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 59

Question 580

An Azure AI Fundamentals solution has this requirement: A product Q&A assistant needs frequently revised catalog facts. Which implementation most directly addresses it?

  1. Use text-to-speech synthesis with a supported voice
  2. Use clustering or another appropriate unsupervised technique
  3. Use retrieval-augmented generation from current approved catalog sources
  4. Use a held-out evaluation and investigate overfitting or data leakage

Explanation

Correct Answer

C:

External retrieval updates factual context without retraining model weights. This implementation addresses the precise goal rather than an adjacent AI task.

Incorrect Answers

A:

Speech synthesis generates audio from text rather than transcribing audio. That is not sufficient here: External retrieval updates factual context without retraining model weights.

B:

Unsupervised learning finds structure without supplied target labels. That is not sufficient here: External retrieval updates factual context without retraining model weights.

D:

A model can memorize training examples without generalizing. That is not sufficient here: External retrieval updates factual context without retraining model weights.

Question 581

An Azure-based AI team must satisfy this objective: A medication assistant may suggest unsafe dosing from incomplete patient context. Which design choice is best?

  1. Measure recognition errors across representative accent and speaking groups
  2. Treat retrieved content as untrusted evidence and prevent it from changing tool authority
  3. Set explicit uncertainty thresholds and route high-risk decisions to qualified review
  4. Maintain a governed feature inventory and trace decisions to policy

Explanation

Correct Answer

C:

Safety-critical recommendations require verified constraints outside open-ended generation. This implementation addresses the precise goal rather than an adjacent AI task.

Incorrect Answers

A:

Inclusive voice systems require performance evaluation beyond a single accent. That is not sufficient here: Safety-critical recommendations require verified constraints outside open-ended generation.

B:

Lower-trust text is not a privileged command channel. That is not sufficient here: Safety-critical recommendations require verified constraints outside open-ended generation.

D:

Accountability requires interpretable reasons for consequential decisions. That is not sufficient here: Safety-critical recommendations require verified constraints outside open-ended generation.

Question 582

An Azure AI solution uses the following approved control: Assign accountable operational and risk owners with escalation responsibilities. What observable result should pass QA?

  1. A restricted summary omits or masks nonessential identifying features
  2. A residency map confirms every protected copy remains in allowed locations
  3. A prompt-injection test remains unable to trigger a prohibited tool after retrieval
  4. An incident simulation reaches a named owner with clear containment authority

Explanation

Correct Answer

D:

An AI model cannot be its own accountable incident responder. The expected acceptance evidence is: An incident simulation reaches a named owner with clear containment authority. This observation directly tests the stated behavior.

Incorrect Answers

A:

Privacy applies to image content as well as text transcripts. This observation does not establish the required result: An incident simulation reaches a named owner with clear containment authority.

B:

Privacy and compliance require controlling the full data path. This observation does not establish the required result: An incident simulation reaches a named owner with clear containment authority.

C:

Threats can enter later stages even when the original user request was harmless. This observation does not establish the required result: An incident simulation reaches a named owner with clear containment authority.

Question 583

An Azure-based AI team must satisfy this objective: A generative model creates fake statistics without citing a dataset. Which design choice is best?

  1. Use validated data retrieval and source-backed numerical responses
  2. Maintain a governed feature inventory and trace decisions to policy
  3. Require governance evidence with reviewers, test outcomes, and documented exceptions
  4. Record the actual decision output, input reference, model and prompt versions in integrity-protected audit records

Explanation

Correct Answer

A:

A numeric-looking answer is not proof that a measure was computed. This implementation addresses the precise goal rather than an adjacent AI task.

Incorrect Answers

B:

Accountability requires interpretable reasons for consequential decisions. That is not sufficient here: A numeric-looking answer is not proof that a measure was computed.

C:

Self-attestation is not independent proof of regulatory controls. That is not sufficient here: A numeric-looking answer is not proof that a measure was computed.

D:

Auditability requires the original decision evidence and version identifiers; regenerating model text is not guaranteed to be identical. That is not sufficient here: A numeric-looking answer is not proof that a measure was computed.

Question 584

An engineering review documents this requirement: A binary alert model performs well on a very imbalanced dataset. Select the most appropriate solution.

  1. Review precision, recall and the confusion matrix rather than accuracy alone
  2. Choose a text-classification or sentiment-analysis capability rather than an image generator
  3. Use the same compatible embedding version and dimensions for stored items and queries
  4. Evaluate domain adaptation or fine-tuning only after baseline tests show the gap

Explanation

Correct Answer

A:

High accuracy can conceal failure to detect rare positive events. This implementation addresses the precise goal rather than an adjacent AI task.

Incorrect Answers

B:

Different model tasks require compatible outputs and labeled evaluation data. That is not sufficient here: High accuracy can conceal failure to detect rare positive events.

C:

Incompatible vector spaces can make nearest-neighbor rankings meaningless. That is not sufficient here: High accuracy can conceal failure to detect rare positive events.

D:

Model specialization may help task behavior but requires measured benefit. That is not sufficient here: High accuracy can conceal failure to detect rare positive events.

Question 585

An Azure AI Fundamentals solution has this requirement: A financial office must extract invoice totals and vendor names from scanned PDFs. Which implementation most directly addresses it?

  1. Use an appropriate search-oriented document analyzer with meaningful chunk boundaries
  2. Use an appropriate Content Understanding or invoice-specific analyzer with a typed field schema
  3. Classify or segment the combined PDF into logical documents before extraction
  4. Use supported locale-aware OCR or language handling with evaluation

Explanation

Correct Answer

B:

Document extraction turns unstructured pages into validated structured fields. This implementation addresses the precise goal rather than an adjacent AI task.

Incorrect Answers

A:

RAG preparation should preserve document structure and source references. That is not sufficient here: Document extraction turns unstructured pages into validated structured fields.

C:

Document boundaries prevent values from unrelated invoices being joined. That is not sufficient here: Document extraction turns unstructured pages into validated structured fields.

D:

Recognition accuracy depends on script and language support. That is not sufficient here: Document extraction turns unstructured pages into validated structured fields.

Question 586

An AI operations team is testing two distinct capabilities. Failure (1): The system books travel after a generated suggestion alone. Failure (2): The team evaluates mean squared distance of vehicle IDs. Which TWO specific corrections should the team apply?

  1. Select visual sign recognition rather than text-only extraction
  2. Require source-linked summarization evaluation
  3. Use entity recognition and verify sensitive-entity recall
  4. Train a binary failure classifier using historical labeled cases
  5. Require a user-approved booking and payment action

Explanation

Correct Answers

D:

Correct for failure (2). A failure probability or class is a discrete decision under a defined horizon. It independently corrects the second defect.

E:

Correct for failure (1). An agent may propose actions but payment must remain permission-controlled. The proposed correction directly addresses that reported defect.

Incorrect Answers

A:

The task chooses a label, with localization needed only when multiple signs appear. The stated pair of failures instead requires “Train a binary failure classifier using historical labeled cases” and “Require a user-approved booking and payment action”; this option cannot substitute for both.

B:

Summarization compresses content without independently validating missing facts. It does not complete the two decisions required in this question.

C:

Entity extraction locates sensitive spans for privacy handling. It does not complete the two decisions required in this question.

Question 587

An AI operations team is testing two distinct capabilities. Failure (1): All non-English characters are dropped during extraction. Failure (2): The model produces a plausible answer with no traceable source span. Which TWO specific corrections should the team apply?

  1. Enable document segmentation and per-document extraction
  2. Configure required invoice fields and validate extracted values
  3. Configure supported language handling and test each script
  4. Keep temporal and content provenance in the retrieval result
  5. Select a supported region or documented alternative

Explanation

Correct Answers

C:

Correct for failure (1). Recognition accuracy depends on script and language support. The proposed correction directly addresses that reported defect.

D:

Correct for failure (2). Multimodal grounding should preserve how the answer maps to source media. It independently corrects the second defect.

Incorrect Answers

A:

Document boundaries prevent values from unrelated invoices being joined. The stated pair of failures instead requires “Configure supported language handling and test each script” and “Keep temporal and content provenance in the retrieval result”; this option cannot substitute for both.

B:

Document extraction turns unstructured pages into validated structured fields. The stated pair of failures instead requires “Configure supported language handling and test each script” and “Keep temporal and content provenance in the retrieval result”; this option cannot substitute for both.

E:

Some modalities and features depend on resource and API version. The stated pair of failures instead requires “Configure supported language handling and test each script” and “Keep temporal and content provenance in the retrieval result”; this option cannot substitute for both.

Question 588

An acceptance tester needs to validate: Evaluate domain adaptation or fine-tuning only after baseline tests show the gap. Which observed result proves the intended behavior?

  1. The resulting cohorts are evaluated for cohesion and business usefulness
  2. Early stopping selects a model before validation performance degrades
  3. A held-out specialty test improves over prompting and retrieval baselines
  4. A user request cannot override the controlling system-level instructions in tests

Explanation

Correct Answer

C:

Model specialization may help task behavior but requires measured benefit. The expected acceptance evidence is: A held-out specialty test improves over prompting and retrieval baselines. This observation directly tests the stated behavior.

Incorrect Answers

A:

Unsupervised learning finds structure without supplied target labels. This observation does not establish the required result: A held-out specialty test improves over prompting and retrieval baselines.

B:

More optimization on training data can worsen generalization. This observation does not establish the required result: A held-out specialty test improves over prompting and retrieval baselines.

D:

Instruction roles help distinguish application policy from user task text. This observation does not establish the required result: A held-out specialty test improves over prompting and retrieval baselines.

Question 589

The approved design action is: Explain that model responses depend on the prompt and available context window. Which test result would BEST establish that it worked?

  1. A new approved price appears in answers after source refresh
  2. A tokenization example shows one word can occupy multiple tokens
  3. Reintroducing relevant context changes the answer in a controlled test
  4. A typed sentence is rendered as understandable speech

Explanation

Correct Answer

C:

A model conditions its output on supplied instructions and conversation text. The expected acceptance evidence is: Reintroducing relevant context changes the answer in a controlled test. This observation directly tests the stated behavior.

Incorrect Answers

A:

External retrieval updates factual context without retraining model weights. This observation does not establish the required result: Reintroducing relevant context changes the answer in a controlled test.

B:

Tokenization splits text into subword or other units for model processing. This observation does not establish the required result: Reintroducing relevant context changes the answer in a controlled test.

D:

Speech synthesis generates audio from text rather than transcribing audio. This observation does not establish the required result: Reintroducing relevant context changes the answer in a controlled test.

Question 590

A solution team has selected this control: Use timestamped multi-segment video analysis for visible actions, with separate higher-frame-rate review when brief security events matter. Which observation is the most direct acceptance evidence?

  1. Observed actions link to supported source intervals while unobserved short events are not claimed as detected
  2. Quoted text matches the time-aligned source transcript
  3. A test invoice yields correct merchant, date and total with source anchors
  4. The decoded barcode matches the printed test label

Explanation

Correct Answer

A:

Content Understanding samples approximately one frame per second, so brief or single-frame actions can be missed despite valid timestamps. The expected acceptance evidence is: Observed actions link to supported source intervals while unobserved short events are not claimed as detected. This observation directly tests the stated behavior.

Incorrect Answers

B:

Generated summaries are paraphrases and not guaranteed verbatim quotes. This observation does not establish the required result: Observed actions link to supported source intervals while unobserved short events are not claimed as detected.

C:

Document extraction turns unstructured pages into validated structured fields. This observation does not establish the required result: Observed actions link to supported source intervals while unobserved short events are not claimed as detected.

D:

A barcode is a structured visual symbol not ordinary OCR text. This observation does not establish the required result: Observed actions link to supported source intervals while unobserved short events are not claimed as detected.

Microsoft Azure AI Fundamentals (AI-901) Practice Tests