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 35

Question 340

A safety threshold was calibrated in English; the agent also serves several regional languages. What must a pre-release evaluation include?

  1. Exclude the most difficult language without changing supported-language claims.
  2. Assume model multilingual fluency proves equal content-filter performance.
  3. Test safety controls in all supported languages with labeled cases.
  4. Use only response latency to certify moderation across languages.

Explanation

Correct Answer

C:

Moderation behavior can vary by language, so English-only results are insufficient.

Incorrect Answers

A:

Omitted cohorts conceal potential safety gaps.

B:

Language generation ability does not establish safety classifier behavior.

D:

Speed is not content classification accuracy.

Question 341

A CI pipeline builds a new agent version. The owner defines minimum task completion and groundedness requirements. What should the pipeline enforce?

  1. Gate CI/CD on task-success and groundedness thresholds.
  2. Evaluate only how fast the build completes.
  3. Increase allowed failures after each regression without approval.
  4. Promote every build if its JSON parses, regardless of task errors.

Explanation

Correct Answer

A:

Automated gates prevent a version with measured quality regressions from being promoted.

Incorrect Answers

B:

Build speed is unrelated to answer groundedness.

C:

Silently relaxing thresholds defeats the owner-defined acceptance rules.

D:

Schema validity alone does not guarantee agent success or factual support.

Question 342

A finance agent sometimes calls privileged transfer tools. Monitoring currently samples only final responses, losing execution context. What should be traced?

  1. Keep raw payment credentials in every public trace to reproduce calls.
  2. Capture protected high-risk tool spans, approvals, outcomes and correlation IDs with sensitive fields minimized.
  3. Disable traces for expensive operations to reduce dashboard noise.
  4. Log only the agent's greeting text and ignore transfer calls.

Explanation

Correct Answer

B:

Execution-level telemetry supports detection and investigation of critical side effects.

Incorrect Answers

A:

Unnecessary secret logging increases financial exposure.

C:

High-impact actions require stronger rather than weaker accountability.

D:

Friendly responses cannot prove correct transaction execution.

Question 343

An API tool's retry count grows from one to ten per successful agent request, though success rates have not yet dropped. What warning should operations create?

  1. Alert on rising retry counts and latency tails.
  2. Alert only when the last tool request fails permanently, ignoring retry trends and elapsed time.
  3. Ignore retries until every customer request returns a 500.
  4. Automatically double retry limits every time volume increases.

Explanation

Correct Answer

A:

Retries can mask emerging reliability problems before requests begin failing completely.

Incorrect Answers

B:

Increasing retry counts may signal degradation before final failure rates increase; monitoring only terminal errors misses early warning.

C:

Waiting for total failure forfeits early intervention.

D:

More retries can overload a degraded service.

Question 344

An agent action group calls CRM and billing endpoints. Error rates increase only for billing operations. What alert design is most informative?

  1. Alert solely on the agent's overall number of chat sessions.
  2. Combine all actions into one global average and ignore the billing-specific trend.
  3. Turn off action outcome logging to simplify operations.
  4. Set per-action-group/operation alerts on error, latency and task-success indicators.

Explanation

Correct Answer

D:

Targeted alerts identify the failing business dependency without obscuring healthy CRM actions.

Incorrect Answers

A:

Conversation volume does not measure billing tool quality.

B:

Aggregation may hide a concentrated operational regression.

C:

Removing outcomes prevents identifying the failing group.

Question 345

Inference cost jumps after a prompt rewrite. Request counts are unchanged, but generated responses became much longer. Which metric distinguishes the cause?

  1. Infer token consumption from rendered response length rather than provider-measured input and output usage.
  2. Look only at requests-per-minute because tokens cannot vary across calls.
  3. Measure output tokens separately from request rate and input tokens.
  4. Check search index storage bytes as the first explanation of more output tokens.

Explanation

Correct Answer

C:

Output-token growth explains cost increases even when calls per minute stay flat.

Incorrect Answers

A:

Displayed text length does not capture hidden retrieval context or billed token metrics.

B:

Response lengths can differ dramatically among equal call counts.

D:

Index storage is separate from billed model generation usage.

Question 346

A support agent answers most questions quickly, but one in twenty takes 40 seconds. Median response time looks acceptable. What latency statistic exposes the tail?

  1. Replace latency measurement with response word count.
  2. Count only the shortest request in each hour.
  3. Measure p95 latency when median performance hides tail failures.
  4. Use median latency alone and declare the long tail irrelevant.

Explanation

Correct Answer

C:

The 95th percentile highlights slow-tail experiences hidden by a central tendency.

Incorrect Answers

A:

Output length alone does not quantify end-to-end tail delay.

B:

Best-case samples hide tail failures.

D:

A good median can coexist with unacceptable slow user sessions.

Question 347

A helpdesk bot can cache a static public office address, but not a personalized benefits balance. Which cache rule fits?

  1. Cache obsolete policies forever even after revision.
  2. Cache all benefits balances globally under the same question string.
  3. Cache only scoped, nonpersonalized read-only responses with expiry.
  4. Use stored access tokens as part of user-visible cache output.

Explanation

Correct Answer

C:

Caching stable public answers saves inference without leaking user-specific state.

Incorrect Answers

A:

Unbounded retention can serve invalid content.

B:

Shared cache keys can expose another user's private information.

D:

Secrets must not become cached response content.

Question 348

A RAG app caches a search result for tenant A. Tenant B sends an identical question and receives A's confidential passage. What should cache lookup enforce?

  1. Use one permanent global cache key for all company searches.
  2. Hide the confidential document title after serving its passage from cache.
  3. Treat identical question text as proof users have identical document rights.
  4. Prevent a cache hit from bypassing tenant authorization checks.

Explanation

Correct Answer

D:

Cached evidence must obey the same access controls as fresh retrieval.

Incorrect Answers

A:

A shared key increases cross-tenant leakage risk.

B:

Post-display masking cannot undo exposure.

C:

Semantic question equality does not imply equal data permissions.

Question 349

An agent integrates with a downstream case-management API that expects a stable record shape. Model outputs occasionally switch between bullet lists and prose. What should be enforced?

  1. Let the model decide its own response format on each request.
  2. Use a defined structured response schema with fixed field names and application-side validation.
  3. Use only a writing-style instruction without validating the actual fields.
  4. Treat a grammatically correct paragraph as equivalent to typed JSON.

Explanation

Correct Answer

B:

Stable output contracts prevent format drift from breaking automated consumers.

Incorrect Answers

A:

Free-form variation is incompatible with a strict receiving API.

C:

Style guidance is not deterministic schema enforcement.

D:

Natural-language fluency does not satisfy a machine contract.

Question 350

A model-generated JSON record contains all required fields plus several unknown optional properties. The integration must stay compatible without silently accepting dangerous fields. What should tests cover?

  1. Accept every unknown field and pass it directly into privileged database writes.
  2. Reject or explicitly validate unknown schema fields before accepting output.
  3. Reject every record solely because its field order differs.
  4. Parse only the first property and assume the rest are valid.

Explanation

Correct Answer

B:

Schema behavior for additional fields must be deterministic and reviewed, not accidental.

Incorrect Answers

A:

Unexpected keys may inject unauthorized operations or break downstream mapping.

C:

JSON object key order is not generally a data-contract semantic.

D:

Ignoring fields avoids neither missing requirements nor unsafe additions.

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