McAfee Secure

Exam Code: AI-901

Exam Name: Microsoft Azure AI Fundamentals

Certification Provider: Microsoft

Corresponding Certification: Microsoft Certified: Azure AI Fundamentals

Microsoft AI-901 Questions & Answers

Study with Up-To-Date REAL Exam Questions and Answers from the ACTUAL Test

166 Questions & Answers with Testing Engine

"Microsoft Azure AI Fundamentals Exam", also known as AI-901 exam, is a Microsoft certification exam.

Pass your tests with the always up-to-date AI-901 Exam Engine. Your AI-901 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-901 Practice Test 37

Question 360

An Azure AI solution uses the following approved control: Evaluate language safety on dialect-representative benign and harmful examples. What observable result should pass QA?

  1. Scanner-specific evaluation identifies the weak cohort and blocks unsupported rollout
  2. Eligible memory entries disappear from active stores after the verified request
  3. Logs contain operational metadata but not raw customer numbers
  4. Dialect-balanced tests show benign requests are allowed without unsafe leakage

Explanation

Correct Answer

D:

Fairness requires discriminating actual harm from linguistic variety. The expected acceptance evidence is: Dialect-balanced tests show benign requests are allowed without unsafe leakage. This observation directly tests the stated behavior.

Incorrect Answers

A:

Reliability requires detecting distribution shifts hidden in aggregate accuracy. This observation does not establish the required result: Dialect-balanced tests show benign requests are allowed without unsafe leakage.

B:

Accountability requires evidence of actual data lifecycle actions. This observation does not establish the required result: Dialect-balanced tests show benign requests are allowed without unsafe leakage.

C:

Troubleshooting does not justify reproducing sensitive application records. This observation does not establish the required result: Dialect-balanced tests show benign requests are allowed without unsafe leakage.

Question 361

Troubleshooting report: The pipeline uses an image captioning endpoint for plain text feedback. Which correction is MOST directly targeted to this failure?

  1. Replace it with a suitable text-analysis model or service
  2. Create compatible text embeddings and evaluate vector retrieval
  3. Invoke text-read or document OCR capability
  4. Process appropriate video frames and the audio track

Explanation

Correct Answer

A:

This correction directly addresses the reported failure. Different model tasks require compatible outputs and labeled evaluation data. The targeted change is: Replace it with a suitable text-analysis model or service.

Incorrect Answers

B:

Embeddings encode learned semantic relationships rather than literal string identity. In the stated case, troubleshooting report: The pipeline uses an image captioning endpoint for plain text feedback; “Create compatible text embeddings and evaluate vector retrieval” does not target that failure.

C:

OCR produces readable text from pixels. In the stated case, troubleshooting report: The pipeline uses an image captioning endpoint for plain text feedback; “Invoke text-read or document OCR capability” does not target that failure.

D:

Different media channels provide complementary scene and dialogue evidence. In the stated case, troubleshooting report: The pipeline uses an image captioning endpoint for plain text feedback; “Process appropriate video frames and the audio track” does not target that failure.

Question 362

An AI operations team is testing two distinct capabilities. Failure (1): A malformed numeral is silently replaced with zero. Failure (2): Every sentence is attributed to the first speaker. Which TWO specific corrections should the team apply?

  1. Poll the documented operation status and handle failures
  2. Use table-aware layout extraction and numeric validation
  3. Enable appropriate diarization and handle ambiguous overlap
  4. Mark the field unreadable and request manual correction
  5. Use schema definitions and grounding evidence for each field

Explanation

Correct Answers

C:

Correct for failure (2). Conversation extraction needs more than a single flattened text stream. It independently corrects the second defect.

D:

Correct for failure (1). Low-quality inputs can make extraction genuinely uncertain. The proposed correction directly addresses that reported defect.

Incorrect Answers

A:

Long-running media analysis may be asynchronous. It does not complete the two decisions required in this question.

B:

Cell relationships are necessary to check item totals. It does not complete the two decisions required in this question.

E:

Custom extraction maps semantically relevant phrases to typed fields. The stated pair of failures instead requires “Enable appropriate diarization and handle ambiguous overlap” and “Mark the field unreadable and request manual correction”; this option cannot substitute for both.

Question 363

An Azure AI solution uses the following approved control: Use the same compatible embedding version and dimensions for stored items and queries. What observable result should pass QA?

  1. A held-out set receives the correct whole-image classes and empirically calibrated class probabilities
  2. The execution trace shows a permitted tool called with validated inputs
  3. The same facts are consistently presented in the required style
  4. Queries and indexed vectors carry matching model-version metadata

Explanation

Correct Answer

D:

Incompatible vector spaces can make nearest-neighbor rankings meaningless. The expected acceptance evidence is: Queries and indexed vectors carry matching model-version metadata. This observation directly tests the stated behavior.

Incorrect Answers

A:

Classification assigns image-wide categories rather than bounding-box coordinates. This observation does not establish the required result: Queries and indexed vectors carry matching model-version metadata.

B:

Agent workflows can reason about tool selection but tools still need authorization. This observation does not establish the required result: Queries and indexed vectors carry matching model-version metadata.

C:

Clear prompting can control tone without model retraining. This observation does not establish the required result: Queries and indexed vectors carry matching model-version metadata.

Question 364

An engineering review documents this requirement: A classifier provides rejection decisions without a way for affected users to contest errors. Select the most appropriate solution.

  1. Record decision reasons and give a human review route for important outcomes
  2. Require independent transaction authorization and approval outside prompts
  3. Enforce tenant identity and record authorization before retrieval and extraction
  4. Disclose the automated nature and supported escalation route in the interface

Explanation

Correct Answer

A:

Accountability includes traceable responsibility and appropriate contestability. This implementation addresses the precise goal rather than an adjacent AI task.

Incorrect Answers

B:

Accountable high-impact operations need verifiable authorized signatures. That is not sufficient here: Accountability includes traceable responsibility and appropriate contestability.

C:

Model instructions cannot grant access beyond the underlying data permission boundary. That is not sufficient here: Accountability includes traceable responsibility and appropriate contestability.

D:

Transparency helps users assess the scope and limitations of the assistance. That is not sufficient here: Accountability includes traceable responsibility and appropriate contestability.

Question 365

A project needs the following AI behavior: A product photograph contains printed lot codes. What should the designer implement?

  1. Use a video-capable Content Understanding analyzer with timestamps and supported fields
  2. Run supported image text extraction with field schema and source location
  3. Preserve text spans and associated source regions during OCR
  4. Use field confidence and source review for ambiguous numeric values

Explanation

Correct Answer

B:

Visual information extraction can map pixels to structured identifiers. This implementation addresses the precise goal rather than an adjacent AI task.

Incorrect Answers

A:

Multimodal extraction draws on visual and spoken evidence over time. That is not sufficient here: Visual information extraction can map pixels to structured identifiers.

C:

Source-location evidence helps verify recognition on mixed-layout documents. That is not sufficient here: Visual information extraction can map pixels to structured identifiers.

D:

A low-confidence amount must not be silently treated as definitive. That is not sufficient here: Visual information extraction can map pixels to structured identifiers.

Question 366

The approved design action is: Measure false positives on approved accessibility scenarios and refine thresholds. Which test result would BEST establish that it worked?

  1. The returned metric reconciles to the approved source table
  2. An unverified testimonial is labeled illustrative rather than quoted as authentic
  3. A deaf tester receives the complete equivalent content as text
  4. Labeled accessible-use examples pass while harmful variants remain blocked

Explanation

Correct Answer

D:

Safety controls should not overblock permitted assistance. The expected acceptance evidence is: Labeled accessible-use examples pass while harmful variants remain blocked. This observation directly tests the stated behavior.

Incorrect Answers

A:

A numeric-looking answer is not proof that a measure was computed. This observation does not establish the required result: Labeled accessible-use examples pass while harmful variants remain blocked.

B:

Transparency prohibits presenting invented testimony as genuine. This observation does not establish the required result: Labeled accessible-use examples pass while harmful variants remain blocked.

C:

Accessible information must not depend on hearing audio. This observation does not establish the required result: Labeled accessible-use examples pass while harmful variants remain blocked.

Question 367

Troubleshooting report: The analyzer produces only a generic image caption. Which correction is MOST directly targeted to this failure?

  1. Configure supported language handling and test each script
  2. Configure required invoice fields and validate extracted values
  3. Verify the actual source signature region or use explicitly approved preview detection, never infer signing from a typed name
  4. Configure OCR-derived field extraction for the lot code

Explanation

Correct Answer

D:

This correction directly addresses the reported failure. Visual information extraction can map pixels to structured identifiers. The targeted change is: Configure OCR-derived field extraction for the lot code.

Incorrect Answers

A:

Recognition accuracy depends on script and language support. In the stated case, troubleshooting report: The analyzer produces only a generic image caption; “Configure supported language handling and test each script” does not target that failure.

B:

Document extraction turns unstructured pages into validated structured fields. In the stated case, troubleshooting report: The analyzer produces only a generic image caption; “Configure required invoice fields and validate extracted values” does not target that failure.

C:

Selection-mark extraction is available with GA document analysis, but dedicated signature-location detection is available only in the 2026-06-01-preview API and cannot establish legal validity. In the stated case, troubleshooting report: The analyzer produces only a generic image caption; “Verify the actual source signature region or use explicitly approved preview detection, never infer signing from a typed name” does not target that failure.

Question 368

An AI operations team is testing two distinct capabilities. Failure (1): Prompt changes are edited in place with no revision history. Failure (2): The model follows a user instruction to expose private internal prompts. Which TWO specific corrections should the team apply?

  1. Use environment-specific configuration and validate resource IDs
  2. Version prompt and deployment settings and retain the actual evaluation output
  3. Reinforce role boundaries and enforce secrets outside prompts
  4. Rewrite the prompt with explicit acceptance criteria
  5. Add adversarial cases and verify configured safety block and escalation behavior

Explanation

Correct Answers

B:

Correct for failure (1). Recorded settings support investigation, but generative models can still vary on an apparently identical rerun. The proposed correction directly addresses that reported defect.

C:

Correct for failure (2). Developer and system constraints govern allowed operations over user requests. It independently corrects the second defect.

Incorrect Answers

A:

A correct code path can still target the wrong resource after migration. The stated pair of failures instead requires “Version prompt and deployment settings and retain the actual evaluation output” and “Reinforce role boundaries and enforce secrets outside prompts”; this option cannot substitute for both.

D:

Specific instructions reduce ambiguity about what constitutes success. It does not complete the two decisions required in this question.

E:

Adversarial testing must measure actual configured policy enforcement rather than treating a prompt instruction as a guaranteed safety boundary. The stated pair of failures instead requires “Version prompt and deployment settings and retain the actual evaluation output” and “Reinforce role boundaries and enforce secrets outside prompts”; this option cannot substitute for both.

Question 369

An AI operations team is testing two distinct capabilities. Failure (1): The app incorrectly treats object coordinates as the requested category. Failure (2): Raw nonimage bytes are sent with an image MIME label. Which TWO specific corrections should the team apply?

  1. Request the needed visual features explicitly
  2. Use the documented feature list for the deployed API version
  3. Map the task to a whole-image classifier
  4. Constrain the prompt and validate the final artwork
  5. Decode and re-encode to a supported image format

Explanation

Correct Answers

C:

Correct for failure (1). Image-level classification is distinct from object localization. The proposed correction directly addresses that reported defect.

E:

Correct for failure (2). Vision endpoints impose supported input formats and limits. It independently corrects the second defect.

Incorrect Answers

A:

Different visual outputs describe text, scene classes and detected people. The stated pair of failures instead requires “Map the task to a whole-image classifier” and “Decode and re-encode to a supported image format”; this option cannot substitute for both.

B:

Feature names and availability vary across versions. It does not complete the two decisions required in this question.

D:

Explicit visual guidance can influence generated appearance but is not absolute. The stated pair of failures instead requires “Map the task to a whole-image classifier” and “Decode and re-encode to a supported image format”; this option cannot substitute for both.

Question 370

An Azure AI Fundamentals solution has this requirement: A Speech service client fails because its configured region differs from its resource. Which implementation most directly addresses it?

  1. Match the speech resource region or endpoint to the SDK configuration
  2. Evaluate word error rate on the same held-out audio samples
  3. Use word or utterance timestamps from a supported recognition output
  4. Combine recognition or audio understanding with verified text and text-to-speech

Explanation

Correct Answer

A:

Region or endpoint mismatch prevents requests reaching the correct resource. This implementation addresses the precise goal rather than an adjacent AI task.

Incorrect Answers

B:

Comparable labeled audio prevents misleading quality claims. That is not sufficient here: Region or endpoint mismatch prevents requests reaching the correct resource.

C:

Timecodes align text segments to audio playback. That is not sufficient here: Region or endpoint mismatch prevents requests reaching the correct resource.

D:

Input transcription and output synthesis are distinct steps. That is not sufficient here: Region or endpoint mismatch prevents requests reaching the correct resource.

Microsoft Azure AI Fundamentals (AI-901) Practice Tests