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 64

Question 630

An acceptance tester needs to validate: Use the box coordinates to locate detected people in the source image. Which observed result proves the intended behavior?

  1. A drawn overlay aligns each box with the corresponding person
  2. The result includes supported text, tags and person boxes without conflation
  3. The product outline remains intact after background editing
  4. The summary preserves the order of observed demonstration steps and identifies gaps or uncertain actions

Explanation

Correct Answer

A:

Localization metadata is tied to pixel regions, not just labels. The expected acceptance evidence is: A drawn overlay aligns each box with the corresponding person. This observation directly tests the stated behavior.

Incorrect Answers

B:

Different visual outputs describe text, scene classes and detected people. This observation does not establish the required result: A drawn overlay aligns each box with the corresponding person.

C:

A mask defines which image area may change. This observation does not establish the required result: A drawn overlay aligns each box with the corresponding person.

D:

A single still cannot convey action order; sampled video frames may also omit fleeting intermediate steps. This observation does not establish the required result: A drawn overlay aligns each box with the corresponding person.

Question 631

Troubleshooting report: The agent invents an article title after an empty retrieval. Which correction is MOST directly targeted to this failure?

  1. Constrain request size and trim or chunk input before sending
  2. Prune irrelevant history and preserve the minimum required state
  3. Create and test a simple one-agent tool workflow
  4. Use an abstention or clarification branch when no source matches

Explanation

Correct Answer

D:

This correction directly addresses the reported failure. An empty result is not positive evidence for a generated factual claim. The targeted change is: Use an abstention or clarification branch when no source matches.

Incorrect Answers

A:

Capabilities and limits vary by model deployment. In the stated case, troubleshooting report: The agent invents an article title after an empty retrieval; “Constrain request size and trim or chunk input before sending” does not target that failure.

B:

Long irrelevant context consumes tokens and may increase latency. In the stated case, troubleshooting report: The agent invents an article title after an empty retrieval; “Prune irrelevant history and preserve the minimum required state” does not target that failure.

C:

A single agent can orchestrate a bounded tool without a multi-agent system. In the stated case, troubleshooting report: The agent invents an article title after an empty retrieval; “Create and test a simple one-agent tool workflow” does not target that failure.

Question 632

Troubleshooting report: The app posts a total solely because the extracted confidence is high. Which correction is MOST directly targeted to this failure?

  1. Select a supported region or documented alternative
  2. Run a supported OCR or visual text extraction feature
  3. Keep transcript anchors and validate extracted tasks
  4. Run deterministic line-item reconciliation before acceptance

Explanation

Correct Answer

D:

This correction directly addresses the reported failure. Independent business rules catch plausible extraction errors. The targeted change is: Run deterministic line-item reconciliation before acceptance.

Incorrect Answers

A:

Some modalities and features depend on resource and API version. In the stated case, troubleshooting report: The app posts a total solely because the extracted confidence is high; “Select a supported region or documented alternative” does not target that failure.

B:

OCR is appropriate for characters embedded in a photograph. In the stated case, troubleshooting report: The app posts a total solely because the extracted confidence is high; “Run a supported OCR or visual text extraction feature” does not target that failure.

C:

Transcription and summarization are separate results with different validation. In the stated case, troubleshooting report: The app posts a total solely because the extracted confidence is high; “Keep transcript anchors and validate extracted tasks” does not target that failure.

Question 633

An Azure-based AI team must satisfy this objective: A generative model is evaluated only on common customer messages. Which design choice is best?

  1. Include rare but consequential safety edge cases in acceptance tests
  2. Record the actual decision output, input reference, model and prompt versions in integrity-protected audit records
  3. Evaluate language safety on dialect-representative benign and harmful examples
  4. Require governance evidence with reviewers, test outcomes, and documented exceptions

Explanation

Correct Answer

A:

A high overall score can conceal severe failure on unusual inputs. This implementation addresses the precise goal rather than an adjacent AI task.

Incorrect Answers

B:

Auditability requires the original decision evidence and version identifiers; regenerating model text is not guaranteed to be identical. That is not sufficient here: A high overall score can conceal severe failure on unusual inputs.

C:

Fairness requires discriminating actual harm from linguistic variety. That is not sufficient here: A high overall score can conceal severe failure on unusual inputs.

D:

Self-attestation is not independent proof of regulatory controls. That is not sufficient here: A high overall score can conceal severe failure on unusual inputs.

Question 634

The approved design action is: Run grounded summarization after speech recognition and preserve source segments. Which test result would BEST establish that it worked?

  1. Spanish test utterances are transcribed correctly and consistently
  2. Subtitle cues match the spoken phrases at measured times
  3. The acronym is spoken according to the approved pronunciation sample
  4. Each action item maps to an utterance and assigned speaker

Explanation

Correct Answer

D:

Transcription obtains words; summarization extracts decisions and tasks. The expected acceptance evidence is: Each action item maps to an utterance and assigned speaker. This observation directly tests the stated behavior.

Incorrect Answers

A:

Recognition models need the appropriate supported language settings. This observation does not establish the required result: Each action item maps to an utterance and assigned speaker.

B:

Timecodes align text segments to audio playback. This observation does not establish the required result: Each action item maps to an utterance and assigned speaker.

C:

SSML can guide pronunciation beyond plain-text guessing. This observation does not establish the required result: Each action item maps to an utterance and assigned speaker.

Question 635

An Azure AI solution uses the following approved control: Inspect whether the tool is registered and appropriately described to the agent. What observable result should pass QA?

  1. Execution traces show the intended tool invoked with valid arguments
  2. A valid identity with required resource role succeeds in the test
  3. A test run shows correct tool selection, arguments and user-visible response
  4. The target model deploys in the approved region and meets residency controls

Explanation

Correct Answer

A:

Tools must be discoverable with valid schemas before selection. The expected acceptance evidence is: Execution traces show the intended tool invoked with valid arguments. This observation directly tests the stated behavior.

Incorrect Answers

B:

Authentication proves caller identity and authorization grants resource access. This observation does not establish the required result: Execution traces show the intended tool invoked with valid arguments.

C:

An answer that sounds plausible may conceal failed tool calls. This observation does not establish the required result: Execution traces show the intended tool invoked with valid arguments.

D:

Not every model or feature is available in every location. This observation does not establish the required result: Execution traces show the intended tool invoked with valid arguments.

Question 636

Troubleshooting report: The model follows a user instruction to expose private internal prompts. Which correction is MOST directly targeted to this failure?

  1. Add adversarial cases and verify configured safety block and escalation behavior
  2. Apply case and role filters before providing model context
  3. Move the model call behind a secured server endpoint
  4. Reinforce role boundaries and enforce secrets outside prompts

Explanation

Correct Answer

D:

This correction directly addresses the reported failure. Developer and system constraints govern allowed operations over user requests. The targeted change is: Reinforce role boundaries and enforce secrets outside prompts.

Incorrect Answers

A:

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 model follows a user instruction to expose private internal prompts; “Add adversarial cases and verify configured safety block and escalation behavior” does not target that failure.

B:

Source scoping enforces privacy and also improves answer relevance. In the stated case, troubleshooting report: The model follows a user instruction to expose private internal prompts; “Apply case and role filters before providing model context” does not target that failure.

C:

Backend controls protect credentials, quotas and data policy. In the stated case, troubleshooting report: The model follows a user instruction to expose private internal prompts; “Move the model call behind a secured server endpoint” does not target that failure.

Question 637

An engineering review documents this requirement: A model rejects all unfamiliar dialects as unsafe language. Select the most appropriate solution.

  1. Evaluate instructions and fallback quality in each supported language
  2. Evaluate language safety on dialect-representative benign and harmful examples
  3. Separate informational assistance from qualified professional decisions and express uncertainty
  4. Identify the virtual agent as synthetic and avoid deceptive identity cues

Explanation

Correct Answer

B:

Fairness requires discriminating actual harm from linguistic variety. This implementation addresses the precise goal rather than an adjacent AI task.

Incorrect Answers

A:

Inclusion requires equivalent safety and meaning across supported locales. That is not sufficient here: Fairness requires discriminating actual harm from linguistic variety.

C:

Transparency reduces mistaken reliance on unsupported professional guarantees. That is not sufficient here: Fairness requires discriminating actual harm from linguistic variety.

D:

Transparency prevents users from falsely assuming they are speaking with a human. That is not sufficient here: Fairness requires discriminating actual harm from linguistic variety.

Question 638

An acceptance tester needs to validate: Implement a network timeout/retry strategy that avoids duplicate tool effects. Which observed result proves the intended behavior?

  1. A network interruption does not create duplicate business records
  2. Red-team tests verify the configured safety layer blocks the required high-severity cases after the update
  3. The same task passes with reduced token use and no quality loss
  4. The service processes role-tagged messages with the expected application behavior

Explanation

Correct Answer

A:

Client retries should distinguish read requests from side-effecting tool calls. The expected acceptance evidence is: A network interruption does not create duplicate business records. This observation directly tests the stated behavior.

Incorrect Answers

B:

Adversarial testing must measure actual configured policy enforcement rather than treating a prompt instruction as a guaranteed safety boundary. This observation does not establish the required result: A network interruption does not create duplicate business records.

C:

Long irrelevant context consumes tokens and may increase latency. This observation does not establish the required result: A network interruption does not create duplicate business records.

D:

Chat APIs use structured messages rather than concatenating untyped command strings. This observation does not establish the required result: A network interruption does not create duplicate business records.

Question 639

A project needs the following AI behavior: An agent invocation fails when the tool response is delayed. What should the designer implement?

  1. Use typed state fields and validate identifiers at each tool boundary
  2. Use representative test prompts and inspect execution and tool traces
  3. Handle tool timeout as a typed error and preserve partial workflow state
  4. Inspect whether the tool is registered and appropriately described to the agent

Explanation

Correct Answer

C:

Timeout handling should not be reported as a successful action. This implementation addresses the precise goal rather than an adjacent AI task.

Incorrect Answers

A:

Structured state prevents subtle substitution of unrelated customer records. That is not sufficient here: Timeout handling should not be reported as a successful action.

B:

An answer that sounds plausible may conceal failed tool calls. That is not sufficient here: Timeout handling should not be reported as a successful action.

D:

Tools must be discoverable with valid schemas before selection. That is not sufficient here: Timeout handling should not be reported as a successful action.

Question 640

An Azure AI Fundamentals solution has this requirement: A team wants to replace a photograph's background while preserving the original product. Which implementation most directly addresses it?

  1. Use image classification for the requested whole-image class
  2. Use segmentation or masked image editing that locks product pixels
  3. Provide the actual image to a capable multimodal or OCR pipeline
  4. Apply bounded retries and respect the service's throughput limit

Explanation

Correct Answer

B:

A mask defines which image area may change. This implementation addresses the precise goal rather than an adjacent AI task.

Incorrect Answers

A:

Image-level classification is distinct from object localization. That is not sufficient here: A mask defines which image area may change.

C:

Visual evidence is needed to read sign content. That is not sufficient here: A mask defines which image area may change.

D:

Too many concurrent requests can exceed resource quota. That is not sufficient here: A mask defines which image area may change.

Microsoft Azure AI Fundamentals (AI-901) Practice Tests