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 37

Question 360

A trace pipeline records full customer invoices and auth headers for every agent call. Investigators need timings and tool outcomes, not raw documents. What should change?

  1. Reduce data leakage by omitting sensitive details from trace payloads.
  2. Disable all telemetry instead of stripping sensitive fields.
  3. Keep all invoice attachments forever in operational logs for convenience.
  4. Share full trace dumps with every agent in the organization.

Explanation

Correct Answer

A:

Structured safe metadata supports troubleshooting without persistent exposure of confidential contents.

Incorrect Answers

B:

Complete logging removal sacrifices necessary operational evidence.

C:

Unneeded raw records expand breach impact.

D:

Broad access violates data minimization and role boundaries.

Question 361

An autonomous claims agent begins submitting approvals even though grounding evaluation detects unsupported policy citations. What safety gate should be applied?

  1. Trigger guarded escalation on degraded grounding quality.
  2. Ignore evaluator failures if the agent's language is persuasive.
  3. Allow actions to continue because response HTTP status is successful.
  4. Disable citation checks to reduce monitoring alerts.

Explanation

Correct Answer

A:

A validated content-quality control can prevent unsubstantiated decisions from causing external side effects.

Incorrect Answers

B:

Fluency cannot justify unsupported policy actions.

C:

Transport success does not establish source-backed claim correctness.

D:

Suppressing evidence failures does not make the claims grounded.

Question 362

Two research agents query different pricing APIs in parallel. One returns a cached figure from yesterday and one returns a current figure. What should the orchestrator verify?

  1. Average the two numbers without checking effective dates.
  2. Discard both source timestamps to simplify the report.
  3. Validate consistency of tool result timestamps across parallel agents.
  4. Prefer whichever agent finished first regardless of its timestamp.

Explanation

Correct Answer

C:

Parallel outputs can be individually valid yet inconsistent in time.

Incorrect Answers

A:

Mixing stale and current data can create an unsupported result.

B:

Removing time provenance prevents a reliable reconciliation.

D:

Completion order does not establish freshness.

Question 363

A RAG assistant's final answer style is stable, but the team suspects missing documents after a search change. Which metric should be prioritized?

  1. Measure page-load CSS timing rather than query results.
  2. Evaluate retrieval recall@k against labeled expected document results.
  3. Count generation tokens without inspecting returned documents.
  4. Score only the friendly tone of the final answer.

Explanation

Correct Answer

B:

Evidence selection metrics isolate retrieval performance from consistent answer formatting.

Incorrect Answers

A:

Frontend paint time does not establish document recall.

C:

Usage data is not retrieval quality evidence.

D:

Style cannot reveal absent relevant passages.

Question 364

Users give an agent high satisfaction ratings, but audited invoices contain incorrect tax amounts. Which reporting separation is necessary?

  1. Separate user satisfaction surveys from objective answer-accuracy measures.
  2. Reduce response verbosity instead of measuring calculation errors.
  3. Remove verified incorrect cases from the dataset because ratings are high.
  4. Treat positive feedback as proof all invoice calculations are correct.

Explanation

Correct Answer

A:

Perception and independently verified task accuracy are distinct evidence streams.

Incorrect Answers

B:

Shortness is not a substitute for objective correctness.

C:

Discarding hard evidence conceals factual defects.

D:

Users may like an answer without checking its arithmetic.

Question 365

A new agent version fails its accuracy gate. A rollback is proposed, but reviewers need to know why. What evidence should be recorded?

  1. Record the model's self-confidence as the only failure explanation.
  2. Document only that a reviewer felt the agent was worse.
  3. Document regression failure evidence before rolling back an agent version.
  4. Delete the failed test inputs so the record stays concise.

Explanation

Correct Answer

C:

Concrete regression evidence supports reproducibility and controlled recovery.

Incorrect Answers

A:

Confidence does not identify an actual incorrect answer or action.

B:

An impression cannot reconstruct the failing cases.

D:

Removing evidence makes root cause verification difficult.

Question 366

A router chooses between a cheap compact model and an expensive reasoning model. Regulated cases carry greater risk than general FAQ traffic. Which decision policy fits?

  1. Always use the most expensive model even for static low-risk facts.
  2. Use cost and risk together to set a model routing policy.
  3. Use the cheapest model for all decisions regardless of measured failures.
  4. Route only by user message character count without testing complexity.

Explanation

Correct Answer

B:

Risk-adjusted selection balances operational expense against the need for demonstrably adequate capability.

Incorrect Answers

A:

Uniform maximum cost ignores proven simpler options.

C:

Pure cost optimization can violate high-stakes quality requirements.

D:

Message length alone does not measure consequence or reasoning difficulty.

Question 367

A retailer wants a completely new illustrated banner based on a written description of a seasonal scene. It has no source image to modify. Which capability should the designer invoke?

  1. Use only image captioning to describe an existing photo.
  2. Run document OCR to identify table cells in the campaign brief.
  3. Use an approved text-to-image generation model with the campaign prompt and appropriate output settings.
  4. Use reference-guided image editing despite having no existing source image.

Explanation

Correct Answer

C:

Text-to-image generation directly creates a new image from a textual scene description.

Incorrect Answers

A:

Captioning analyzes supplied imagery; it does not create a new banner from the written description.

B:

OCR extracts text from images or documents and is not the requested visual creation step.

D:

Image editing is appropriate when modifying an image; the task calls for creating original artwork solely from a written scene.

Question 368

A brand manager has an approved product photograph. The new advertisement must keep the exact product shape while changing background aesthetics based on that reference. Which workflow should the team evaluate?

  1. Use image-to-image generation or editing with the approved product reference and preservation constraints.
  2. Start from text alone and never provide the product image to the model.
  3. Apply an Azure AI Search query to the product SKU as the final edited image.
  4. Convert the photo to a meeting transcript and translate it into another language.

Explanation

Correct Answer

A:

A reference-guided editing workflow can constrain product appearance while transforming the surrounding creative presentation.

Incorrect Answers

B:

Without the approved reference, preserving precise product identity and geometry is less reliable.

C:

Search can retrieve assets but does not perform the specified creative background modification.

D:

Audio and language transformation do not edit the source image background.

Question 369

A promotional photo has a small damaged area on a painted wall. The editor must replace only that patch and preserve the subject and rest of the background. What technique is most appropriate?

  1. Use outpainting to extend all four borders without targeting the damaged region.
  2. Use an inpainting mask to replace only a damaged image region.
  3. Rerun broad text-to-image generation without the original photograph.
  4. Choose an OCR-plus-captioning pipeline to describe an existing illustration rather than synthesize new pixels.

Explanation

Correct Answer

B:

A localized mask constrains the edit to the intended region while preserving unmasked pixels as much as supported.

Incorrect Answers

A:

Border expansion is designed for adding canvas outside the source image, not correcting a precise interior patch.

C:

Regenerating the entire image can alter the subject and many undamaged details.

D:

Extraction and captioning cannot generate the new image needed for the creative request.

Question 370

An approved landscape image is too narrow for a panoramic website header. The team wants to extend the scenery beyond the photo edges rather than change the central composition. Which edit is designed for this?

  1. Apply a small interior inpainting mask over the main subject.
  2. Horizontally stretch the source image to the new header width without adding any scenery.
  3. Crop both sides of the photograph to reduce its width.
  4. Use outpainting to extend a photo beyond its original boundaries.

Explanation

Correct Answer

D:

Outpainting generates plausible surrounding imagery beyond the existing borders while retaining the central reference.

Incorrect Answers

A:

Interior patch replacement does not add pixels beyond the existing canvas edges.

B:

Rescaling existing pixels distorts the center content instead of synthesizing new imagery beyond original boundaries.

C:

Cropping removes content and cannot extend the image to the requested wider aspect.

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