identifies insufficient escalation coverage, not inadequate training duration, as the limitation addressed by the successful experiment.

D. Replacing the model is not supported because improving dataset coverage already raised escalation performance above the target.

Question 46

Topic: Generative AI fundamentals

An ecommerce company lets an AI agent process refund requests over $500.

Approval rule: A finance approver must inspect the order evidence and exact refund amount. Approval must occur before the payment tool makes the refund effective. The agent may gather evidence and draft the amount while approval is pending.

Current workflow:

Which workflow change satisfies the approval rule?

Options:

• A. Require finance approval of the drafted amount before any payment call.

• B. Require finance approval at intake before evidence gathering and amount drafting.

• C. Require finance approval of the evidence before the amount is drafted.

• D. Require finance approval after payment acceptance but before customer notification.

Best answer: A

Explanation: A human approval checkpoint must govern the action path before a consequential action becomes effective. Here, the agent may safely gather evidence and prepare the proposed refund, but the payment tool call must remain blocked until finance reviews and approves that exact amount. This placement gives the approver the information required by policy and preserves a meaningful intervention path if approval is denied. Reviewing the transaction after payment acceptance may support auditing or recovery, but it cannot serve as authorization because the refund is already effective. The key distinction is between preparing an action and permitting its execution.

Why each option fits or fails:

A. Finance can inspect the evidence and exact amount, while the approval gate prevents the consequential tool call unless permission is granted.

B. Approval at intake cannot be based on the required evidence or be bound to the exact refund amount.

C. Finance cannot approve the exact refund amount before it exists, so this checkpoint does not satisfy the stated review requirement.

D. Payment acceptance has already made the refund effective, so later approval is a review rather than prior authorization.

Question 47

Topic: Generative AI fundamentals

A retailer is selecting a generative model to create structured product descriptions directly from text attributes and a product photo. All models support the required output format.

Mandatory requirements:

• Accept text and image input directly
• Achieve at least 90% reviewer approval
• Provide p95 latency of 2 seconds or less
• Process data only in an approved Region

After meeting these requirements, lower cost is preferred. Results are from the same representative workload; costs are per request.

Which conclusions are supported by the evidence? Select TWO.

Options:

• A. Prefer Delta because low cost and latency outweigh the Region restriction.

• B. Prefer Atlas over Echo because its higher quality overrides the cost preference.

• C. Prefer Echo over Atlas because both qualify and Echo has lower cost.

• D. Exclude Cygnus because its p95 latency exceeds the required maximum.

• E. Prefer Boreal because lower cost and latency compensate for unsupported image input.

Correct answers: C and D

Explanation: Model selection should apply non-negotiable constraints before comparing preferences. Boreal fails the required image modality, Cygnus exceeds the latency limit, and Delta violates the processing-Region restriction. Their favorable quality, latency, or cost results cannot compensate for those failures. Atlas and Echo both satisfy the modality, quality, latency, and Region requirements. Because the scenario treats quality as a threshold rather than a preference and then favors lower cost, Echo is preferred at $0.014 per request. Strong performance on one metric does not establish workload suitability when another mandatory operating constraint is unmet.

Why each option fits or fails:

A. Delta processes data outside the approved Region, which disqualifies it regardless of its favorable cost and latency.

B. Both models exceed the quality threshold, after which the scenario explicitly makes lower cost the deciding preference.

C. Atlas and Echo satisfy every mandatory requirement, so the stated cost preference favors Echo’s lower per-request price.

D. Cygnus leads in quality but exceeds the non-negotiable 2-second p95 latency limit.

E. Boreal cannot directly accept the required image input, and favorable cost or latency cannot offset a mandatory modality failure.

Question 48

Topic: Generative AI fundamentals

A company is designing a travel-planning assistant. Policy checks, inventory searches, and entry-requirement research use distinct data and permissions and can usually run in parallel. Availability conflicts require dynamic replanning.

Pilot results:

Production requires at least 90% success, P95 latency no greater than 25 seconds, and cost no greater than $0.20 per request. Which approach is best supported?

Options:

• A. Use a fixed workflow that calls all tools sequentially.

• B. Deploy the three specialists with coordinator synthesis.

• C. Expand to one specialist per data source with coordinator synthesis.

• D. Keep one agent and optimize its shared context.

Best answer: B

Explanation: Specialist agents add value when work divides into distinct areas of expertise, tasks can proceed independently, and coordination produces a measurable improvement. Here, policy, inventory, and entry-requirement work have separate information and permission needs. The three-specialist design improves success from 84% to 93% while remaining within the latency and cost limits. A coordinator is also justified because availability conflicts require results to be integrated and plans to be revised dynamically. Adding still more agents is not automatically beneficial because each additional handoff can increase latency, cost, duplicated effort, and error propagation. The measured three-specialist architecture is therefore better supported than either the underperforming single agent or an inflexible workflow.

Why each option fits or fails:

A. A fixed sequence does not fit variable tool selection and dynamic replanning, and it prevents independent tasks from running in parallel.

B. This design matches the separable tasks and meets all measured success, latency, and cost requirements.

C. The pilot provides no evidence that further decomposition improves success, while additional handoffs would increase coordination cost and failure paths.

D. The single agent is faster and cheaper, but its measured 84% success rate misses the 90% production requirement.

Question 49

Topic: Foundation-model applications

A company will use 2,000 support records to fine-tune a foundation model. Production serves English and Spanish, retail and enterprise customers, and short and long inputs. Account-lockout escalations are uncommon but consequential and occur in every language and customer segment.

Approved requirement: Represent every usage combination and include at least 100 validated lockout examples spanning all affected language and customer segments. A separate held-out evaluation set will mirror production traffic.

Which preparation action does NOT satisfy the requirement?

Options:

• A. Use 100 English retail lockouts and stratify the remaining routine cases across all combinations.

• B. Start with stratified traffic, then oversample lockouts within each affected segment to 100 total.

• C. Rebalance every usage combination, replacing frequent routine cases with 100 cross-segment lockouts.

• D. Stratify every usage combination and allocate 100 lockout cases across affected segments.

Best answer: A

Explanation: Representativeness applies both to general user groups and to important scenarios within those groups. A dataset can include every group in routine interactions yet remain unrepresentative if consequential lockout examples come only from the dominant segment. The model could improve overall while still performing poorly on Spanish or enterprise lockouts.

Because lockouts occur across language and customer segments, validated examples must span those segments. Oversampling these rare cases is appropriate even when it makes the fine-tuning distribution differ from natural traffic. The separate production-proportional evaluation set can then measure expected deployment performance. Reaching 100 examples alone is insufficient when they are concentrated in one segment.

Why each option fits or fails:

A. The total reaches 100, but English retail examples do not represent lockout behavior in Spanish or enterprise segments.

B. The stratified base represents normal traffic, while within-segment oversampling covers consequential lockout behavior across the served population.

C. Rebalancing limits dominance by routine cases while preserving broad usage coverage and sufficient rare-case representation.

D. This covers ordinary usage combinations and provides the required cross-segment representation of validated lockout cases.

Question 50

Topic: Generative AI fundamentals

A team uses Kiro to draft unit tests and propose a dependency upgrade. Before merging, team policy requires independent developer review of functional correctness, dependency compatibility, and security, followed by meaningful testing. Which proposed action does NOT satisfy the policy?

Options:

• A. Treat Kiro’s explanation as the review, then merge after its generated tests pass.

• B. Inspect the diff, release notes, and scan findings, then run the full CI suite before merging.

• C. Verify generated tests cover edge cases, review compatibility and security risks, then execute them before merging.

• D. Compare behavior and dependency changes with requirements, then run regression and security tests before merging.

Best answer: A

Explanation: Code-generation tools such as Kiro can accelerate drafting tests, explaining code, and proposing dependency changes, but their output remains unverified. A developer must assess whether the change matches requirements, whether dependencies remain compatible, whether known security concerns are addressed, and whether tests cover relevant behavior and edge cases. CI, regression, compatibility, and security testing can provide supporting evidence. Passing only AI-generated tests may miss incorrect shared assumptions or cases the assistant omitted.

AI assistance improves productivity but does not transfer accountability for validation to the tool.

Why each option fits or fails:

A. An assistant’s explanation and its own generated tests do not replace independent developer review of correctness, dependency compatibility, and security.

B. The diff, release notes, and scan findings support the required reviews, while the CI suite tests the proposed change.

C. This action evaluates the generated tests and dependency risks instead of assuming the assistant’s output is already correct.

D. Comparison with requirements provides human review of correctness and compatibility, while regression and security tests provide additional validation evidence.

Questions 51-65

Question 51

Topic: AI and ML fundamentals

A facilities company receives inspection videos containing equipment images, visible asset labels, and technicians’ spoken observations. Asset IDs are already stored as metadata. The company must first create searchable, time-stamped written records of exactly what the technicians said. Which AI capability primarily supports this requirement?

Options:

• A. Automatic speech recognition to convert technician narration into text

• B. Optical character recognition to convert visible asset labels into text

• C. Speech synthesis to convert inspection records into spoken audio

• D. Natural language processing to classify technician observations by topic

Best answer: A

Explanation: The source modality is recorded speech, and the required output is a written record of the spoken words. Automatic speech recognition, also called speech-to-text, directly performs this conversion and can provide timestamps for searchable transcripts. Computer vision and optical character recognition would analyze visual frames or visible text. Natural language processing could later classify, summarize, or extract information from the transcript, while speech synthesis would generate audio from text. The decisive factor is the transformation from spoken audio to written text, not the presence of images or labels in the video.

Why each option fits or fails:

A. Automatic speech recognition processes the audio track and produces the required searchable written representation of the spoken observations.

B. Optical character recognition processes text visible in video frames, but the asset labels are not the required spoken observations.

C. Speech synthesis performs the opposite transformation, producing audio from text rather than producing text from recorded speech.

D. Topic classification can analyze written observations after transcription, but it does not create text from the original audio.

Question 52

Topic: AI and ML fundamentals

A lending team records this workflow:

08:00 Training: labeled historical applications -> M7
09:00 Deployment: M7
10:00 Inference: M7, income 72,000 -> approve
10:05 Input correction: income 52,000
10:06 Inference: M7, income 52,000 -> review
10:07 Corrected application queued for nightly training

Which interpretation is INCORRECT?

Options:

• A. The training run used labeled examples to learn M7.

• B. Both scoring requests used M7 to perform inference.

• C. The corrected input retrained M7 before the second score.

• D. M7 is the learned model artifact deployed for scoring.

Best answer: C

Explanation: Training is the learning procedure that adjusts model parameters from examples. Its output is a learned model artifact, identified here as M7. Inference occurs when that artifact processes a new input to produce a prediction. The second prediction changed because the income value changed, not because M7 was retrained. Queuing the corrected application makes it available for a future training run but does not immediately modify the deployed model. A changed output alone is therefore not evidence that training occurred.

Why each option fits or fails:

A. The 08:00 event applies a learning procedure to labeled historical applications and produces the learned artifact M7.

B. Each scoring event applies the already-trained M7 artifact to an input, which is inference rather than training.

C. M7 remained unchanged during rescoring, while the corrected application was only queued for a later training run.

D. M7 is the output of training and is subsequently deployed to make predictions on application inputs.

Question 53

Topic: Generative AI fundamentals

An online retailer launched a generative AI shopping assistant. Each column covers all website activity during one eight-week period.

A marketing campaign began with the launch, no control group was used, and the assistant’s operating costs have not been compiled.

Which TWO conclusions are supported by the evidence?

Options:

• A. Unique-customer conversion declined from 4.0% to 3.8%.

• B. The comparison does not establish that the assistant caused the revenue increase.

• C. The assistant produced a positive ROI because revenue increased by $112,000.

• D. Customer lifetime value increased because revenue per purchasing customer rose.

• E. The postlaunch unique-customer conversion rate was 4.4%.

Correct answers: A and B

Explanation: Unique-customer conversion must use a consistent population, time period, numerator, and denominator. Here, the rates are 4,000 / 100,000 = 4.0% before launch and 4,940 / 130,000 = 3.8% after launch. Transactions cannot replace unique customers because one customer may complete multiple transactions.

Revenue increased, but the before-and-after comparison does not isolate the assistant’s effect. Traffic also increased, and a marketing campaign began simultaneously. Establishing ROI would additionally require attributable financial benefit and all relevant costs, preferably using profit or contribution margin rather than revenue alone. Revenue per purchasing customer over eight weeks also does not establish customer lifetime value.

Usage and revenue growth alone do not prove causation or positive ROI.

Why each option fits or fails:

A. Dividing unique purchasing customers by unique visitors within each period gives 4.0% before launch and 3.8% after launch.

B. The simultaneous marketing campaign and absence of a control group prevent the revenue increase from being attributed specifically to the assistant.

C. Positive ROI cannot be established without attributable financial benefit and relevant costs, including the assistant’s uncompiled operating costs.

D. Revenue per purchasing customer during one eight-week period does not measure the customer’s expected value over the full relationship.

E. The 4.4% figure uses transactions rather than unique purchasing customers, so it is a transaction rate rather than customer conversion.

Question 54

Topic: Security, compliance and governance

An agent proposes a $25,000 supplier payment using an approved payment tool.

Execution evidence:

The agent states that the manager’s approval permits it to proceed. Which decision is supported by the evidence?

Options:

• A. Retry using the approving manager’s session because approval conveys delegated authority.

• B. Execute now because valid business approval overrides the resource-policy denial.

• C. Obtain a second human approval, then retry using the same agent identity.

• D. Do not execute until an effectively authorized identity submits the approved request.

Best answer: D

Explanation: Authorization and human approval are independent controls. The matching manager approval satisfies the required business workflow, but effective authorization still fails because the resource policy explicitly denies payments over $10,000 for the agent role. An identity-policy allow does not override that explicit deny, and neither a human approval nor the agent’s statement can grant permissions. The payment must remain blocked until it is submitted through a properly authorized execution path while retaining the required approval evidence. Conversely, effective tool permission alone would not replace a missing business approval for a consequential action.

Why each option fits or fails:

A. Approving a transaction does not automatically delegate the manager’s credentials or permissions to the agent.

B. Human approval cannot override an authorization denial imposed on the identity attempting the tool action.

C. Additional approval does not resolve the explicit deny affecting the agent identity’s effective permissions.

D. The approval satisfies the business workflow, but the explicit resource-policy deny prevents the agent role from executing the payment.

Question 55

Topic: Responsible AI

A retailer is preparing an intent classifier for support chats in Mexico. The expected traffic and current evidence are shown below.

Which action is best supported before rollout?

Options:

• A. Collect local Spanish chats for training and evaluate on a separate local test set.

• B. Translate more US English chats and evaluate on a larger translated Spanish set.

• C. Retain current training data and reweight test results using expected traffic shares.

• D. Collect more US English chats and evaluate each existing subgroup separately.

Best answer: A

Explanation: Training and evaluation data should reflect the population and conditions where the model will be used. Mexico Spanish users represent 40% of expected traffic but only 5% of both datasets, and those records are translations rather than local conversations. The lower observed accuracy also signals that performance may differ for this group, although the translated test set cannot reliably measure real deployment performance. The retailer should obtain production-like Mexico Spanish interactions for training and reserve separate, representative local data for evaluation. Adding more examples from the same US source, even after translation, increases sample size without closing the representation gap. Reweighting similarly cannot repair unrepresentative evidence.

Why each option fits or fails:

A. Local, production-like data would improve training coverage and provide independent evidence for the population representing 40% of intended use.

B. A larger translated sample reduces sampling uncertainty but preserves the source bias and may miss regional language and intent patterns.

C. Reweighting changes the aggregate calculation but cannot make translated subgroup evidence representative or correct the training-data coverage gap.

D. Separate reporting could reveal the gap, but additional US English data would not improve coverage of Mexico Spanish users or local usage conditions.

Question 56

Topic: AI and ML fundamentals

A quality-control team tests four image-classification models on the same representative day of labeled production data. Missing defective products creates a safety risk.

Operating requirements:

• Recall must be at least 90%.
• Review capacity is 300 alerts per day.
• Every predicted positive generates an alert.

Which model best satisfies both operating requirements?

Options:

• A. Model Gamma

• B. Model Beta

• C. Model Alpha

• D. Model Delta

Best answer: C

Explanation: Recall measures the proportion of actual positive cases successfully detected: true positives divided by true positives plus false negatives. Each model is evaluated on 200 actual defective products. Alpha detects 180, giving 90% recall. It also generates 280 alerts because alert volume equals true positives plus false positives. Beta and Delta exceed the recall threshold but create more than 300 alerts. Gamma stays within review capacity but misses too many defective products. High recall can therefore coexist with an unsustainable number of false positives; operational selection must consider both missed-case consequences and review capacity.

Why each option fits or fails:

A. Gamma generates only 260 alerts, but its 85% recall does not meet the required detection rate.

B. Beta achieves 95% recall, but its 400 alerts exceed the daily review capacity.

C. Alpha achieves 90% recall and generates 280 alerts, meeting both the detection threshold and review-capacity limit.

D. Delta achieves 92.5% recall, but its 315 alerts exceed the daily review capacity.

Question 57

Topic: Foundation-model applications

A company is selecting how to maintain an internal policy assistant. Policies change weekly, and the assistant’s response style is already acceptable. Each approved revision must be available within 24 hours, and at least 90% of answers must cite a supporting clause.

Which deployment approach is best supported by the evidence?

Options:

• A. Maintain policy excerpts in a few-shot template and revise it after changes.

• B. Continue pretraining on the policy archive and request citations during inference.

• C. Deploy RAG over approved policies and refresh the index after changes.

• D. Fine-tune the model on each revision and generate citations during inference.

Best answer: C

Explanation: Retrieval-augmented generation (RAG) connects the assistant to an approved, updateable knowledge source and supplies relevant passages during inference. The retrieval pilot makes revisions available in four hours and achieves 94% valid citations, satisfying both requirements. Updating the index changes the available evidence without changing the foundation model’s learned parameters. Fine-tuning and continued pretraining modify model behavior or knowledge in its weights, but the results show slower updates and weak attribution. Few-shot excerpts provide acceptable citations but miss the freshness requirement by six hours. Therefore, retrieval is the best mechanism for current, attributable organizational knowledge.

Why each option fits or fails:

A. Few-shot prompting meets the citation threshold, but its 30-hour revision delay exceeds the required 24-hour limit.

B. Continued pretraining takes seven days and does not reliably ground citations in the applicable policy clauses.

C. Retrieval-augmented generation meets both the 24-hour availability requirement and the 90% valid-citation threshold without retraining model weights.

D. Fine-tuning takes three days and produces only 42% valid citations, so it fails both stated requirements.

Question 58

Topic: AI and ML fundamentals

A company must automate travel reimbursements. The authoritative policy has 18 documented rules:

• A claim is eligible only if its purpose, category, approval, and submission date satisfy specified conditions.
• Payment equals the lower of the verified amount and the applicable location cap.
• Auditors require identical results for identical validated inputs and a reference to the governing rule.

The company has 80,000 historical decisions, but the policy owner changes rules and caps twice yearly. Which TWO conclusions are supported? Select TWO.

Options:

• A. Use fixed model settings to make learned inference equivalent to policy execution.

• B. Use deterministic logic for eligibility and payment decisions.

• C. Use a supervised classifier trained on prior approvals for eligibility.

• D. Use a regression model trained on prior payments for reimbursement amounts.

• E. Use versioned rule updates and automated tests after policy changes.

Correct answers: B and E

Explanation: An authoritative policy with precisely defined conditions and calculations is a deterministic specification, not a relationship that must be learned. Explicit logic can produce the required result, identify the governing rule, and be tested directly when requirements change. Historical examples could contain errors or reflect previous policy versions, so models trained on them would approximate decisions rather than establish compliance. Fixed inference settings may improve repeatability, but repeating a prediction is not equivalent to executing the authoritative rules. AI can still assist with tasks such as extracting receipt fields or flagging unusual claims, while validated inputs and deterministic logic control the final decision.

Why each option fits or fails:

A. Repeatable inference can reproduce a learned prediction, but it does not prove that the prediction implements every policy rule exactly.

B. Deterministic logic directly implements the authoritative conditions and calculation, producing traceable results from validated inputs.

C. A classifier would approximate historical decisions rather than reliably execute the current authoritative eligibility conditions.

D. Regression estimates numeric outcomes, but the required payment is an exact calculation defined by the policy.

E. Explicit rule maintenance aligns updates with policy changes and avoids retraining a learned approximation whenever requirements change.

Question 59

Topic: Foundation-model applications

A retailer runs a controlled 30-day pilot of a foundation-model shopping assistant with comparable traffic in both periods.

Requirement: Increase completed purchases while maintaining or improving successful task resolution for every customer group.

Which interpretation is NOT supported by the pilot results?

Options:

• A. Higher conversion shows sales progress despite the subgroup resolution decline.

• B. Higher repeat use reflects engagement, not necessarily business success.

• C. Longer conversations establish greater value delivered during each interaction.

• D. Higher satisfaction reflects favorable perceptions, not necessarily task completion.

Best answer: C

Explanation: Satisfaction, engagement, and business outcomes measure different aspects of an application. Satisfaction scores capture user perceptions, while repeat use and conversation length describe engagement. Neither category alone establishes successful task completion or business value. Purchase conversion directly supports the retailer’s sales goal, but the decline in resolution for screen-reader users shows that the complete requirement was not met. Longer conversations are especially ambiguous: they may indicate useful engagement, but they can also reflect confusion or repeated attempts to obtain a satisfactory response. Business evaluation should therefore combine goal-aligned outcomes with task-success and subgroup measures rather than treating increased interaction as proof of value.

Why each option fits or fails:

A. Purchase conversion directly measures progress toward completed purchases, although the screen-reader subgroup result means the full requirement was not achieved.

B. Repeat use indicates continued engagement, but it does not independently show completed purchases, successful resolutions, or other business outcomes.

C. Conversation length is ambiguous because additional turns can result from confusion, recovery attempts, or difficulty completing a task.

D. Satisfaction captures users’ reported perceptions and can remain favorable even when some users do not successfully resolve their tasks.

Question 60

Topic: Foundation-model applications

A company uses the same foundation model for two applications:

• A document assistant must consistently extract records into a fixed structure.
• A marketing assistant must produce varied brainstorming candidates.

The model documentation states that temperature ranges from 0.0 to 1.0, with higher values increasing variation. Each application can use separate inference settings, and outputs are validated independently.

Which initial configuration best aligns with these goals?

Options:

• A. Set both to 0.8 and add extraction examples.

• B. Set both to 0.1 and lengthen brainstorming responses.

• C. Set extraction to 0.8 and brainstorming to 0.1.

• D. Set extraction to 0.1 and brainstorming to 0.8.

Best answer: D

Explanation: Temperature influences how broadly a model samples possible continuations. Lower temperature favors more probable continuations and generally improves consistency, making it suitable for repeatable document extraction. Higher temperature allows less probable continuations to be selected more often, which can increase variety for brainstorming. Temperature does not guarantee identical output, valid structure, or factual correctness, so independent validation remains necessary. Its exact behavior and supported range also depend on the selected model. Prompt examples may improve format adherence, but they do not eliminate the variability introduced by a high temperature.

Why each option fits or fails:

A. Examples can reinforce the extraction format, but high temperature still introduces unnecessary variation into the structured task.

B. Longer responses can contain more content, but they do not provide the sampling diversity encouraged by a higher temperature.

C. This reverses the desired behavior by increasing extraction variability while reducing diversity in brainstorming responses.

D. Low temperature supports repeatable extraction structure, while higher temperature promotes the variation desired for brainstorming candidates.

Question 61

Topic: Generative AI fundamentals

On February 10, 2026, a support assistant must answer:

Is order C-184 eligible for an expedited refund, and by what date must the request be submitted?

The requester can access Support-General and case C-184, but not Finance-Restricted. Authorization filtering occurs before records enter the model context, which is limited to two records.

Which TWO context-selection decisions are supported by the exhibit? Select TWO.

Options:

• A. Admit P0 for an approved expedited-refund rule.

• B. Admit C1 for the authorized order-specific facts.

• C. Admit F1 for the current public refund timing.

• D. Admit P1 for the current expedited-refund rule.

• E. Admit D1 for the newest proposed refund window.

Correct answers: B and D

Explanation: Context selection should consider relevance, authority, freshness, and authorization before information reaches the model. P1 supplies the current approved eligibility rule, while C1 supplies the authorized facts needed to apply that rule. Together, they fit the two-record budget and provide evidence for both eligibility and the deadline.

A newer document is not automatically suitable: D1 is only a draft and is unauthorized. Redacting its content from the final response would not correct exposing it to the model. F1 is current and accessible but does not address the requested decision, while P0 contains stale guidance. The goal is the smallest authorized evidence set sufficient for the business request.

Why each option fits or fails:

A. P0 is accessible and approved but has been superseded, making its 14-day rule stale for the current request.

B. C1 is accessible and supplies the purchase date, payment method, and refund status needed to apply the policy.

C. F1 describes standard processing duration, not expedited-refund eligibility or the order-specific submission deadline.

D. P1 is the current approved policy and provides the authoritative eligibility period needed to determine the submission deadline.

E. D1 is an unapproved draft outside the requester’s permissions, so it must be excluded before model context construction.

Question 62

Topic: Generative AI fundamentals

A brand wants to turn licensed product photos and approved translated copy into short social videos. The videos must animate the photos, preserve packaging and logo details, and narrate the copy. Which approach best fits the task and provides the evidence needed before publication?

Options:

• A. Use image-conditioned video generation and text-to-speech; verify AI-derivative permissions, packaging fidelity, and claims against approved sources.

• B. Use image-conditioned video generation and text-to-speech; verify visual realism, narration fluency, and favorable audience reactions.

• C. Use visual media analysis and text-to-speech; verify AI-derivative permissions, detected labels, and claims against approved sources.

• D. Use text-only video generation and text-to-speech; verify AI-derivative permissions, packaging fidelity, and claims against approved sources.

Best answer: A

Explanation: Image-conditioned video generation fits a task that must animate specific source photos while retaining their recognizable details. Text-to-speech can produce narration from the approved translated copy. Before public use, the brand should confirm that the photo licenses and relevant service terms permit the intended AI-derived marketing content. Reviewers should also compare packaging, logos, narration, and product claims with the approved source assets and copy. A polished or convincing result is not evidence that its details are accurate or that the organization has permission to publish it. Visual analysis is useful for examining existing media, but it does not perform the requested media creation.

Why each option fits or fails:

A. Image-conditioned generation uses the permitted source assets, while rights verification and source-based review address publication permissions, visual fidelity, and factual accuracy.

B. Realism, fluency, and audience reactions do not establish permission for AI-derived media or accuracy against the approved product sources.

C. Visual analysis can identify content in existing photos, but it does not create the required animated video from those assets.

D. Text-only generation does not directly edit the licensed photos, making reliable preservation of their specific packaging and logo details less likely.

Question 63

Topic: Generative AI fundamentals

A retailer compares 100 manually written product descriptions with a generative AI pilot covering 100 descriptions of matched complexity and traffic. Management wants to know whether total effort decreased without reducing quality. Labor for post-publication corrections was not recorded.

Which interpretation is best supported by the evidence?

Options:

• A. Report no labor saving, since additional checking and rework consumed the entire drafting-time reduction.

• B. Report a net 45-hour labor saving, while treating checking time as a separate quality measure.

• C. Report a provisional 15-hour prepublication saving, pending measurement of correction effort and quality impact.

• D. Report a net 15-hour labor saving, while tracking correction counts separately as a quality measure.

Best answer: C

Explanation: Productivity should be assessed across the complete workflow, not only the generated first draft. For the same output volume, drafting decreased by 45 hours, checking and rework increased by 30 hours, and publication time remained unchanged. This produces a 15-hour prepublication saving. However, post-publication corrections increased from 2 to 7, indicating a possible quality decline and additional work whose duration was not measured. The pilot therefore shows faster production through publication, but it does not yet establish a net productivity gain or maintained quality. Correction effort, severity, and business impact must be included in the final comparison.

Why each option fits or fails:

A. Drafting fell by 45 hours while checking and rework rose by 30 hours, leaving a 15-hour saving through publication.

B. The drafting reduction excludes the 30-hour increase in checking and rework, which is part of the content-production workflow.

C. Labor through publication fell from 100 to 85 hours, but increased corrections and unmeasured correction effort prevent a complete productivity conclusion.

D. Post-publication correction work belongs in the complete workflow, so separating it cannot establish a net saving when its effort is unknown.

Question 64

Topic: Foundation-model applications

On August 15, 2025, an evaluator reviews a grounded support response. Only the retrieved context was provided to the model.

Which conclusions does the evidence support? Select TWO.

Options:

• A. The 30-day limit is factually correct but unfaithful to the retrieved passage.

• B. The full response is faithful because its receipt requirement matches the passage.

• C. The entire response is factually incorrect because the cited article is outdated.

• D. Citation [1] supports the receipt requirement but not the 30-day limit.

• E. The evidence shows that retrieval selected the current authoritative policy record.

Correct answers: A and D

Explanation: Factual correctness compares a claim with the authoritative real-world state. Faithfulness checks whether the generated claim is supported by the context provided to the model. Citation support checks whether the cited passage actually supports the associated claim.

Here, the complete response matches the current policy, so it is factually correct. However, the model received an archived FAQ specifying 45 days, making the generated 30-day limit unfaithful to its context. Citation [1] supports the receipt requirement but contradicts the return window. A response can therefore be correct while still being insufficiently grounded or only partially supported by its citation.

Why each option fits or fails:

A. The current policy establishes 30 days, while the model’s retrieved passage states 45 days, separating correctness from faithfulness.

B. Matching the receipt condition does not establish full faithfulness because the generated 30-day limit conflicts with the retrieved 45-day limit.

C. An outdated citation does not make the content incorrect; the generated return window and receipt condition match the current policy.

D. The cited FAQ requires a receipt but specifies 45 days, so it supports only part of the generated sentence.

E. The model received only the archived FAQ; the current policy was evaluator evidence rather than retrieved model context.

Question 65

Topic: AI and ML fundamentals

A payment company wants to prioritize 500 daily transactions for fraud investigation.

Evidence:

Which approach is best supported by this evidence?

Options:

• A. Rank transactions with a supervised classifier trained on transaction attributes.

• B. Rank transactions with unsupervised anomaly detection for investigator review.

• C. Rank transactions with a classifier trained on rarity-based fraud labels.

• D. Rank transactions by membership in the smallest discovered cluster.

Best answer: B

Explanation: Unsupervised anomaly detection is appropriate when historical records have useful attributes but lack confirmed outcome labels. It can assign anomaly scores based on how much transactions differ from learned patterns, allowing the company to prioritize a limited investigation queue. The scores do not prove fraud because rare transactions can be legitimate, as the evidence explicitly indicates. Investigators therefore provide the human follow-up needed to determine outcomes and can create reliable labels for possible future supervised learning.

The key distinction is that anomaly detection identifies unusual activity, while fraud classification requires evidence connecting activity to confirmed fraud outcomes.

Why each option fits or fails:

A. A supervised fraud classifier requires historical outcome labels that identify fraudulent and legitimate training examples, but those labels are unavailable.

B. Anomaly detection can identify unusual attribute patterns without labeled outcomes, while investigators determine whether the prioritized transactions are actually fraudulent.

C. Creating fraud labels from rarity assumes that unusual activity is harmful, reproducing the unsupported conclusion the investigation should determine.

D. The smallest cluster may represent uncommon but legitimate behavior, so cluster size alone is not a reliable investigation priority.

Record your review

Review updates and feedback for the release history or to report a specific question. An active IT Mastery pass or subscription includes the complete current AIF-C01 bank and all other IT Mastery banks during the access period.

Continue in the web app

Use IT Mastery for interactive AWS AIF-C01 practice with mixed sets, timed mocks, topic drills, explanations, and progress tracking.

Try AWS AIF-C01 on Web