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 79

Question 780

An AI operations team is testing two distinct capabilities. Failure (1): A single image classification label cannot identify defect positions. Failure (2): The team calls speech-to-text on already written text. Which TWO specific corrections should the team apply?

  1. Set the API request to the correct deployment name and endpoint
  2. Test a higher diversity setting and compare evaluated outputs
  3. Invoke text-read or document OCR capability
  4. Switch to object detection with coordinate outputs
  5. Use a speech synthesis endpoint and compatible output format

Explanation

Correct Answers

D:

Correct for failure (1). An appropriate spatial detection model must localize the defect classes, not merely assign an overall picture label. The proposed correction directly addresses that reported defect.

E:

Correct for failure (2). Speech synthesis generates audio from text rather than transcribing audio. It independently corrects the second defect.

Incorrect Answers

A:

Applications call a deployed model resource rather than a generic marketing model name. The stated pair of failures instead requires “Switch to object detection with coordinate outputs” and “Use a speech synthesis endpoint and compatible output format”; this option cannot substitute for both.

B:

More varied token sampling can increase creative alternatives. It does not complete the two decisions required in this question.

C:

This action would address another problem: the app returns only visual tags such as paper and desk. OCR produces readable text from pixels. It does not resolve either specified failure.

Question 781

An Azure AI Fundamentals solution has this requirement: A law firm wants to find clauses in hundreds of contracts. Which implementation most directly addresses it?

  1. Use text embeddings with nearest-neighbor search and business filters
  2. Use sentiment analysis to estimate positive, negative, or neutral tone
  3. Use document information extraction with a typed field schema
  4. Use semantic document indexing and retrieval rather than only keyword equality

Explanation

Correct Answer

D:

Semantic search can retrieve conceptually related passages. This implementation addresses the precise goal rather than an adjacent AI task.

Incorrect Answers

A:

Semantic vectors capture related concepts while filters enforce eligibility. That is not sufficient here: Semantic search can retrieve conceptually related passages.

B:

Sentiment analysis infers expressed attitude rather than extracting names. That is not sufficient here: Semantic search can retrieve conceptually related passages.

C:

Document extraction maps unstructured layouts into structured values. That is not sufficient here: Semantic search can retrieve conceptually related passages.

Question 782

Troubleshooting report: The team tests only whether the chat interface renders text. Which correction is MOST directly targeted to this failure?

  1. Exercise negative and expected tool paths before release
  2. Require a completed tool result and error-path response
  3. Isolate request state and bind responses to their originating session
  4. Add approval state and verified role checks

Explanation

Correct Answer

A:

This correction directly addresses the reported failure. An answer that sounds plausible may conceal failed tool calls. The targeted change is: Exercise negative and expected tool paths before release.

Incorrect Answers

B:

Timeout handling should not be reported as a successful action. In the stated case, troubleshooting report: The team tests only whether the chat interface renders text; “Require a completed tool result and error-path response” does not target that failure.

C:

Concurrent inference must not mix user data or response ownership. In the stated case, troubleshooting report: The team tests only whether the chat interface renders text; “Isolate request state and bind responses to their originating session” does not target that failure.

D:

High-impact side effects need permissions independent of conversational language. In the stated case, troubleshooting report: The team tests only whether the chat interface renders text; “Add approval state and verified role checks” does not target that failure.

Question 783

An acceptance tester needs to validate: Use evaluated confidence thresholds aligned to the business error cost. Which observed result proves the intended behavior?

  1. The thumbnail preserves subject proportions and important content
  2. An ambiguous frame triggers review rather than an irreversible action
  3. A labeled test shows the chosen threshold achieves target precision/recall
  4. A drawn overlay aligns each box with the corresponding person

Explanation

Correct Answer

C:

Confidence scores require calibration and contextual acceptance policy. The expected acceptance evidence is: A labeled test shows the chosen threshold achieves target precision/recall. This observation directly tests the stated behavior.

Incorrect Answers

A:

Smart cropping differs from simply stretching pixels. This observation does not establish the required result: A labeled test shows the chosen threshold achieves target precision/recall.

B:

Generated visual interpretations can be wrong even when phrased fluently. This observation does not establish the required result: A labeled test shows the chosen threshold achieves target precision/recall.

D:

Localization metadata is tied to pixel regions, not just labels. This observation does not establish the required result: A labeled test shows the chosen threshold achieves target precision/recall.

Question 784

An AI operations team is testing two distinct capabilities. Failure (1): The pipeline only summarizes the document in prose. Failure (2): Every upload is forced through the invoice schema. Which TWO specific corrections should the team apply?

  1. Evaluate noise conditions and request clarification at low confidence
  2. Introduce a document classifier and route by type
  3. Define extraction fields and validate values
  4. Use document layout extraction with typed checkbox outputs
  5. Add retrieval and grounded citation generation

Explanation

Correct Answers

B:

Correct for failure (2). Classification selects the extraction process matching the document category. It independently corrects the second defect.

C:

Correct for failure (1). Document extraction maps unstructured layouts into structured values. The proposed correction directly addresses that reported defect.

Incorrect Answers

A:

Speech input quality affects subsequent language classification. It does not complete the two decisions required in this question.

D:

Different form elements require more than ordinary running-text OCR. The stated pair of failures instead requires “Introduce a document classifier and route by type” and “Define extraction fields and validate values”; this option cannot substitute for both.

E:

RAG retrieves current external evidence for a generated response. The stated pair of failures instead requires “Introduce a document classifier and route by type” and “Define extraction fields and validate values”; this option cannot substitute for both.

Question 785

A solution team has selected this control: Use a compatible code-capable generative model with constrained context and tests. Which observation is the most direct acceptance evidence?

  1. An unconfirmed action remains pending despite a persuasive request
  2. The answer mentions relevant visual details unique to the uploaded image
  3. A proposed snippet passes syntax and behavioral tests before merge
  4. A follow-up refers to the correct permitted order from the same session

Explanation

Correct Answer

C:

Generated code must be validated, not treated as production-ready from fluency. The expected acceptance evidence is: A proposed snippet passes syntax and behavioral tests before merge. 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 proposed snippet passes syntax and behavioral tests before merge.

B:

The model must receive the actual visual content to reason about it. This observation does not establish the required result: A proposed snippet passes syntax and behavioral tests before merge.

D:

Conversation continuity should preserve identifiers without crossing user sessions. This observation does not establish the required result: A proposed snippet passes syntax and behavioral tests before merge.

Question 786

Troubleshooting report: The code uses only a model marketing name with no deployment routing. Which correction is MOST directly targeted to this failure?

  1. Create a supported deployment and confirm resource access
  2. Require explicit identifiers and reject invented IDs
  3. Attach idempotency keys and verify tool receipts after reconnecting
  4. Set the correct endpoint and deployed model name in the client

Explanation

Correct Answer

D:

This correction directly addresses the reported failure. An inferencing request is routed to a concrete deployment configured in the resource. The targeted change is: Set the correct endpoint and deployed model name in the client.

Incorrect Answers

A:

The portal provides model selection, deployment and test interfaces for supported resources. In the stated case, troubleshooting report: The code uses only a model marketing name with no deployment routing; “Create a supported deployment and confirm resource access” does not target that failure.

B:

Prompts can define required preconditions but tools must still enforce them. In the stated case, troubleshooting report: The code uses only a model marketing name with no deployment routing; “Require explicit identifiers and reject invented IDs” does not target that failure.

C:

Client retries should distinguish read requests from side-effecting tool calls. In the stated case, troubleshooting report: The code uses only a model marketing name with no deployment routing; “Attach idempotency keys and verify tool receipts after reconnecting” does not target that failure.

Question 787

Troubleshooting report: The model returns only an image-wide severity label. Which correction is MOST directly targeted to this failure?

  1. Preserve speaker identities and confirm relevant transcript spans
  2. Evaluate noise conditions and request clarification at low confidence
  3. Invoke image segmentation and compositing or a controlled image editor
  4. Choose segmentation with validated pixel-level damage regions

Explanation

Correct Answer

D:

This correction directly addresses the reported failure. Pixel-level region masks localize building damage more precisely than an overall image class label. The targeted change is: Choose segmentation with validated pixel-level damage regions.

Incorrect Answers

A:

Attribution distinguishes who committed to each action. In the stated case, troubleshooting report: The model returns only an image-wide severity label; “Preserve speaker identities and confirm relevant transcript spans” does not target that failure.

B:

Speech input quality affects subsequent language classification. In the stated case, troubleshooting report: The model returns only an image-wide severity label; “Evaluate noise conditions and request clarification at low confidence” does not target that failure.

C:

Editing changes selected pixels while leaving product regions intact. In the stated case, troubleshooting report: The model returns only an image-wide severity label; “Invoke image segmentation and compositing or a controlled image editor” does not target that failure.

Question 788

Troubleshooting report: The app merges handwritten and printed fields with no context. Which correction is MOST directly targeted to this failure?

  1. Verify the actual source signature region or use explicitly approved preview detection, never infer signing from a typed name
  2. Enable supported layout and field extraction features
  3. Enable supported document confidence and evaluate thresholds against labeled results and business rules
  4. Retain region metadata and use field review where needed

Explanation

Correct Answer

D:

This correction directly addresses the reported failure. Source-location evidence helps verify recognition on mixed-layout documents. The targeted change is: Retain region metadata and use field review where needed.

Incorrect Answers

A:

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. In the stated case, troubleshooting report: The app merges handwritten and printed fields with no context; “Verify the actual source signature region or use explicitly approved preview detection, never infer signing from a typed name” does not target that failure.

B:

Ordinary plain-text OCR may lose table relationships and checkbox states. In the stated case, troubleshooting report: The app merges handwritten and printed fields with no context; “Enable supported layout and field extraction features” does not target that failure.

C:

Content Understanding confidence applies to supported document analyzers; confidence estimates alone do not prove field correctness. In the stated case, troubleshooting report: The app merges handwritten and printed fields with no context; “Enable supported document confidence and evaluate thresholds against labeled results and business rules” does not target that failure.

Question 789

An engineering review documents this requirement: A chatbot uses a retrieved web page as if it could change the app's instructions. Select the most appropriate solution.

  1. Preserve the higher-priority application constraints in the message hierarchy
  2. Use representative test prompts and inspect execution and tool traces
  3. Keep retrieved documents in lower-trust context and filter tool requests independently
  4. Specify the output contract and validate returned fields against the schema

Explanation

Correct Answer

C:

External content cannot replace privileged system behavior. This implementation addresses the precise goal rather than an adjacent AI task.

Incorrect Answers

A:

Developer and system constraints govern allowed operations over user requests. That is not sufficient here: External content cannot replace privileged system behavior.

B:

An answer that sounds plausible may conceal failed tool calls. That is not sufficient here: External content cannot replace privileged system behavior.

D:

Model output constraints help downstream parsing but still require validation. That is not sufficient here: External content cannot replace privileged system behavior.

Question 790

An Azure-based AI team must satisfy this objective: An application must route generation to the correct model deployment after a model upgrade. Which design choice is best?

  1. Send the image in a supported multimodal message along with the question
  2. Archive the actual evaluated response together with the prompt, model deployment, input and generation settings
  3. Write an instruction to identify missing inputs and request clarification before tools
  4. Update the configured deployment name and verify compatibility with the SDK

Explanation

Correct Answer

D:

A model family label is not necessarily the concrete deployed endpoint. This implementation addresses the precise goal rather than an adjacent AI task.

Incorrect Answers

A:

The model must receive the actual visual content to reason about it. That is not sufficient here: A model family label is not necessarily the concrete deployed endpoint.

B:

Recorded settings support investigation, but generative models can still vary on an apparently identical rerun. That is not sufficient here: A model family label is not necessarily the concrete deployed endpoint.

C:

Prompts can define required preconditions but tools must still enforce them. That is not sufficient here: A model family label is not necessarily the concrete deployed endpoint.

Microsoft Azure AI Fundamentals (AI-901) Practice Tests