cert
cert-1
cert-2

Pass Microsoft AI-103 Exam in First Attempt Guaranteed!

Get 100% Latest Exam Questions, Accurate & Verified Answers to Pass the Actual Exam!
30 Days Free Updates, Instant Download!

cert-5
cert-6
AI-103 Exam - Verified By Experts
AI-103 Premium File

AI-103 Premium File

$89.99
$98.99
  • Premium File 187 Questions & Answers. Last Update: Nov 15, 2026

Whats Included:

  • Latest Questions
  • 100% Accurate Answers
  • Fast Exam Updates
 
$98.99
$89.99
accept 397 downloads in the last 7 days
block-screenshots
Microsoft AI-103 Practice Test 46

Question 450

An agent retrieves a restricted HR policy because the index is shared across teams. What must be enforced?

  1. Use a shared search index as permission to display HR passages to every department
  2. Apply source ACL and tenant authorization at retrieval time
  3. Remove or invalidate corresponding vectors and cached excerpts during deletion handling
  4. Store source version identifiers with retrieved chunks and citations

Explanation

Correct Answer

B:

A shared vector index must not bypass permissions present on the authoritative documents. This directly supports the need to prevent restricted documents crossing retrieval boundaries, unlike a mechanism primarily intended to attach traceable evidence to generated answers.

Incorrect Answers

A:

Co-location in an index is not an authorization grant. Retrieval must enforce the source's ACL and tenant scope.

C:

Derived retrieval artifacts must follow source retention and deletion requirements. Here the needed outcome is to prevent restricted documents crossing retrieval boundaries; an option designed to remove expired source-derived vectors and snippets would not directly provide it.

D:

Document version lineage enables auditing which exact evidence informed the response. Its primary purpose is to identify referenced document revisions; the stated constraint instead demands a way to prevent restricted documents crossing retrieval boundaries.

Question 451

A fixed-format response service needs consistent JSON fields, low stylistic variation and protection from instructions inside retrieved content. Which THREE techniques address those needs? Choose THREE.

  1. Separate trusted instructions from retrieved untrusted evidence and forbid instruction promotion
  2. Increase generation randomness and accept any syntactically valid JSON, even when fields or values vary across runs.
  3. Specify task priority and an ordered output schema for separate subrequests
  4. Use lower sampling variability with a strict response template
  5. Use supported schema-constrained output and deterministic post-parse validation
  6. Evaluate diverse examples and hard negatives before locking the prompt template

Explanation

Correct Answers

A:

Prompt structure should not let a third-party document override trusted application rules. Lower variability, trustworthy prompt separation and schema validation jointly constrain output and authority.

D:

Constrained generation and a defined schema reduce style variation for deterministic business tasks.

E:

Required keys and types should be enforced with a machine-readable contract, not example text alone.

Incorrect Answers

B:

A valid JSON document can still violate the fixed business schema. Constrain field names and permitted values, reduce unnecessary variation and run deterministic validation.

C:

Clear task decomposition reduces ambiguity when one input contains multiple decisions. That mechanism concerns organize multi-intent generation deterministically, whereas this case tests how to stabilize generation, protect prompt boundaries and enforce structured output.

F:

Prompt examples can bias the model toward narrow patterns, so edge-case coverage matters. Here, however, the stated objective is to stabilize generation, protect prompt boundaries and enforce structured output; this option focuses on validate prompt behavior on diverse difficult cases.

Question 452

An autonomous assistant can request invoice payments but has no business authorization to execute them alone. What control is required?

  1. Require a named human approval before side-effecting payment tools may run.
  2. Let the LLM approve the payment when its confidence exceeds 90 percent.
  3. Automatically execute any payment under a user-specified amount without verifying authorization.
  4. Ask the assistant to write manager approved into its own audit log.

Explanation

Correct Answer

A:

A payment action changes external state. Human-in-the-loop approval should be tied to the specific amount, payee and audit record.

Incorrect Answers

B:

Model confidence does not constitute authorized human approval for financial side effects.

C:

A threshold alone cannot replace the workflow's required named approver and transaction approval.

D:

Self-generated log text cannot demonstrate a real human authorized the amount and payee.

Question 453

A service uses unverified online image references that could include malicious or private content. What ingestion boundary is needed?

  1. Screen incoming media with appropriate visual safety classification
  2. Validate input sources, permissions and content safety before multimodal processing
  3. Pass text found inside remote images directly to the agent as trusted instructions after an OCR scan.
  4. Calibrate visual safety thresholds on reviewed allowed and prohibited examples

Explanation

Correct Answer

B:

Third-party media should be assessed for authorization and harmful payloads before agent use. The practical requirement is to screen untrusted external media sources; a solution built to moderate disallowed input imagery would not meet that same need.

Incorrect Answers

A:

Input moderation reduces exposure to disallowed visual content before later processing. That is relevant to moderate disallowed input imagery, but the present scenario requires an ability to screen untrusted external media sources.

C:

A successful OCR extraction does not make online content trustworthy. Validate allowed sources and provenance, scan for unsafe content and isolate retrieved text from agent instructions and privileges.

D:

Moderation requires measuring both overblocking and missed harmful content. That capability is appropriate for reduce visual safety false positives, not this separate task of finding how to screen untrusted external media sources.

Question 454

A speech translation workflow omits the exact invoice amount in one target language. What comparison should the quality gate perform?

  1. Judge translation quality solely from fluent spoken pronunciation.
  2. Check translated speech transcription against source numbers, currencies and named entities.
  3. Compare only the number of spoken words between languages.
  4. Confirm that both languages mention an invoice without checking the amount.

Explanation

Correct Answer

B:

Multilingual audio quality is not only naturalness; critical financial facts must remain intact.

Incorrect Answers

A:

Natural-sounding audio may still omit an invoice amount or change its currency.

C:

Two spoken translations can have similar word counts while one omits an invoice amount. Compare the recognized numeric value, currency and units against the source transcript and use targeted review when they diverge.

D:

Topic overlap is not enough when the exact monetary value must remain intact.

Question 455

An agent answers a legal question with a citation to a document's title but no supporting sentence. What provenance should be added?

  1. Link to the legal document's cover page as proof of every generated claim.
  2. Generate a likely paragraph number without comparing it to extracted source text.
  3. Include the exact supporting page or chunk span and confirm it entails the generated claim.
  4. Use the model's confidence score to replace source excerpts.

Explanation

Correct Answer

C:

A document title is not enough to verify a particular assertion; passage-level citations support review.

Incorrect Answers

A:

A cover page identifies a document but rarely substantiates a specific legal assertion.

B:

A plausible invented span is not a verified citation to the underlying legal passage.

D:

Confidence is not provenance and cannot demonstrate textual support for a legal answer.

Question 456

A model generates a solution and then rates its own answer as perfect without checking evidence. What additional validation is needed?

  1. Cap critique-and-revision passes using quality improvement thresholds
  2. Use independent source-grounded checks rather than relying solely on self-confidence
  3. Accept the model’s self-reported perfect score without any labeled comparison or source check.
  4. Re-run a held-out evaluation set with measured before-and-after outcomes

Explanation

Correct Answer

B:

Self-critique without external evidence can reinforce the same mistaken claim. The decisive difference is whether the design can verify model assertions against independent sources; the competing function serves to stop unproductive self-critique loops.

Incorrect Answers

A:

Bounded reflection prevents token waste when repeated attempts no longer improve output. That capability is appropriate for stop unproductive self-critique loops, not this separate task of finding how to verify model assertions against independent sources.

C:

The model’s confidence is not independent evidence of correctness. Evaluate the answer against external references or validated tests rather than trusting self-grading.

D:

Independent examples reveal whether revision improves correctness beyond cosmetic changes. That is relevant to evaluate objective improvement after a critique pass, but the present scenario requires an ability to verify model assertions against independent sources.

Question 457

A search-grounded assistant cites correct documents but omits essential facts asked by users. What assessment is missing?

  1. Use groundedness or faithfulness evaluation against the source document
  2. Measure answer completeness and task relevance against expected criteria
  3. Archive evaluation-dataset revision IDs while scoring only whether each cited document link is reachable, never whether the answer covers required facts.
  4. Count the number of citations without checking whether the answer covered every requested fact.

Explanation

Correct Answer

B:

Citation correctness is not sufficient if the response does not fulfill the information need. In this scenario the team must evaluate whether answers actually fulfill the task. A feature used to score source-supported generated statements does not substitute for it.

Incorrect Answers

A:

A fluent summary can still invent claims; groundedness tests whether assertions are supported by the reference. Here the needed outcome is to evaluate whether answers actually fulfill the task; an option designed to score source-supported generated statements would not directly provide it.

C:

Versioning benchmarks aids repeatability but does not substitute for an answer completeness metric. Evaluate required facts against the expected answer, including omissions despite accurate citations.

D:

Correct document links may coexist with missing required information. Score answer completeness against a labeled requirement checklist, not citation presence alone.

Question 458

A generative image system depicts a hazardous medical procedure convincingly but may be inaccurate. What review gate is prudent?

  1. Apply brand-usage policy checks and require approved assets before publication
  2. Require subject-matter review for high-impact synthetic instructional visuals
  3. Apply clear provenance labeling or watermarking where appropriate
  4. Block OCR-based tool calls but publish the generated medical procedure imagery without clinical validation.

Explanation

Correct Answer

B:

Photorealism is not evidence of correctness; risky instructional media needs expert validation. Here the relevant capability must let the team prevent unsafe high-stakes synthetic guidance. An alternative targeting enforce approved trademark usage in generated media solves another problem.

Incorrect Answers

A:

Visual generation does not override intellectual-property or brand governance requirements. This helps when a team needs to enforce approved trademark usage in generated media; it does not establish how to prevent unsafe high-stakes synthetic guidance.

C:

Synthetic-media disclosure reduces confusion over whether visuals show real products or events. That capability is appropriate for disclose generated-media provenance, not this separate task of finding how to prevent unsafe high-stakes synthetic guidance.

D:

OCR instruction controls do not establish medical accuracy. Use subject-matter and safety review before releasing persuasive medical-procedure imagery that may depict unsafe steps.

Question 459

An agent gives spoken instructions before a user finishes dictating their request. What interaction control is appropriate?

  1. Speak the final answer as soon as the first partial audio token arrives
  2. Use endpoint detection and turn-taking signals before final response synthesis
  3. Ask for explicit confirmation of the recognized command before side effects
  4. Normalize modality-specific evidence into tagged multimodal conversation turns

Explanation

Correct Answer

B:

Voice interaction needs explicit speech-turn boundaries to avoid interrupting or mishearing requests. This resolves the specific need to detect completion of a spoken user turn, while a measure that helps handle mid-session language changes safely addresses a different requirement.

Incorrect Answers

A:

A partial utterance may not represent the completed request. Endpoint and turn-taking signals should determine when the assistant may respond.

C:

Voice recognition errors can be costly, so important tool actions require verified user intent. This helps when a team needs to confirm uncertain spoken commands before actions; it does not establish how to detect completion of a spoken user turn.

D:

Tagged input blocks preserve source and modality while maintaining ordered task context. The function serves to integrate different input modalities in one conversation, whereas the actual requirement is to detect completion of a spoken user turn.

Question 460

A document is deleted for a legal retention request but its embedding remains in the index. What action is needed?

  1. Remove the source document record but keep its embeddings and cached passages active
  2. Return source-grounded citations that identify exact evidence spans
  3. Remove or invalidate corresponding vectors and cached excerpts during deletion handling
  4. Apply source ACL and tenant authorization at retrieval time

Explanation

Correct Answer

C:

Derived retrieval artifacts must follow source retention and deletion requirements. The decisive difference is whether the design can remove expired source-derived vectors and snippets; the competing function serves to identify referenced document revisions.

Incorrect Answers

A:

A deletion workflow must propagate to derived vectors and caches, otherwise removed content may continue to appear in retrieval.

B:

Verifiable passage references let readers assess whether an answer follows the authorized documents. Here the needed outcome is to remove expired source-derived vectors and snippets; an option designed to attach traceable evidence to generated answers would not directly provide it.

D:

A shared vector index must not bypass permissions present on the authoritative documents. That is relevant to prevent restricted documents crossing retrieval boundaries, but the present scenario requires an ability to remove expired source-derived vectors and snippets.

Developing AI Apps and Agents on Azure (AI-103) Practice Tests