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 50

Question 490

An Azure AI solution uses the following approved control: Honor service throttling and use bounded exponential backoff. What observable result should pass QA?

  1. An unconfirmed action remains pending despite a persuasive request
  2. A load test recovers without a retry storm or duplicate effects
  3. Known unsupported claims remain flagged regardless of temperature
  4. The agent selects the expected tool for a supported test query

Explanation

Correct Answer

B:

Rate limits protect the service and uncontrolled retries amplify overload. The expected acceptance evidence is: A load test recovers without a retry storm or duplicate effects. This observation directly tests the stated behavior.

Incorrect Answers

A:

High-impact side effects need permissions independent of conversational language. This observation does not establish the required result: A load test recovers without a retry storm or duplicate effects.

C:

Sampling parameters influence diversity but do not verify knowledge. This observation does not establish the required result: A load test recovers without a retry storm or duplicate effects.

D:

A single agent can orchestrate a bounded tool without a multi-agent system. This observation does not establish the required result: A load test recovers without a retry storm or duplicate effects.

Question 491

Troubleshooting report: Exact keyword matching is the only retrieval method. Which correction is MOST directly targeted to this failure?

  1. Use classification outputs for the entire image
  2. Preserve speaker identities and confirm relevant transcript spans
  3. Combine vector similarity with structured catalog filtering
  4. Use sentiment analysis instead of named-entity extraction

Explanation

Correct Answer

C:

This correction directly addresses the reported failure. Semantic vectors capture related concepts while filters enforce eligibility. The targeted change is: Combine vector similarity with structured catalog filtering.

Incorrect Answers

A:

Image classification predicts image-level labels. In the stated case, troubleshooting report: Exact keyword matching is the only retrieval method; “Use classification outputs for the entire image” does not target that failure.

B:

Attribution distinguishes who committed to each action. In the stated case, troubleshooting report: Exact keyword matching is the only retrieval method; “Preserve speaker identities and confirm relevant transcript spans” does not target that failure.

D:

Sentiment analysis infers expressed attitude rather than extracting names. In the stated case, troubleshooting report: Exact keyword matching is the only retrieval method; “Use sentiment analysis instead of named-entity extraction” does not target that failure.

Question 492

An AI operations team is testing two distinct capabilities. Failure (1): A query filters by document title instead of authenticated tenant ID. Failure (2): The agent can call a payment API directly with no server-side limit. Which TWO specific corrections should the team apply?

  1. Move authorization into server-side record filtering
  2. Mask or exclude unauthorized identity details before publishing
  3. Gate the payment tool with verified role and amount checks
  4. Run the frozen adversarial and ordinary acceptance suite against the candidate
  5. Propagate typed tool errors and use an explicit unavailable response

Explanation

Correct Answers

A:

Correct for failure (1). Model instructions cannot grant access beyond the underlying data permission boundary. The proposed correction directly addresses that reported defect.

C:

Correct for failure (2). A prompt-level request cannot override a transaction policy. It independently corrects the second defect.

Incorrect Answers

B:

Privacy applies to image content as well as text transcripts. The stated pair of failures instead requires “Move authorization into server-side record filtering” and “Gate the payment tool with verified role and amount checks”; this option cannot substitute for both.

D:

Reliability includes preventing behavior regression across model versions. It does not complete the two decisions required in this question.

E:

A fluent generation cannot substitute for missing authoritative transaction data. The stated pair of failures instead requires “Move authorization into server-side record filtering” and “Gate the payment tool with verified role and amount checks”; this option cannot substitute for both.

Question 493

An AI operations team is testing two distinct capabilities. Failure (1): State from a previous customer is reused for a new login. Failure (2): Only total monthly bill is visible with no per-call metrics. Which TWO specific corrections should the team apply?

  1. Filter data at the tool boundary using approved policies
  2. Reinforce role boundaries and enforce secrets outside prompts
  3. Capture per-request usage and attribute it to versioned deployments
  4. Add adversarial cases and verify configured safety block and escalation behavior
  5. Partition conversation context by identity and expiration

Explanation

Correct Answers

C:

Correct for failure (2). Usage telemetry reveals context-growth and unexpected repetition. It independently corrects the second defect.

E:

Correct for failure (1). Conversation continuity should preserve identifiers without crossing user sessions. The proposed correction directly addresses that reported defect.

Incorrect Answers

A:

Tool descriptions cannot override data-loss-prevention boundaries. It does not complete the two decisions required in this question.

B:

Developer and system constraints govern allowed operations over user requests. The stated pair of failures instead requires “Capture per-request usage and attribute it to versioned deployments” and “Partition conversation context by identity and expiration”; this option cannot substitute for both.

D:

Adversarial testing must measure actual configured policy enforcement rather than treating a prompt instruction as a guaranteed safety boundary. The stated pair of failures instead requires “Capture per-request usage and attribute it to versioned deployments” and “Partition conversation context by identity and expiration”; this option cannot substitute for both.

Question 494

An acceptance tester needs to validate: Explain token-based next-token prediction conditioned on preceding context. Which observed result proves the intended behavior?

  1. A tokenization example shows one word can occupy multiple tokens
  2. A regression test records model version and key generation settings
  3. The output is produced as a sequence of tokens conditioned on the prompt
  4. A permitted larger output budget allows the complete response

Explanation

Correct Answer

C:

A generative language model estimates likely continuations rather than reading a stored answer paragraph. The expected acceptance evidence is: The output is produced as a sequence of tokens conditioned on the prompt. This observation directly tests the stated behavior.

Incorrect Answers

A:

Tokenization splits text into subword or other units for model processing. This observation does not establish the required result: The output is produced as a sequence of tokens conditioned on the prompt.

B:

Reproducibility depends on more than the visible prompt text. This observation does not establish the required result: The output is produced as a sequence of tokens conditioned on the prompt.

D:

Generation stops when an output budget or stop condition is reached. This observation does not establish the required result: The output is produced as a sequence of tokens conditioned on the prompt.

Question 495

Troubleshooting report: A blurry graph yields fabricated precise percentages. Which correction is MOST directly targeted to this failure?

  1. Use the documented feature list for the deployed API version
  2. Request a clearer image or report uncertainty
  3. Use image input and OCR or vision grounding
  4. Use image captioning with verified visual input

Explanation

Correct Answer

B:

This correction directly addresses the reported failure. Visual reasoning is not proof of an exact value when axes are unclear. The targeted change is: Request a clearer image or report uncertainty.

Incorrect Answers

A:

Feature names and availability vary across versions. In the stated case, troubleshooting report: A blurry graph yields fabricated precise percentages; “Use the documented feature list for the deployed API version” does not target that failure.

C:

Visual evidence is needed to read sign content. In the stated case, troubleshooting report: A blurry graph yields fabricated precise percentages; “Use image input and OCR or vision grounding” does not target that failure.

D:

Captioning produces a sentence-level description rather than a coordinate list. In the stated case, troubleshooting report: A blurry graph yields fabricated precise percentages; “Use image captioning with verified visual input” does not target that failure.

Question 496

A project needs the following AI behavior: A classifier performs poorly on specialized industry terminology despite strong retrieval. What should the designer implement?

  1. Use a supervised classification model with representative labeled cases
  2. Evaluate domain adaptation or fine-tuning only after baseline tests show the gap
  3. Review precision, recall and the confusion matrix rather than accuracy alone
  4. Measure expected tokens and requests per minute, then choose an approved throughput tier or capacity adjustment before launch

Explanation

Correct Answer

B:

Model specialization may help task behavior but requires measured benefit. This implementation addresses the precise goal rather than an adjacent AI task.

Incorrect Answers

A:

Classification predicts discrete labels rather than a continuous numeric target. That is not sufficient here: Model specialization may help task behavior but requires measured benefit.

C:

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

D:

Quota sizing is a deployment-capacity decision; retrying a predictable sustained excess cannot add model throughput. That is not sufficient here: Model specialization may help task behavior but requires measured benefit.

Question 497

Troubleshooting report: The tool definitions have ambiguous identical descriptions. Which correction is MOST directly targeted to this failure?

  1. Keep publication behind a separate approved deployment step
  2. Poll or await the documented completion status before reporting success
  3. Separate tool names, arguments and method scopes
  4. Respect retry guidance and cap concurrent calls

Explanation

Correct Answer

C:

This correction directly addresses the reported failure. Clear tool contracts reduce confusion between read and write operations. The targeted change is: Separate tool names, arguments and method scopes.

Incorrect Answers

A:

A draft or tested agent is not automatically a publicly released service. In the stated case, troubleshooting report: The tool definitions have ambiguous identical descriptions; “Keep publication behind a separate approved deployment step” does not target that failure.

B:

Submission is not the same as successful completion. In the stated case, troubleshooting report: The tool definitions have ambiguous identical descriptions; “Poll or await the documented completion status before reporting success” does not target that failure.

D:

Rate limits protect the service and uncontrolled retries amplify overload. In the stated case, troubleshooting report: The tool definitions have ambiguous identical descriptions; “Respect retry guidance and cap concurrent calls” does not target that failure.

Question 498

An AI operations team is testing two distinct capabilities. Failure (1): The team attempts inference before deploying or selecting a model. Failure (2): The code uses only a model marketing name with no deployment routing. Which TWO specific corrections should the team apply?

  1. Reject claims with no matching source evidence
  2. Tighten the output contract and apply post-response validation
  3. Version prompt and deployment settings and retain the actual evaluation output
  4. Create a supported deployment and confirm resource access
  5. Set the correct endpoint and deployed model name in the client

Explanation

Correct Answers

D:

Correct for failure (1). The portal provides model selection, deployment and test interfaces for supported resources. The proposed correction directly addresses that reported defect.

E:

Correct for failure (2). An inferencing request is routed to a concrete deployment configured in the resource. It independently corrects the second defect.

Incorrect Answers

A:

Prompt instructions must be paired with evidence verification for consequential facts. The stated pair of failures instead requires “Create a supported deployment and confirm resource access” and “Set the correct endpoint and deployed model name in the client”; this option cannot substitute for both.

B:

Explicit examples can improve adherence to bounded output conventions. The stated pair of failures instead requires “Create a supported deployment and confirm resource access” and “Set the correct endpoint and deployed model name in the client”; this option cannot substitute for both.

C:

Recorded settings support investigation, but generative models can still vary on an apparently identical rerun. The stated pair of failures instead requires “Create a supported deployment and confirm resource access” and “Set the correct endpoint and deployed model name in the client”; this option cannot substitute for both.

Question 499

An AI operations team is testing two distinct capabilities. Failure (1): The team reports only class accuracy for numeric predictions. Failure (2): The team fabricates purchase labels to run a supervised classifier. Which TWO specific corrections should the team apply?

  1. Apply clustering to relevant normalized features
  2. Use early stopping and investigate excessive model capacity
  3. Re-embed the corpus or use a versioned compatible index
  4. Evaluate minority-class recall and threshold tradeoffs
  5. Evaluate mean absolute error or another appropriate regression metric

Explanation

Correct Answers

A:

Correct for failure (2). Unsupervised learning finds structure without supplied target labels. It independently corrects the second defect.

E:

Correct for failure (1). Regression predicts continuous quantities under a defined error metric. The proposed correction directly addresses that reported defect.

Incorrect Answers

B:

More optimization on training data can worsen generalization. It does not complete the two decisions required in this question.

C:

Incompatible vector spaces can make nearest-neighbor rankings meaningless. It does not complete the two decisions required in this question.

D:

High accuracy can conceal failure to detect rare positive events. The stated pair of failures instead requires “Apply clustering to relevant normalized features” and “Evaluate mean absolute error or another appropriate regression metric”; this option cannot substitute for both.

Question 500

A solution team has selected this control: Use a video-capable Content Understanding analyzer with timestamps and supported fields. Which observation is the most direct acceptance evidence?

  1. The answer resolves to a specific approved source document version
  2. The extracted steps map to the correct clips and source times
  3. Quoted text matches the time-aligned source transcript
  4. Observed actions link to supported source intervals while unobserved short events are not claimed as detected

Explanation

Correct Answer

B:

Multimodal extraction draws on visual and spoken evidence over time. The expected acceptance evidence is: The extracted steps map to the correct clips and source times. 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 extracted steps map to the correct clips and source times.

C:

Generated summaries are paraphrases and not guaranteed verbatim quotes. This observation does not establish the required result: The extracted steps map to the correct clips and source times.

D:

Content Understanding samples approximately one frame per second, so brief or single-frame actions can be missed despite valid timestamps. This observation does not establish the required result: The extracted steps map to the correct clips and source times.

Microsoft Azure AI Fundamentals (AI-901) Practice Tests