cert
cert-1
cert-2

Pass Microsoft AI-103 Exam in First Attempt Guaranteed!

Get 100% Latest Exam Questions, Accurate & Verified Answers to Pass the Actual Exam!
30 Days Free Updates, Instant Download!

cert-5
cert-6
AI-103 Exam - Verified By Experts
AI-103 Premium File

AI-103 Premium File

$89.99
$98.99
  • Premium File 187 Questions & Answers. Last Update: Nov 15, 2026

Whats Included:

  • Latest Questions
  • 100% Accurate Answers
  • Fast Exam Updates
 
$98.99
$89.99
accept 397 downloads in the last 7 days
block-screenshots
Microsoft AI-103 Practice Test 48

Question 470

An agent revises an answer endlessly with small wording changes. What control limits reflection overhead?

  1. Allow unlimited critique-and-rewrite cycles until the model says its answer is flawless.
  2. Check cited statements against the retrieved source passages
  3. Cap critique-and-revision passes using quality improvement thresholds
  4. Record observable rule violations, test inputs and trace spans rather than internal hidden reasoning

Explanation

Correct Answer

C:

Bounded reflection prevents token waste when repeated attempts no longer improve output. The practical requirement is to stop unproductive self-critique loops; a solution built to repair generated code from actual test failures would not meet that same need.

Incorrect Answers

A:

Unbounded reflection can waste cost and repeatedly change phrasing without improvement. Apply a stopping budget, measurable tests and a termination condition.

B:

A second opinion is not enough when both share the same fabricated evidence. This helps when a team needs to validate citations at the source level; it does not establish how to stop unproductive self-critique loops.

D:

External events and policy criteria can explain failures without requiring private reasoning traces. That capability is appropriate for audit model failures using observable evidence, not this separate task of finding how to stop unproductive self-critique loops.

Question 471

A new agent uses three tools and an evaluator labels its final answer accurate, but the tool path was unsafe. What should be evaluated separately?

  1. Measure answer completeness and task relevance against expected criteria
  2. Evaluate only the final written answer and ignore the tools invoked and their side effects.
  3. Version evaluation datasets, scoring rubrics and results alongside release artifacts
  4. Inspect agent trajectory and policy compliance in addition to final-answer quality

Explanation

Correct Answer

D:

A correct final sentence does not establish that intermediate tools were used safely or with authorization. This directly supports the need to evaluate agent actions as well as final responses, unlike a mechanism primarily intended to evaluate whether answers actually fulfill the task.

Incorrect Answers

A:

Citation correctness is not sufficient if the response does not fulfill the information need. The function serves to evaluate whether answers actually fulfill the task, whereas the actual requirement is to evaluate agent actions as well as final responses.

B:

A plausible final answer can follow unauthorized or unsafe tool use. Inspect tool selection, permissions and execution traces separately from answer accuracy.

C:

Fixed benchmarks and recorded methods make regressions reproducible across candidates. Here the needed outcome is to evaluate agent actions as well as final responses; an option designed to make evaluation runs reproducible over time would not directly provide it.

Question 472

A system uses an LLM to authorize an irreversible refund solely because the prompt says 'manager approved.' What is missing?

  1. Accept the prompt phrase manager approved as evidence of authorization.
  2. Have the LLM sign its own approval record and execute the refund.
  3. Perform the refund first and query the approval database afterward.
  4. Validate approval against a trusted workflow record before running the refund tool.

Explanation

Correct Answer

D:

A claimed approval in user text is not verified authorization for a side effect; backend records must confirm it.

Incorrect Answers

A:

User text may be inaccurate or adversarial; it cannot confer backend approval.

B:

An LLM-created approval record is not evidence that an authorized manager consented. Verify the approver's identity and decision through a trusted approval workflow before invoking the irreversible refund function.

C:

A post-refund database check cannot prevent an unauthorized irreversible payment. Obtain and verify the required manager approval before the write, preserving an auditable authorization reference.

Question 473

An engineer needs a repeatable report of model quality before each release. Which practice is appropriate?

  1. Run privacy-preserving subgroup performance and fairness checks on appropriate test data
  2. Use groundedness or faithfulness evaluation against the source document
  3. Run one informal evaluator prompt and publish its result without saving dataset or model versions.
  4. Version evaluation datasets, scoring rubrics and results alongside release artifacts

Explanation

Correct Answer

D:

Fixed benchmarks and recorded methods make regressions reproducible across candidates. The decisive difference is whether the design can make evaluation runs reproducible over time; the competing function serves to check fairness across relevant cohorts.

Incorrect Answers

A:

Aggregate subgroup metrics can expose disparate performance without publishing individual records. That capability is appropriate for check fairness across relevant cohorts, not this separate task of finding how to make evaluation runs reproducible over time.

B:

A fluent summary can still invent claims; groundedness tests whether assertions are supported by the reference. Here the needed outcome is to make evaluation runs reproducible over time; an option designed to score source-supported generated statements would not directly provide it.

C:

An unversioned ad hoc test cannot support repeatable release comparisons. Maintain a representative evaluation set, scoring criteria and versioned results.

Question 474

A safety reviewer needs to understand why generated output failed a policy test without collecting hidden private reasoning. What evidence should be requested?

  1. Check cited statements against the retrieved source passages
  2. Record observable rule violations, test inputs and trace spans rather than internal hidden reasoning
  3. Use independent source-grounded checks rather than relying solely on self-confidence
  4. Request and store the model’s hidden private reasoning as the compliance record for every policy rejection.

Explanation

Correct Answer

B:

External events and policy criteria can explain failures without requiring private reasoning traces. Here the relevant capability must let the team audit model failures using observable evidence. An alternative targeting validate citations at the source level solves another problem.

Incorrect Answers

A:

A second opinion is not enough when both share the same fabricated evidence. This helps when a team needs to validate citations at the source level; it does not establish how to audit model failures using observable evidence.

C:

Self-critique without external evidence can reinforce the same mistaken claim. That capability is appropriate for verify model assertions against independent sources, not this separate task of finding how to audit model failures using observable evidence.

D:

Hidden reasoning is not an appropriate or reliably available audit artifact. Record observable inputs, policy rule identifiers and test results needed to reproduce the failure.

Question 475

A compliance review must reconstruct which model deployment and policy rules produced a customer recommendation. What evidence should be retained?

  1. Store only the final recommendation text without release or source identifiers.
  2. Record immutable deployment version, policy version, authorized source IDs and correlated tool decisions.
  3. Retain the model deployment name but discard its revision and tool events.
  4. Collect unrelated application page-view analytics instead of decision provenance.

Explanation

Correct Answer

B:

Traceable provenance makes high-impact outputs reproducible and reviewable without pretending a plain final answer contains the full decision history.

Incorrect Answers

A:

The recommendation alone cannot identify which policy and model configurations led to it.

C:

A friendly name may change while source and policy decisions are essential to reconstruction.

D:

Traffic metrics cannot reconstruct the model version and evidence that affected an individual case.

Question 476

A model produces code with no error handling for a tool timeout. How can a targeted reflection rubric help?

  1. Use independent source-grounded checks rather than relying solely on self-confidence
  2. Regenerate the entire answer without exercising the timed-out tool path or checking recovery behavior.
  3. Check generated code against explicit failure-mode and retry-safety criteria
  4. Check cited statements against the retrieved source passages

Explanation

Correct Answer

C:

Rubric-based critique focuses on known reliability requirements instead of vague style suggestions. This directly supports the need to identify missing resilience behaviors in code review, unlike a mechanism primarily intended to verify model assertions against independent sources.

Incorrect Answers

A:

Self-critique without external evidence can reinforce the same mistaken claim. The function serves to verify model assertions against independent sources, whereas the actual requirement is to identify missing resilience behaviors in code review.

B:

A generic rewrite does not prove that timeout handling now works. Use a targeted rubric and test that simulates tool delay, retry and graceful failure.

D:

A second opinion is not enough when both share the same fabricated evidence. Its primary purpose is to validate citations at the source level; the stated constraint instead demands a way to identify missing resilience behaviors in code review.

Question 477

A high-risk assistant must support review of its evidence, compare outputs fairly, and assess unsafe tool behavior. Which THREE evaluation layers are appropriate? Choose THREE.

  1. Run privacy-preserving subgroup performance and fairness checks on appropriate test data
  2. Provide traceable source references and decision-factor provenance for review
  3. Inspect agent trajectory and policy compliance in addition to final-answer quality
  4. Treat automatically generated evaluator scores as independent truth without any reference examples or validation against reviewed outcomes.
  5. Discard evaluation dataset and scoring-rubric versions after each run so results cannot be reproduced or compared.
  6. Count source citations without testing whether any cited passage supports the actual claim.

Explanation

Correct Answers

A:

Aggregate subgroup metrics can expose disparate performance without publishing individual records. Traceable evidence, responsible subgroup evaluation and trajectory inspection cover explanation, fairness and agent process safety.

B:

Auditable evidence and factor summaries explain why the system recommended an outcome without exposing private hidden reasoning.

C:

A correct final sentence does not establish that intermediate tools were used safely or with authorization.

Incorrect Answers

D:

Uncalibrated evaluator output does not establish fairness, explain decisions or demonstrate safe tool behavior for a high-risk assistant.

E:

Without a versioned dataset and rubric, the team cannot make reproducible cross-release assessments or audit fairness and safety findings.

F:

A raw citation count is not evidence of grounded claims; it can reward fabricated references instead of validating them.

Question 478

A team wants to know whether a repair prompt truly improved answers rather than just their fluency. What testing method fits?

  1. Re-run a held-out evaluation set with measured before-and-after outcomes
  2. Declare prompt repair successful because the revised answer sounds more confident
  3. Treat longer repaired answers as improved accuracy without testing the original failures.
  4. Cap critique-and-revision passes using quality improvement thresholds

Explanation

Correct Answer

A:

Independent examples reveal whether revision improves correctness beyond cosmetic changes. The decisive difference is whether the design can evaluate objective improvement after a critique pass; the competing function serves to verify model assertions against independent sources.

Incorrect Answers

B:

Confidence of prose is not a measurement of factual improvement. A controlled held-out evaluation with before-and-after outcomes is required.

C:

Fluency and verbosity are not correctness evidence. Re-run held-out failure cases and compare objective task outcomes after the repair.

D:

Bounded reflection prevents token waste when repeated attempts no longer improve output. That is relevant to stop unproductive self-critique loops, but the present scenario requires an ability to evaluate objective improvement after a critique pass.

Question 479

An AI service unexpectedly edited a customer record and investigators need to identify the triggering request. What audit linkage is essential?

  1. Save retrieval document IDs and timestamps but omit the authenticated caller, action tool name and write-event trace that changed the customer record.
  2. Use a human-in-the-loop draft-and-approve workflow for writes
  3. Trigger an emergency stop for future agent actions and delete the request, actor and tool-call logs needed to investigate the earlier edit.
  4. Record the request trace ID alongside authenticated actor and tool-write event

Explanation

Correct Answer

D:

Correlated IDs and identities let reviewers reconstruct which specific request produced an action. This resolves the specific need to attribute a tool side effect to a request and actor, while a measure that helps record knowledge provenance for replay addresses a different requirement.

Incorrect Answers

A:

Retrieval lineage can identify supporting context but does not prove which authorized caller or tool wrote the record. Correlate a request trace and actor with the actual side-effect event.

B:

Approval gates preserve human control over irreversible or regulated output. That is relevant to stage side effects until review approval, but the present scenario requires an ability to attribute a tool side effect to a request and actor.

C:

A kill switch can prevent additional writes, but deleting traces destroys the causal audit trail. Preserve immutable request IDs, principal identity, action details and write outcomes for investigation.

Question 480

An app receives model responses from multiple regions and must determine whether one region is causing more safety refusals. What monitoring key helps?

  1. Label refusal metrics by deployment region, model version and safety-policy version.
  2. Compare regional refusal counts without accounting for model and policy versions.
  3. Monitor only global refusal totals and hide the deployment region label.
  4. Track server uptime alone and infer that safety policies are identical.

Explanation

Correct Answer

A:

A single global refusal total obscures regional configuration and version-specific behavior.

Incorrect Answers

B:

A higher regional total may reflect different traffic or settings rather than location-specific behavior.

C:

Aggregation removes the dimension needed to diagnose one region's higher rejection rate.

D:

Availability does not reveal refusal distribution or safety-filter configuration drift.

Developing AI Apps and Agents on Azure (AI-103) Practice Tests