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 7

Question 60

An Azure-based AI team must satisfy this objective: A generative assistant for public services omits captions for audio explanations. Which design choice is best?

  1. Evaluate safety again after tool and retrieval results are incorporated
  2. Provide synchronized transcripts or captions for spoken content
  3. Apply retention and deletion controls subject to recorded legal exceptions
  4. Assign accountable operational and risk owners with escalation responsibilities

Explanation

Correct Answer

B:

Accessible information must not depend on hearing audio. This implementation addresses the precise goal rather than an adjacent AI task.

Incorrect Answers

A:

Threats can enter later stages even when the original user request was harmless. That is not sufficient here: Accessible information must not depend on hearing audio.

C:

Accountability requires evidence of actual data lifecycle actions. That is not sufficient here: Accessible information must not depend on hearing audio.

D:

An AI model cannot be its own accountable incident responder. That is not sufficient here: Accessible information must not depend on hearing audio.

Question 61

An acceptance tester needs to validate: Use named entity recognition with entity types. Which observed result proves the intended behavior?

  1. Suggested products are relevant and meet inventory/authorization rules
  2. An article yields correctly labeled organizations and products
  3. The transcript contains the actual spoken issue and timestamp
  4. The timeline marks arrival frames with appropriate timestamps

Explanation

Correct Answer

B:

NER locates and categorizes named entities in text. The expected acceptance evidence is: An article yields correctly labeled organizations and products. This observation directly tests the stated behavior.

Incorrect Answers

A:

Semantic vectors capture related concepts while filters enforce eligibility. This observation does not establish the required result: An article yields correctly labeled organizations and products.

C:

Speech recognition converts speech into words suitable for search. This observation does not establish the required result: An article yields correctly labeled organizations and products.

D:

Video understanding retains order and time, unlike single-image tagging. This observation does not establish the required result: An article yields correctly labeled organizations and products.

Question 62

A project needs the following AI behavior: A claims form has a checked insurance option. What should the designer implement?

  1. Extract selection marks as explicit typed values
  2. Extract currency code and monetary amount as separate validated fields
  3. Use field confidence and source review for ambiguous numeric values
  4. Use timestamped evidence-linked video segments and explicitly account for frame-sampling limits

Explanation

Correct Answer

A:

Checkbox state is a structured signal rather than arbitrary prose. This implementation addresses the precise goal rather than an adjacent AI task.

Incorrect Answers

B:

A currency symbol or amount alone may not establish financial meaning. That is not sufficient here: Checkbox state is a structured signal rather than arbitrary prose.

C:

A low-confidence amount must not be silently treated as definitive. That is not sufficient here: Checkbox state is a structured signal rather than arbitrary prose.

D:

Temporal evidence supports the timing of observed events, but approximately one frame per second cannot guarantee subsecond event capture. That is not sufficient here: Checkbox state is a structured signal rather than arbitrary prose.

Question 63

Troubleshooting report: The team uses regression loss and interprets continuous scores as categories without thresholds. Which correction is MOST directly targeted to this failure?

  1. Select a classification output and calibrated decision boundary
  2. Implement and test fixed numeric rules in application code
  3. Chunk and retrieve relevant source sections before inference
  4. Reduce sampling variability and validate outputs against a schema

Explanation

Correct Answer

A:

This correction directly addresses the reported failure. Classification predicts discrete labels rather than a continuous numeric target. The targeted change is: Select a classification output and calibrated decision boundary.

Incorrect Answers

B:

Generative models are not a substitute for auditable deterministic computation. In the stated case, troubleshooting report: The team uses regression loss and interprets continuous scores as categories without thresholds; “Implement and test fixed numeric rules in application code” does not target that failure.

C:

A finite context window limits how much material can be processed at once. In the stated case, troubleshooting report: The team uses regression loss and interprets continuous scores as categories without thresholds; “Chunk and retrieve relevant source sections before inference” does not target that failure.

D:

Lower randomness often improves consistency but does not guarantee correctness. In the stated case, troubleshooting report: The team uses regression loss and interprets continuous scores as categories without thresholds; “Reduce sampling variability and validate outputs against a schema” does not target that failure.

Question 64

An Azure AI Fundamentals solution has this requirement: A voice assistant misunderstands accents used by a significant user population. Which implementation most directly addresses it?

  1. Evaluate outcomes with counterfactual name swaps and remove unjustified proxy effects
  2. Measure recognition errors across representative accent and speaking groups
  3. Maintain a governed feature inventory and trace decisions to policy
  4. Provide synchronized transcripts or captions for spoken content

Explanation

Correct Answer

B:

Inclusive voice systems require performance evaluation beyond a single accent. This implementation addresses the precise goal rather than an adjacent AI task.

Incorrect Answers

A:

Fairness checks compare materially equivalent cases rather than raw group averages alone. That is not sufficient here: Inclusive voice systems require performance evaluation beyond a single accent.

C:

Accountability requires interpretable reasons for consequential decisions. That is not sufficient here: Inclusive voice systems require performance evaluation beyond a single accent.

D:

Accessible information must not depend on hearing audio. That is not sufficient here: Inclusive voice systems require performance evaluation beyond a single accent.

Question 65

An engineering review documents this requirement: A claims OCR system processes blurry scans with unreadable handwriting. Select the most appropriate solution.

  1. Run supported image text extraction with field schema and source location
  2. Use an appropriate search-oriented document analyzer with meaningful chunk boundaries
  3. Extract selection marks as explicit typed values
  4. Apply confidence-based review instead of forcing every field into a number

Explanation

Correct Answer

D:

Low-quality inputs can make extraction genuinely uncertain. This implementation addresses the precise goal rather than an adjacent AI task.

Incorrect Answers

A:

Visual information extraction can map pixels to structured identifiers. That is not sufficient here: Low-quality inputs can make extraction genuinely uncertain.

B:

RAG preparation should preserve document structure and source references. That is not sufficient here: Low-quality inputs can make extraction genuinely uncertain.

C:

Checkbox state is a structured signal rather than arbitrary prose. That is not sufficient here: Low-quality inputs can make extraction genuinely uncertain.

Question 66

An Azure AI Fundamentals solution has this requirement: An AI model is updated and begins producing harmful content in previously safe scenarios. Which implementation most directly addresses it?

  1. Assign accountable operational and risk owners with escalation responsibilities
  2. Disclose collection purpose and retention period before capturing sensitive audio
  3. Record the actual decision output, input reference, model and prompt versions in integrity-protected audit records
  4. Run versioned safety regressions and block promotion on failed high-severity cases

Explanation

Correct Answer

D:

Reliability includes preventing behavior regression across model versions. This implementation addresses the precise goal rather than an adjacent AI task.

Incorrect Answers

A:

An AI model cannot be its own accountable incident responder. That is not sufficient here: Reliability includes preventing behavior regression across model versions.

B:

Privacy requires meaningful notice and bounded handling of personal data. That is not sufficient here: Reliability includes preventing behavior regression across model versions.

C:

Auditability requires the original decision evidence and version identifiers; regenerating model text is not guaranteed to be identical. That is not sufficient here: Reliability includes preventing behavior regression across model versions.

Question 67

An Azure AI solution uses the following approved control: Use supervised classification with labeled maintenance outcomes. What observable result should pass QA?

  1. The product shape remains unchanged against the new background
  2. Held-out events produce measured recall and precision for failures
  3. Names in a labeled sample are detected and redacted
  4. Test feedback returns relevant phrases found in the message

Explanation

Correct Answer

B:

A failure probability or class is a discrete decision under a defined horizon. The expected acceptance evidence is: Held-out events produce measured recall and precision for failures. This observation directly tests the stated behavior.

Incorrect Answers

A:

Editing changes selected pixels while leaving product regions intact. This observation does not establish the required result: Held-out events produce measured recall and precision for failures.

C:

Entity extraction locates sensitive spans for privacy handling. This observation does not establish the required result: Held-out events produce measured recall and precision for failures.

D:

Key phrase extraction finds prominent topics without inventing new source material. This observation does not establish the required result: Held-out events produce measured recall and precision for failures.

Question 68

Troubleshooting report: The system returns organization names but no attitude labels. Which correction is MOST directly targeted to this failure?

  1. Train a binary failure classifier using historical labeled cases
  2. Apply entity extraction and validate spans and categories
  3. Use sentiment analysis instead of named-entity extraction
  4. Interpret the actual image first and synthesize the verified description

Explanation

Correct Answer

C:

This correction directly addresses the reported failure. Sentiment analysis infers expressed attitude rather than extracting names. The targeted change is: Use sentiment analysis instead of named-entity extraction.

Incorrect Answers

A:

A failure probability or class is a discrete decision under a defined horizon. In the stated case, troubleshooting report: The system returns organization names but no attitude labels; “Train a binary failure classifier using historical labeled cases” does not target that failure.

B:

NER locates and categorizes named entities in text. In the stated case, troubleshooting report: The system returns organization names but no attitude labels; “Apply entity extraction and validate spans and categories” does not target that failure.

D:

Multimodal input can describe the graphic and speech output can read that description. In the stated case, troubleshooting report: The system returns organization names but no attitude labels; “Interpret the actual image first and synthesize the verified description” does not target that failure.

Question 69

An AI operations team is testing two distinct capabilities. Failure (1): A text-only model describes the edit but returns no image. Failure (2): An unrelated generic summary is generated from the webinar title only. Which TWO specific corrections should the team apply?

  1. Treat absent fields as unknown and request a better image
  2. Choose segmentation with validated pixel-level damage regions
  3. Call the authorized ticket tool and verify the receipt
  4. Invoke image segmentation and compositing or a controlled image editor
  5. Ground the summary in extracted transcript segments

Explanation

Correct Answers

D:

Correct for failure (1). Editing changes selected pixels while leaving product regions intact. The proposed correction directly addresses that reported defect.

E:

Correct for failure (2). Meeting summaries require temporal and spoken source content. It independently corrects the second defect.

Incorrect Answers

A:

Images of receipts need OCR/layout extraction before typed business values. In this case the first outcome requires the team to use image editing or segmentation with object-mask preservation, and the second requires it to process the audio transcript and optionally key video scenes before summarizing. This option does not replace those two separate controls.

B:

Pixel-level region masks localize building damage more precisely than an overall image class label. In this case the first outcome requires the team to use image editing or segmentation with object-mask preservation, and the second requires it to process the audio transcript and optionally key video scenes before summarizing. This option does not replace those two separate controls.

C:

Agentic workloads combine reasoning and bounded external actions. In this case the first outcome requires the team to use image editing or segmentation with object-mask preservation, and the second requires it to process the audio transcript and optionally key video scenes before summarizing. This option does not replace those two separate controls.

Question 70

A project needs the following AI behavior: A planned peak workload exceeds the selected model deployment throughput quota before launch. What should the designer implement?

  1. Choose a model or pipeline that supports both visual and audio modalities
  2. Choose a speech-recognition model or audio-capable multimodal endpoint with a documented audio-input contract
  3. Measure expected tokens and requests per minute, then choose an approved throughput tier or capacity adjustment before launch
  4. Choose a compatible multimodal model and provide supported text/image inputs

Explanation

Correct Answer

C:

Quota sizing is a deployment-capacity decision; retrying a predictable sustained excess cannot add model throughput.

Incorrect Answers

A:

Different media channels provide complementary scene and dialogue evidence. That is not sufficient here: Quota sizing is a deployment-capacity decision; retrying a predictable sustained excess cannot add model throughput.

B:

Model capabilities must include the required audio modality before the client can obtain speech-to-text results. That is not sufficient here: Quota sizing is a deployment-capacity decision; retrying a predictable sustained excess cannot add model throughput.

D:

Multimodal models can reason over more than one supported input modality. That is not sufficient here: Quota sizing is a deployment-capacity decision; retrying a predictable sustained excess cannot add model throughput.

Microsoft Azure AI Fundamentals (AI-901) Practice Tests