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 39

Question 380

An Azure AI solution uses the following approved control: Detect contradictory extracted values and route uncertain cases for review. What observable result should pass QA?

  1. The answer resolves to a specific approved source document version
  2. A reply links its claim to a transcript time or relevant video segment
  3. The result reports both candidate values and cites their locations
  4. The output returns correct limits with supporting clauses

Explanation

Correct Answer

C:

A model should not silently choose among competing contractual evidence. The expected acceptance evidence is: The result reports both candidate values and cites their locations. This observation directly tests the stated behavior.

Incorrect Answers

A:

Verifiable provenance differentiates a cited answer from guesswork. This observation does not establish the required result: The result reports both candidate values and cites their locations.

B:

Multimodal grounding should preserve how the answer maps to source media. This observation does not establish the required result: The result reports both candidate values and cites their locations.

D:

Custom extraction maps semantically relevant phrases to typed fields. This observation does not establish the required result: The result reports both candidate values and cites their locations.

Question 381

An Azure-based AI team must satisfy this objective: A team's API client gets a deployment-not-found response. Which design choice is best?

  1. Audit training labels and evaluate corrected outcome criteria
  2. Check output-token limits and context-budget allocation
  3. Review precision, recall and the confusion matrix rather than accuracy alone
  4. Confirm the model deployment identifier and endpoint resource match

Explanation

Correct Answer

D:

Applications call a deployed model resource rather than a generic marketing model name. This implementation addresses the precise goal rather than an adjacent AI task.

Incorrect Answers

A:

Bias can enter through the targets themselves, not just inputs. That is not sufficient here: Applications call a deployed model resource rather than a generic marketing model name.

B:

Generation stops when an output budget or stop condition is reached. That is not sufficient here: Applications call a deployed model resource rather than a generic marketing model name.

C:

High accuracy can conceal failure to detect rare positive events. That is not sufficient here: Applications call a deployed model resource rather than a generic marketing model name.

Question 382

A project needs the following AI behavior: An agent needs to decide which tool to invoke after reading a customer request. What should the designer implement?

  1. Choose a text-classification or sentiment-analysis capability rather than an image generator
  2. Use an agent-capable orchestration flow with bounded tool definitions
  3. Use a carefully constrained prompt and a lower sampling temperature where supported
  4. Measure expected tokens and requests per minute, then choose an approved throughput tier or capacity adjustment before launch

Explanation

Correct Answer

B:

Agent workflows can reason about tool selection but tools still need authorization. This implementation addresses the precise goal rather than an adjacent AI task.

Incorrect Answers

A:

Different model tasks require compatible outputs and labeled evaluation data. That is not sufficient here: Agent workflows can reason about tool selection but tools still need authorization.

C:

Lower randomness often improves consistency but does not guarantee correctness. That is not sufficient here: Agent workflows can reason about tool selection but tools still need authorization.

D:

Quota sizing is a deployment-capacity decision; retrying a predictable sustained excess cannot add model throughput. That is not sufficient here: Agent workflows can reason about tool selection but tools still need authorization.

Question 383

Troubleshooting report: The agent relies on free-form generation when retrieval is empty. Which correction is MOST directly targeted to this failure?

  1. Require source comparison and abstention on missing evidence
  2. Enforce source filters and explicit missing-evidence handling
  3. Use code validation and protected deployment controls
  4. Use an abstention or clarification branch when no source matches

Explanation

Correct Answer

B:

This correction directly addresses the reported failure. Retrieval access and abstention avoid inventing unsupported documentation. The targeted change is: Enforce source filters and explicit missing-evidence handling.

Incorrect Answers

A:

Grounding constrains the answer to the available text and provenance. In the stated case, troubleshooting report: The agent relies on free-form generation when retrieval is empty; “Require source comparison and abstention on missing evidence” does not target that failure.

C:

Generated code must be validated, not treated as production-ready from fluency. In the stated case, troubleshooting report: The agent relies on free-form generation when retrieval is empty; “Use code validation and protected deployment controls” does not target that failure.

D:

An empty result is not positive evidence for a generated factual claim. In the stated case, troubleshooting report: The agent relies on free-form generation when retrieval is empty; “Use an abstention or clarification branch when no source matches” does not target that failure.

Question 384

Troubleshooting report: The workflow writes entire identity scans to public diagnostic storage. Which correction is MOST directly targeted to this failure?

  1. Build citations from stored IDs and source spans
  2. Validate field presence against source evidence
  3. Redact logs and use access-controlled short-lived source storage
  4. Preserve actual section/page references through retrieval

Explanation

Correct Answer

C:

This correction directly addresses the reported failure. Inference outputs and telemetry can contain protected identity information. The targeted change is: Redact logs and use access-controlled short-lived source storage.

Incorrect Answers

A:

Verifiable provenance differentiates a cited answer from guesswork. In the stated case, troubleshooting report: The workflow writes entire identity scans to public diagnostic storage; “Build citations from stored IDs and source spans” does not target that failure.

B:

Extraction should preserve absence and distinguish it from zero. In the stated case, troubleshooting report: The workflow writes entire identity scans to public diagnostic storage; “Validate field presence against source evidence” does not target that failure.

D:

Structured layout and provenance let users verify a quoted clause. In the stated case, troubleshooting report: The workflow writes entire identity scans to public diagnostic storage; “Preserve actual section/page references through retrieval” does not target that failure.

Question 385

An Azure-based AI team must satisfy this objective: A search system needs to compare semantic meaning between document passages. Which design choice is best?

  1. Choose a text-classification or sentiment-analysis capability rather than an image generator
  2. Use an agent-capable orchestration flow with bounded tool definitions
  3. Represent text with embeddings and compare vectors using an appropriate similarity measure
  4. Use retrieval-augmented generation from current approved catalog sources

Explanation

Correct Answer

C:

Embeddings encode learned semantic relationships rather than literal string identity. This implementation addresses the precise goal rather than an adjacent AI task.

Incorrect Answers

A:

Different model tasks require compatible outputs and labeled evaluation data. That is not sufficient here: Embeddings encode learned semantic relationships rather than literal string identity.

B:

Agent workflows can reason about tool selection but tools still need authorization. That is not sufficient here: Embeddings encode learned semantic relationships rather than literal string identity.

D:

External retrieval updates factual context without retraining model weights. That is not sufficient here: Embeddings encode learned semantic relationships rather than literal string identity.

Question 386

Troubleshooting report: The prompt hardcodes pricing that the app never updates. Which correction is MOST directly targeted to this failure?

  1. Keep policy instructions separate from lower-trust data
  2. Add adversarial cases and verify configured safety block and escalation behavior
  3. Enforce permitted methods at the service boundary
  4. Move volatile facts into a governed retrieval source

Explanation

Correct Answer

D:

This correction directly addresses the reported failure. Static prompts are unsuitable as a changing source of commercial truth. The targeted change is: Move volatile facts into a governed retrieval source.

Incorrect Answers

A:

High-priority instructions should describe approved behavior and escalation rules. In the stated case, troubleshooting report: The prompt hardcodes pricing that the app never updates; “Keep policy instructions separate from lower-trust data” does not target that failure.

B:

Adversarial testing must measure actual configured policy enforcement rather than treating a prompt instruction as a guaranteed safety boundary. In the stated case, troubleshooting report: The prompt hardcodes pricing that the app never updates; “Add adversarial cases and verify configured safety block and escalation behavior” does not target that failure.

C:

Least-privilege tool methods reduce accidental external actions. In the stated case, troubleshooting report: The prompt hardcodes pricing that the app never updates; “Enforce permitted methods at the service boundary” does not target that failure.

Question 387

The approved design action is: Use image classification for the requested category. Which test result would BEST establish that it worked?

  1. Known spikes score unusually relative to a validated baseline
  2. Indoor/outdoor labels meet validation accuracy on held-out photos
  3. Test feedback returns relevant phrases found in the message
  4. Totals, dates and merchant names match receipt evidence

Explanation

Correct Answer

B:

Image classification predicts image-level labels. The expected acceptance evidence is: Indoor/outdoor labels meet validation accuracy on held-out photos. This observation directly tests the stated behavior.

Incorrect Answers

A:

Anomaly methods identify deviations rather than assigning preset incident labels. This observation does not establish the required result: Indoor/outdoor labels meet validation accuracy on held-out photos.

C:

Key phrase extraction finds prominent topics without inventing new source material. This observation does not establish the required result: Indoor/outdoor labels meet validation accuracy on held-out photos.

D:

Images of receipts need OCR/layout extraction before typed business values. This observation does not establish the required result: Indoor/outdoor labels meet validation accuracy on held-out photos.

Question 388

An Azure AI solution uses the following approved control: Enable or use supported punctuation formatting with appropriate post-processing. What observable result should pass QA?

  1. Labeled specialty phrases are transcribed with reduced errors
  2. A labeled command yields readable sentence boundaries
  3. A test email returns correct product and organization spans
  4. The sanitized SSML parses and renders the intended characters

Explanation

Correct Answer

B:

Speech text may require punctuation inference distinct from word recognition. The expected acceptance evidence is: A labeled command yields readable sentence boundaries. This observation directly tests the stated behavior.

Incorrect Answers

A:

Domain-specific words may require biasing or adaptation beyond a generic recognizer. This observation does not establish the required result: A labeled command yields readable sentence boundaries.

C:

NER identifies named spans rather than summarizing the email. This observation does not establish the required result: A labeled command yields readable sentence boundaries.

D:

SSML is XML and malformed markup breaks the synthesis request. This observation does not establish the required result: A labeled command yields readable sentence boundaries.

Question 389

An Azure-based AI team must satisfy this objective: A signed contract contains checkboxes and a signature region. Which design choice is best?

  1. Create a schema for policy type, coverage amount and effective dates
  2. Extract checkbox selection marks with supported document analysis, and verify the visible signature separately; use signature detection only if its preview API is approved
  3. Use audio extraction followed by grounded action-item analysis
  4. Preserve table row-column structure and validate numeric totals

Explanation

Correct Answer

B:

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.

Incorrect Answers

A:

Custom extraction maps semantically relevant phrases to typed fields. That is not sufficient here: 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.

C:

Transcription and summarization are separate results with different validation. That is not sufficient here: 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.

D:

Cell relationships are necessary to check item totals. That is not sufficient here: 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.

Question 390

An Azure AI Fundamentals solution has this requirement: A safety filter allows an answer after retrieving a malicious document. Which implementation most directly addresses it?

  1. Evaluate outcomes with counterfactual name swaps and remove unjustified proxy effects
  2. Test representative document layouts and accessibility-oriented inputs
  3. Evaluate safety again after tool and retrieval results are incorporated
  4. Use grounded medical content, high-risk detection and appropriate expert escalation

Explanation

Correct Answer

C:

Threats can enter later stages even when the original user request was harmless. 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: Threats can enter later stages even when the original user request was harmless.

B:

Inclusion requires robust input handling beyond a single document template. That is not sufficient here: Threats can enter later stages even when the original user request was harmless.

D:

Safety mechanisms should reduce unsupported high-impact recommendations. That is not sufficient here: Threats can enter later stages even when the original user request was harmless.

Microsoft Azure AI Fundamentals (AI-901) Practice Tests