McAfee-Secured Website

Exam Code: AI-103

Exam Name: Developing AI Apps and Agents on Azure

Certification Provider: Microsoft

Corresponding Certification: Microsoft Certified: Azure AI Apps and Agents Developer Associate

100% Microsoft AI-103 Guaranteed Success With Testking Training Materials

Get Real AI-103 Exam Questions & Expert Verified Answers as Experienced on the Real Exam!

187 Questions & Answers with Testing Engine

"Developing AI Apps and Agents on Azure Exam", also known as AI-103 exam, is a Microsoft certification exam.

Pass your tests with the always up-to-date AI-103 Exam Engine. Your AI-103 training materials keep you at the head of the pack!

Satisfaction Guaranteed

Money Back Guarantee

Testking provides hassle-free money back guarantee with our products. That is because we have 100% trust in the abilities of our professional and experience product team, and our record is a proof of that.

99.6% PASS RATE
Was: $137.49
Now: $124.99

Microsoft AI-103 Practice Test 66

Question 650

A research agent retrieves a web page containing a list of product specifications and a hidden instruction to send internal documents to an outside endpoint. The legitimate task is to compare specifications. What must the agent do?

  1. Treat the hidden instruction as untrusted page data and restrict tools to the authorized comparison.
  2. Follow the hidden instruction because the search tool returned the page successfully.
  3. Authorize the outbound send if the page appears among the first search results.
  4. Use a higher reasoning model and rely on it to decide whether to leak documents.

Explanation

Correct Answer

A:

External page text is not an instruction source that can expand the agent’s permission scope.

Incorrect Answers

B:

Successful retrieval does not convert document text into trusted commands.

C:

Retrieval rank reflects relevance, not user approval for exfiltration.

D:

Model capability cannot replace explicit tool authorization and trust boundaries.

Question 651

A triage agent classifies an IT incident and hands it to an infrastructure agent. The second agent must know the ticket ID, severity and validated evidence, but not unrelated conversation history. Which handoff design fits?

  1. Send the entire raw chat transcript regardless of its personal or unrelated content.
  2. Pass only the triage agent’s self-assessed confidence score and omit the ticket ID.
  3. Pass a free-form summary and let the specialist infer missing ticket fields if necessary.
  4. Pass a typed minimal handoff record containing the necessary ticket fields and evidence references.

Explanation

Correct Answer

D:

Explicit state transfer avoids unnecessary exposure and gives the specialist a reliable task contract.

Incorrect Answers

A:

Unbounded history may expose data irrelevant to the second agent’s work.

B:

A score alone does not identify which incident the specialist should handle.

C:

Unvalidated inference can lose the ticket identifier or alter required incident facts.

Question 652

An agent identifies a cloud resource with high monthly costs. It can inspect usage but requires human approval to delete the resource. How should the planning and execution tools be configured?

  1. Grant the same unrestricted deletion tool to the initial read-only analysis agent.
  2. Allow read-only diagnosis and require a separately verified approval before deletion.
  3. Assume cost-saving intent implies permission to delete resources without confirmation.
  4. Hide the deletion tool in the interface but keep it callable by any analysis response.

Explanation

Correct Answer

B:

The analysis step can proceed without allowing an unauthorized irreversible side effect.

Incorrect Answers

A:

This bypasses the stated approval boundary between investigation and mutation.

C:

A goal to reduce spend does not authorize any specific destructive operation.

D:

Interface visibility is not enforcement of tool-side access control.

Question 653

A team compares two agent model versions, but different test prompts are used for each. The scores differ and stakeholders want to attribute the change to the model. What comparison method is sound?

  1. Use a smaller easier dataset for the new model to demonstrate improved results.
  2. Report average token count as the only indicator of factual answer quality.
  3. Evaluate both versions on the same representative held-out dataset and metric definitions.
  4. Ignore whether tool mocks and source data changed between evaluation runs.

Explanation

Correct Answer

C:

Controlling test inputs permits a meaningful attribution of performance differences to the deployment change.

Incorrect Answers

A:

Changing difficulty confounds the comparison and can create a false improvement.

B:

Token consumption is not a substitute for correctness or task success.

D:

Different dependencies can alter outcomes independently of the model version.

Question 654

A customer-support agent’s new prompt reduces average latency but causes a measurable increase in unauthorized refund attempts. Which release response meets the safety requirements?

  1. Roll back the prompt change and investigate the refund-action failures before retrying release.
  2. Keep the new prompt because lower latency outweighs any authorization defect.
  3. Disable refund-action logging while keeping the prompt in production.
  4. Increase the number of simultaneous sessions to dilute the failed cases in the average.

Explanation

Correct Answer

A:

An action-policy regression is material even when a separate latency metric improves.

Incorrect Answers

B:

Operational speed cannot justify increased unauthorized financial actions.

C:

Suppressing observations conceals rather than removes the safety regression.

D:

Additional traffic does not repair incorrect action authorization.

Question 655

An evaluation pipeline uses a language model to rate agent responses for groundedness. Its ratings change on repeated runs, even for fixed examples. What validation improves confidence?

  1. Treat the first model-generated rating as an immutable ground truth for every future run.
  2. Increase the evaluator temperature without testing score variance or agreement.
  3. Remove the source evidence while asking the evaluator to judge factual grounding.
  4. Calibrate the evaluator against fixed reviewed examples and examine repeatability on a representative sample.

Explanation

Correct Answer

D:

Reference judgments and stability checks reveal unreliable automated grading behavior.

Incorrect Answers

A:

A probabilistic grader can misjudge cases, so first output is not proof.

B:

Greater sampling variability does not establish reliable or reproducible grading.

C:

Groundedness cannot be assessed credibly without the relevant supporting text.

Question 656

A RAG agent starts fabricating product-policy answers. Engineers are unsure whether retrieval finds the wrong passages or the model ignores correct passages. Which experiment best isolates the cause?

  1. Change retrieval and generation settings simultaneously and compare only total latency.
  2. Inspect retrieved evidence separately, then replay generation with a known-correct evidence set.
  3. Disable citations and grade the responses by fluency alone.
  4. Increase top_k continuously without examining what the additional chunks contain.

Explanation

Correct Answer

B:

Component-level tests separate evidence selection problems from evidence-use failures.

Incorrect Answers

A:

Combined changes make root-cause attribution difficult and do not test factual support.

C:

Fluency can remain high while unsupported answers persist.

D:

Additional irrelevant passages may worsen rather than resolve hallucination.

Question 657

An inventory assistant serves repeated stock queries from a cache. A deployment report shows lower latency but more stale quantities after warehouse transactions. Which metric should be evaluated alongside cache hit rate?

  1. Count only the number of cached entries allocated in the service.
  2. Use model inference latency as a proxy for whether cached stock values remain current.
  3. Measure answer freshness and correctness against the authoritative inventory record.
  4. Measure average cache entry age without comparing returned quantities with updated stock events.

Explanation

Correct Answer

C:

Fast cached answers can violate the live-inventory requirement after updates.

Incorrect Answers

A:

Cache size does not establish that a returned quantity is current or accurate.

B:

Fast model generation does not establish inventory freshness after a warehouse transaction.

D:

Cache age alone cannot verify a particular stock value against recent source changes.

Question 658

A policy-answering agent succeeds on benchmark questions copied verbatim from internal help articles but misses equivalent questions phrased differently. Before a model upgrade, what additional evaluation is most informative?

  1. Test meaning-preserving question paraphrases on a held-out set and compare supported answer consistency.
  2. Reuse only the exact article headings as test queries and ignore variant user phrasing.
  3. Increase generation temperature and evaluate creativity instead of factual retrieval consistency.
  4. Remove difficult paraphrases from the test set to maintain the historical aggregate score.

Explanation

Correct Answer

A:

Paraphrase robustness measures whether the agent retrieves and grounds the same facts despite alternate wording.

Incorrect Answers

B:

The existing benchmark already overestimates performance on familiar wording.

C:

More varied language generation does not establish correct behavior on equivalent user intents.

D:

Excluding realistic queries conceals the actual robustness weakness.

Question 659

An order agent uses a primary stock lookup and a fallback stock lookup. For create_order, the fallback vendor cannot guarantee deduplication. What should the reliability plan do?

  1. Retry create_order across every vendor whenever a network timeout occurs.
  2. Use the fallback lookup to assert that a previously created order certainly failed.
  3. Disable all tool timeout handling and leave the agent waiting indefinitely.
  4. Allow controlled read-only lookup failover but require transaction-status verification before retrying orders.

Explanation

Correct Answer

D:

Reads can be retried safely more often than irreversible writes without idempotency.

Incorrect Answers

A:

An uncertain first result can lead to duplicate purchases after failover.

B:

A stock query does not reveal the authoritative status of a separate transaction.

C:

Unbounded waits do not provide controlled recovery or safe failure.

Question 660

A knowledge agent passes generation tests after a search-index rebuild, but several documents updated yesterday are missing. Which release check directly covers this failure?

  1. Approve based on the model’s ability to write coherent summaries from old documents.
  2. Compare indexed document versions and ingestion completeness with the authoritative source inventory.
  3. Measure only the number of vector dimensions in the rebuilt index.
  4. Increase the answer generation temperature to compensate for missing documents.

Explanation

Correct Answer

B:

Freshness and coverage checks catch missing recent source changes before approving the index.

Incorrect Answers

A:

Fluent outputs do not demonstrate that required new revisions were indexed.

C:

Correct schema dimensions do not prove source freshness or document coverage.

D:

Sampling cannot recover evidence absent from the search index.

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