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 90

Question 890

An acceptance tester needs to validate: Use speech-to-text transcription before indexing call content. Which observed result proves the intended behavior?

  1. Line-item totals and row relationships reconcile to the document
  2. Names in a labeled sample are detected and redacted
  3. Suggested products are relevant and meet inventory/authorization rules
  4. The transcript contains the actual spoken issue and timestamp

Explanation

Correct Answer

D:

Speech recognition converts speech into words suitable for search. The expected acceptance evidence is: The transcript contains the actual spoken issue and timestamp. This observation directly tests the stated behavior.

Incorrect Answers

A:

Table extraction associates quantities, descriptions and prices by row. This observation does not establish the required result: The transcript contains the actual spoken issue and timestamp.

B:

Entity extraction locates sensitive spans for privacy handling. This observation does not establish the required result: The transcript contains the actual spoken issue and timestamp.

C:

Semantic vectors capture related concepts while filters enforce eligibility. This observation does not establish the required result: The transcript contains the actual spoken issue and timestamp.

Question 891

An engineering review documents this requirement: A team changes the system prompt and later cannot reproduce a previously good answer. Select the most appropriate solution.

  1. Run adversarial model evaluations and verify that configured content-safety guardrails enforce the stated high-severity block criteria
  2. Choose a model size using representative quality and cost benchmarks
  3. Archive the actual evaluated response together with the prompt, model deployment, input and generation settings
  4. Treat user content as untrusted and maintain privileged instruction boundaries

Explanation

Correct Answer

C:

Recorded settings support investigation, but generative models can still vary on an apparently identical rerun.

Incorrect Answers

A:

Adversarial testing must measure actual configured policy enforcement rather than treating a prompt instruction as a guaranteed safety boundary. That is not sufficient here: Recorded settings support investigation, but generative models can still vary on an apparently identical rerun.

B:

Lower-cost deployment is acceptable only if accuracy and safety floors remain met. That is not sufficient here: Recorded settings support investigation, but generative models can still vary on an apparently identical rerun.

D:

A user cannot use text to elevate permissions beyond the application policy. That is not sufficient here: Recorded settings support investigation, but generative models can still vary on an apparently identical rerun.

Question 892

An AI operations team is testing two distinct capabilities. Failure (1): Only the English transcript is returned with no translation. Failure (2): Plain text without timestamps is fed to a caption player. Which TWO specific corrections should the team apply?

  1. Use a fixed annotated evaluation set and consistent metrics
  2. Enable supported diarization and validate turn assignments
  3. Request time-aligned recognition details
  4. XML-escape text values before synthesis
  5. Add translation and French synthesis after recognition

Explanation

Correct Answers

C:

Correct for failure (2). Timecodes align text segments to audio playback. It independently corrects the second defect.

E:

Correct for failure (1). Cross-language audio output requires both language transfer and synthesis. The proposed correction directly addresses that reported defect.

Incorrect Answers

A:

Comparable labeled audio prevents misleading quality claims. It does not complete the two decisions required in this question.

B:

Speaker diarization separates turns and identities at a conversation level. The stated pair of failures instead requires “Request time-aligned recognition details” and “Add translation and French synthesis after recognition”; this option cannot substitute for both.

D:

SSML is XML and malformed markup breaks the synthesis request. The stated pair of failures instead requires “Request time-aligned recognition details” and “Add translation and French synthesis after recognition”; this option cannot substitute for both.

Question 893

The approved design action is: Combine speech recognition, language translation, and appropriate speech synthesis. Which test result would BEST establish that it worked?

  1. The timeline marks arrival frames with appropriate timestamps
  2. A consented test delivers equivalent meaning in the listener's language
  3. Held-out events produce measured recall and precision for failures
  4. The product shape remains unchanged against the new background

Explanation

Correct Answer

B:

Speech translation spans audio input, semantic transfer and optional output audio. The expected acceptance evidence is: A consented test delivers equivalent meaning in the listener's language. This observation directly tests the stated behavior.

Incorrect Answers

A:

Video understanding retains order and time, unlike single-image tagging. This observation does not establish the required result: A consented test delivers equivalent meaning in the listener's language.

C:

A failure probability or class is a discrete decision under a defined horizon. This observation does not establish the required result: A consented test delivers equivalent meaning in the listener's language.

D:

Editing changes selected pixels while leaving product regions intact. This observation does not establish the required result: A consented test delivers equivalent meaning in the listener's language.

Question 894

An engineering review documents this requirement: An agent proposes sending customer data to an external tool. Select the most appropriate solution.

  1. Enforce data classification and authorization before tool invocation
  2. Use representative test prompts and inspect execution and tool traces
  3. Require explicit confirmation and backend authorization before tool execution
  4. Write an instruction to identify missing inputs and request clarification before tools

Explanation

Correct Answer

A:

Tool descriptions cannot override data-loss-prevention boundaries. This implementation addresses the precise goal rather than an adjacent AI task.

Incorrect Answers

B:

An answer that sounds plausible may conceal failed tool calls. That is not sufficient here: Tool descriptions cannot override data-loss-prevention boundaries.

C:

High-impact side effects need permissions independent of conversational language. That is not sufficient here: Tool descriptions cannot override data-loss-prevention boundaries.

D:

Prompts can define required preconditions but tools must still enforce them. That is not sufficient here: Tool descriptions cannot override data-loss-prevention boundaries.

Question 895

An acceptance tester needs to validate: Use region-aware captioning or a cropped image focused on that location. Which observed result proves the intended behavior?

  1. The same object is recognized consistently across supported rotations
  2. The returned description matches the selected upper-left object
  3. Holdout images correctly separate the two defect types
  4. The output is a validated outdoor probability or category

Explanation

Correct Answer

B:

Specific regions can carry details missed by a whole-image caption. The expected acceptance evidence is: The returned description matches the selected upper-left object. This observation directly tests the stated behavior.

Incorrect Answers

A:

Rotated inputs can degrade detection or coordinate interpretation. This observation does not establish the required result: The returned description matches the selected upper-left object.

C:

Fine-grained visual classes require representative discriminating images. This observation does not establish the required result: The returned description matches the selected upper-left object.

D:

Image-level classification is distinct from object localization. This observation does not establish the required result: The returned description matches the selected upper-left object.

Question 896

Troubleshooting report: The model labels the entire image as a store without locating items. Which correction is MOST directly targeted to this failure?

  1. Use a detection output with bounding boxes
  2. Apply entity extraction and validate spans and categories
  3. Define extraction fields and validate values
  4. Use entity recognition and verify sensitive-entity recall

Explanation

Correct Answer

A:

This correction directly addresses the reported failure. Object detectors return classes and spatial boundaries. The targeted change is: Use a detection output with bounding boxes.

Incorrect Answers

B:

NER locates and categorizes named entities in text. In the stated case, troubleshooting report: The model labels the entire image as a store without locating items; “Apply entity extraction and validate spans and categories” does not target that failure.

C:

Document extraction maps unstructured layouts into structured values. In the stated case, troubleshooting report: The model labels the entire image as a store without locating items; “Define extraction fields and validate values” does not target that failure.

D:

Entity extraction locates sensitive spans for privacy handling. In the stated case, troubleshooting report: The model labels the entire image as a store without locating items; “Use entity recognition and verify sensitive-entity recall” does not target that failure.

Question 897

A project needs the following AI behavior: A service wants short overviews of lengthy support tickets. What should the designer implement?

  1. Segment authorized audio by task and retention purpose before processing
  2. Run grounded summarization after speech recognition and preserve source segments
  3. Use text summarization with clear length and fidelity constraints
  4. Check the deployment supports the requested audio-input modality and request structure

Explanation

Correct Answer

C:

Summarization compresses written information rather than changing its facts. This implementation addresses the precise goal rather than an adjacent AI task.

Incorrect Answers

A:

Minimization reduces unnecessary personal-data exposure. That is not sufficient here: Summarization compresses written information rather than changing its facts.

B:

Transcription obtains words; summarization extracts decisions and tasks. That is not sufficient here: Summarization compresses written information rather than changing its facts.

D:

Model capabilities and accepted payload types differ across deployments. That is not sufficient here: Summarization compresses written information rather than changing its facts.

Question 898

An Azure AI Fundamentals solution has this requirement: An automated report must return fields with exact names and types. Which implementation most directly addresses it?

  1. Use the same compatible embedding version and dimensions for stored items and queries
  2. Use structured outputs or strict application-side schema validation
  3. Use a carefully constrained prompt and a lower sampling temperature where supported
  4. Use optical character recognition instead of generic image tagging

Explanation

Correct Answer

B:

Natural-language fluency is not a guarantee of machine-readable contract validity. This implementation addresses the precise goal rather than an adjacent AI task.

Incorrect Answers

A:

Incompatible vector spaces can make nearest-neighbor rankings meaningless. That is not sufficient here: Natural-language fluency is not a guarantee of machine-readable contract validity.

C:

Lower randomness often improves consistency but does not guarantee correctness. That is not sufficient here: Natural-language fluency is not a guarantee of machine-readable contract validity.

D:

OCR produces readable text from pixels. That is not sufficient here: Natural-language fluency is not a guarantee of machine-readable contract validity.

Question 899

An Azure AI solution uses the following approved control: Use conversational generative AI with grounded retrieval and scoped dialogue state. What observable result should pass QA?

  1. Follow-up questions resolve consistently to the same approved case documents
  2. The assistant produces a relevant answer with approved citations
  3. A labeled review set produces measured sentiment predictions
  4. Line-item totals and row relationships reconcile to the document

Explanation

Correct Answer

A:

The assistant should incorporate context while remaining connected to reliable sources. The expected acceptance evidence is: Follow-up questions resolve consistently to the same approved case documents. This observation directly tests the stated behavior.

Incorrect Answers

B:

Generation produces novel text, unlike choosing a fixed class label. This observation does not establish the required result: Follow-up questions resolve consistently to the same approved case documents.

C:

Sentiment analysis infers expressed attitude rather than extracting names. This observation does not establish the required result: Follow-up questions resolve consistently to the same approved case documents.

D:

Table extraction associates quantities, descriptions and prices by row. This observation does not establish the required result: Follow-up questions resolve consistently to the same approved case documents.

Question 900

Troubleshooting report: The QA sample includes only benign customer questions. Which correction is MOST directly targeted to this failure?

  1. Construct citations from stored retrieval IDs rather than generated titles
  2. Apply case and role filters before providing model context
  3. Prune irrelevant history and preserve the minimum required state
  4. Add adversarial cases and verify configured safety block and escalation behavior

Explanation

Correct Answer

D:

This correction directly addresses the reported failure. Adversarial testing must measure actual configured policy enforcement rather than treating a prompt instruction as a guaranteed safety boundary. The targeted change is: Add adversarial cases and verify configured safety block and escalation behavior.

Incorrect Answers

A:

Explicit response structure makes evidence easier to verify. In the stated case, troubleshooting report: The QA sample includes only benign customer questions; “Construct citations from stored retrieval IDs rather than generated titles” does not target that failure.

B:

Source scoping enforces privacy and also improves answer relevance. In the stated case, troubleshooting report: The QA sample includes only benign customer questions; “Apply case and role filters before providing model context” does not target that failure.

C:

Long irrelevant context consumes tokens and may increase latency. In the stated case, troubleshooting report: The QA sample includes only benign customer questions; “Prune irrelevant history and preserve the minimum required state” does not target that failure.

Microsoft Azure AI Fundamentals (AI-901) Practice Tests