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 3

Question 20

An Azure-based AI team must satisfy this objective: A support kiosk must respond to spoken input with synthesized speech. Which design choice is best?

  1. Evaluate noise conditions and input quality rather than only clean studio speech
  2. Match the speech resource region or endpoint to the SDK configuration
  3. Apply approved audio retention and access rules and keep only required derived text
  4. Combine recognition or audio understanding with verified text and text-to-speech

Explanation

Correct Answer

D:

Input transcription and output synthesis are distinct steps. This implementation addresses the precise goal rather than an adjacent AI task.

Incorrect Answers

A:

Signal quality directly affects recognizer error rates. That is not sufficient here: Input transcription and output synthesis are distinct steps.

B:

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

C:

Speech workloads still process personal data and require controlled lifecycle. That is not sufficient here: Input transcription and output synthesis are distinct steps.

Question 21

A solution team has selected this control: Mark synthetic examples clearly and require verified consent for attributed statements. Which observation is the most direct acceptance evidence?

  1. Cross-tenant invoice queries are rejected while own-tenant extraction works
  2. An unverified testimonial is labeled illustrative rather than quoted as authentic
  3. A deaf tester receives the complete equivalent content as text
  4. A delete request fails while the permitted read operation succeeds

Explanation

Correct Answer

B:

Transparency prohibits presenting invented testimony as genuine. The expected acceptance evidence is: An unverified testimonial is labeled illustrative rather than quoted as authentic. This observation directly tests the stated behavior.

Incorrect Answers

A:

Model instructions cannot grant access beyond the underlying data permission boundary. This observation does not establish the required result: An unverified testimonial is labeled illustrative rather than quoted as authentic.

C:

Accessible information must not depend on hearing audio. This observation does not establish the required result: An unverified testimonial is labeled illustrative rather than quoted as authentic.

D:

Least privilege reduces damage from compromised prompts or tool output. This observation does not establish the required result: An unverified testimonial is labeled illustrative rather than quoted as authentic.

Question 22

An Azure AI solution uses the following approved control: Use Speech Synthesis Markup Language pauses or breaks where supported. What observable result should pass QA?

  1. The narration pauses at marked section boundaries
  2. A confirmed final transcript triggers the intended action once
  3. Noise-augmented recognition tests meet requirements or request repetition
  4. The model's response reflects the words spoken in the actual recording

Explanation

Correct Answer

A:

SSML controls aspects of prosody and timing for synthesized speech. The expected acceptance evidence is: The narration pauses at marked section boundaries. This observation directly tests the stated behavior.

Incorrect Answers

B:

Interim hypotheses can change as more audio arrives. This observation does not establish the required result: The narration pauses at marked section boundaries.

C:

Signal quality directly affects recognizer error rates. This observation does not establish the required result: The narration pauses at marked section boundaries.

D:

A text-only request cannot interpret sound content by file name alone. This observation does not establish the required result: The narration pauses at marked section boundaries.

Question 23

Troubleshooting report: The application uploads entire unrelated calls for one short keyword. Which correction is MOST directly targeted to this failure?

  1. Extract relevant consented segments instead
  2. Use entity extraction instead of sentiment classification
  3. Add translation and French synthesis after recognition
  4. Control synthesis prosody in SSML

Explanation

Correct Answer

A:

This correction directly addresses the reported failure. Minimization reduces unnecessary personal-data exposure. The targeted change is: Extract relevant consented segments instead.

Incorrect Answers

B:

NER identifies named spans rather than summarizing the email. In the stated case, troubleshooting report: The application uploads entire unrelated calls for one short keyword; “Use entity extraction instead of sentiment classification” does not target that failure.

C:

Cross-language audio output requires both language transfer and synthesis. In the stated case, troubleshooting report: The application uploads entire unrelated calls for one short keyword; “Add translation and French synthesis after recognition” does not target that failure.

D:

Voice selection and prosody can influence delivery without changing the text. In the stated case, troubleshooting report: The application uploads entire unrelated calls for one short keyword; “Control synthesis prosody in SSML” does not target that failure.

Question 24

Troubleshooting report: The scoring pipeline retains a postcode that acts as an unchecked proxy. Which correction is MOST directly targeted to this failure?

  1. Expand the consented voice test corpus and provide fallback interaction
  2. Rebalance safety tests and adjust erroneous dialect associations
  3. Audit correlated features and document the business necessity of each factor
  4. Assess both unsafe misses and harmful overblocking

Explanation

Correct Answer

C:

This correction directly addresses the reported failure. Legitimate credit criteria cannot be replaced by sensitive identity traits. The targeted change is: Audit correlated features and document the business necessity of each factor.

Incorrect Answers

A:

Inclusive voice systems require performance evaluation beyond a single accent. In the stated case, troubleshooting report: The scoring pipeline retains a postcode that acts as an unchecked proxy; “Expand the consented voice test corpus and provide fallback interaction” does not target that failure.

B:

Fairness requires discriminating actual harm from linguistic variety. In the stated case, troubleshooting report: The scoring pipeline retains a postcode that acts as an unchecked proxy; “Rebalance safety tests and adjust erroneous dialect associations” does not target that failure.

D:

Safety controls should not overblock permitted assistance. In the stated case, troubleshooting report: The scoring pipeline retains a postcode that acts as an unchecked proxy; “Assess both unsafe misses and harmful overblocking” does not target that failure.

Question 25

Troubleshooting report: An unrelated generic summary is generated from the webinar title only. Which correction is MOST directly targeted to this failure?

  1. Preserve speaker identities and confirm relevant transcript spans
  2. Ground the summary in extracted transcript segments
  3. Embed or semantically index the documents with source provenance
  4. Evaluate noise conditions and request clarification at low confidence

Explanation

Correct Answer

B:

This correction directly addresses the reported failure. Meeting summaries require temporal and spoken source content. The targeted change is: Ground the summary in extracted transcript segments.

Incorrect Answers

A:

Attribution distinguishes who committed to each action. In the stated case, troubleshooting report: An unrelated generic summary is generated from the webinar title only; “Preserve speaker identities and confirm relevant transcript spans” does not target that failure.

C:

Semantic search can retrieve conceptually related passages. In the stated case, troubleshooting report: An unrelated generic summary is generated from the webinar title only; “Embed or semantically index the documents with source provenance” does not target that failure.

D:

Speech input quality affects subsequent language classification. In the stated case, troubleshooting report: An unrelated generic summary is generated from the webinar title only; “Evaluate noise conditions and request clarification at low confidence” does not target that failure.

Question 26

An AI operations team is testing two distinct capabilities. Failure (1): The prompt asks a model to improvise tax totals without a rules engine. Failure (2): The team sets fully deterministic output and expects broad variation. Which TWO specific corrections should the team apply?

  1. Test a higher diversity setting and compare evaluated outputs
  2. Refresh indexed grounding sources and verify their effective date
  3. Implement and test fixed numeric rules in application code
  4. Select a classification output and calibrated decision boundary
  5. Set the API request to the correct deployment name and endpoint

Explanation

Correct Answers

A:

Correct for failure (2). More varied token sampling can increase creative alternatives. It independently corrects the second defect.

C:

Correct for failure (1). Generative models are not a substitute for auditable deterministic computation. The proposed correction directly addresses that reported defect.

Incorrect Answers

B:

External retrieval updates factual context without retraining model weights. In this case the first outcome requires the team to use a conventional calculation function for exact arithmetic and let the model explain it, and the second requires it to adjust supported sampling parameters while preserving safety and quality tests. This option does not replace those two separate controls.

D:

Classification predicts discrete labels rather than a continuous numeric target. In this case the first outcome requires the team to use a conventional calculation function for exact arithmetic and let the model explain it, and the second requires it to adjust supported sampling parameters while preserving safety and quality tests. This option does not replace those two separate controls.

E:

Applications call a deployed model resource rather than a generic marketing model name. In this case the first outcome requires the team to use a conventional calculation function for exact arithmetic and let the model explain it, and the second requires it to adjust supported sampling parameters while preserving safety and quality tests. This option does not replace those two separate controls.

Question 27

An Azure-based AI team must satisfy this objective: A logistics workflow receives forms with tables and handwritten notes. Which design choice is best?

  1. Use timestamped multi-segment video analysis for visible actions, with separate higher-frame-rate review when brief security events matter
  2. Classify document type before routing to the relevant extraction analyzer
  3. Use layout-aware document analysis that preserves rows, selection marks and text spans
  4. Use an appropriate Content Understanding or invoice-specific analyzer with a typed field schema

Explanation

Correct Answer

C:

Ordinary plain-text OCR may lose table relationships and checkbox states. This implementation addresses the precise goal rather than an adjacent AI task.

Incorrect Answers

A:

Content Understanding samples approximately one frame per second, so brief or single-frame actions can be missed despite valid timestamps. That is not sufficient here: Ordinary plain-text OCR may lose table relationships and checkbox states.

B:

Different document kinds have distinct field schemas and review paths. That is not sufficient here: Ordinary plain-text OCR may lose table relationships and checkbox states.

D:

Document extraction turns unstructured pages into validated structured fields. That is not sufficient here: Ordinary plain-text OCR may lose table relationships and checkbox states.

Question 28

An engineering review documents this requirement: A user asks an AI agent to approve a large money transfer through persuasive language. Select the most appropriate solution.

  1. Minimize unnecessary identifying visual data and enforce approved access
  2. Restrict processing, inference, logs and backups to authorized regions
  3. Require independent transaction authorization and approval outside prompts
  4. Identify the virtual agent as synthetic and avoid deceptive identity cues

Explanation

Correct Answer

C:

Accountable high-impact operations need verifiable authorized signatures. This implementation addresses the precise goal rather than an adjacent AI task.

Incorrect Answers

A:

Privacy applies to image content as well as text transcripts. That is not sufficient here: Accountable high-impact operations need verifiable authorized signatures.

B:

Privacy and compliance require controlling the full data path. That is not sufficient here: Accountable high-impact operations need verifiable authorized signatures.

D:

Transparency prevents users from falsely assuming they are speaking with a human. That is not sufficient here: Accountable high-impact operations need verifiable authorized signatures.

Question 29

The approved design action is: Run adversarial model evaluations and verify that configured content-safety guardrails enforce the stated high-severity block criteria. Which test result would BEST establish that it worked?

  1. Red-team tests verify the configured safety layer blocks the required high-severity cases after the update
  2. The agent selects the expected tool for a supported test query
  3. The answer mentions relevant visual details unique to the uploaded image
  4. A load test recovers without a retry storm or duplicate effects

Explanation

Correct Answer

A:

Adversarial testing must measure actual configured policy enforcement rather than treating a prompt instruction as a guaranteed safety boundary. The expected acceptance evidence is: Red-team tests verify the configured safety layer blocks the required high-severity cases after the update. This observation directly tests the stated behavior.

Incorrect Answers

B:

A single agent can orchestrate a bounded tool without a multi-agent system. This observation does not establish the required result: Red-team tests verify the configured safety layer blocks the required high-severity cases after the update.

C:

The model must receive the actual visual content to reason about it. This observation does not establish the required result: Red-team tests verify the configured safety layer blocks the required high-severity cases after the update.

D:

Rate limits protect the service and uncontrolled retries amplify overload. This observation does not establish the required result: Red-team tests verify the configured safety layer blocks the required high-severity cases after the update.

Question 30

Troubleshooting report: The explanation equates one token to one full English word in all cases. Which correction is MOST directly targeted to this failure?

  1. Chunk and retrieve relevant source sections before inference
  2. Demonstrate tokenization differences across words and scripts
  3. Implement and test fixed numeric rules in application code
  4. Re-embed the corpus or use a versioned compatible index

Explanation

Correct Answer

B:

This correction directly addresses the reported failure. Tokenization splits text into subword or other units for model processing. The targeted change is: Demonstrate tokenization differences across words and scripts.

Incorrect Answers

A:

A finite context window limits how much material can be processed at once. In the stated case, troubleshooting report: The explanation equates one token to one full English word in all cases; “Chunk and retrieve relevant source sections before inference” does not target that failure.

C:

Generative models are not a substitute for auditable deterministic computation. In the stated case, troubleshooting report: The explanation equates one token to one full English word in all cases; “Implement and test fixed numeric rules in application code” does not target that failure.

D:

Incompatible vector spaces can make nearest-neighbor rankings meaningless. In the stated case, troubleshooting report: The explanation equates one token to one full English word in all cases; “Re-embed the corpus or use a versioned compatible index” does not target that failure.

Microsoft Azure AI Fundamentals (AI-901) Practice Tests