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 43

Question 420

A claims pipeline must return a known set of fields from varied claim documents. What analyzer design fits?

  1. Return an unstructured paragraph for every claim, with arbitrary field names that downstream systems must guess.
  2. Configure a Content Understanding analyzer with task-specific structured output fields
  3. Choose a supported multimodal Content Understanding configuration appropriate to mixed-media inputs
  4. Validate extracted amounts against source evidence and business rules before downstream writes

Explanation

Correct Answer

B:

Defined extraction fields make downstream claim validation and automation dependable. The decisive difference is whether the design can produce consistent structured outputs from claim documents; the competing function serves to prepare structured readable document evidence.

Incorrect Answers

A:

A known downstream schema requires consistent typed fields. Configure the analyzer's structured output and verify required values rather than relying on unpredictable prose.

C:

Multimodal analyzers can combine evidence from different media when configured for the task. Its primary purpose is to analyze mixed-media documents and assets; the stated constraint instead demands a way to produce consistent structured outputs from claim documents.

D:

Extraction confidence does not replace checks against the actual source and numeric invariants. That is relevant to verify extracted values before financial side effects, but the present scenario requires an ability to produce consistent structured outputs from claim documents.

Question 421

A customer support system mixes system instructions with retrieved web content and obeys the web page's commands. What prompt assembly boundary is needed?

  1. Separate trusted instructions from retrieved untrusted evidence and forbid instruction promotion
  2. Increase temperature and randomness to resolve the assistant's tendency to follow instructions from untrusted web pages.
  3. Use supported schema-constrained output and deterministic post-parse validation
  4. Use lower sampling variability with a strict response template

Explanation

Correct Answer

A:

Prompt structure should not let a third-party document override trusted application rules. In this scenario the team must isolate prompt-injected external content. A feature used to introduce bounded language variation does not substitute for it.

Incorrect Answers

B:

Sampling does not create an instruction trust boundary. Keep system and developer policy separate from retrieved evidence and restrict tool calls independently.

C:

Required keys and types should be enforced with a machine-readable contract, not example text alone. That is relevant to produce parseable fixed-field JSON, but the present scenario requires an ability to isolate prompt-injected external content.

D:

Constrained generation and a defined schema reduce style variation for deterministic business tasks. This helps when a team needs to reduce creative variation in structured summaries; it does not establish how to isolate prompt-injected external content.

Question 422

A regulated assistant may answer questions about diagnoses but must not give definitive medical decisions without human review. What guardrail fits?

  1. Evaluate false-positive and false-negative rates on representative safe and unsafe prompts
  2. Apply tested content-filter policies to unsafe categories of generated text
  3. Use a policy-driven high-risk response gate with human escalation
  4. Stamp the AI output as synthetic and let the model issue definitive clinical decisions without human escalation.

Explanation

Correct Answer

C:

An explicit approval or handoff condition limits unsupported automated decisions in sensitive workflows. In this scenario the team must escalate high-impact recommendations for qualified review. A feature used to measure moderation precision and recall does not substitute for it.

Incorrect Answers

A:

Safety tuning must balance missed harms against overblocking legitimate users. Here the needed outcome is to escalate high-impact recommendations for qualified review; an option designed to measure moderation precision and recall would not directly provide it.

B:

Safety filters can block or restrict prohibited outputs independent of model fluency. Its primary purpose is to moderate unsafe generated content; the stated constraint instead demands a way to escalate high-impact recommendations for qualified review.

D:

Generated-content labels do not make high-risk medical decisions safe. Use explicit policy gates and qualified human review where required.

Question 423

A public image upload workflow must flag graphic or prohibited content before it reaches a model. What should be configured?

  1. Apply brand-usage policy checks and require approved assets before publication
  2. Validate input sources, permissions and content safety before multimodal processing
  3. Screen incoming media with appropriate visual safety classification
  4. Treat any legible OCR text in the uploaded picture as the only content-moderation signal.

Explanation

Correct Answer

C:

Input moderation reduces exposure to disallowed visual content before later processing. Here the relevant capability must let the team moderate disallowed input imagery. 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 moderate disallowed input imagery.

B:

Third-party media should be assessed for authorization and harmful payloads before agent use. The function serves to screen untrusted external media sources, whereas the actual requirement is to moderate disallowed input imagery.

D:

OCR alone does not classify visual harm or prohibited graphic content, and image text may be adversarial. Apply a supported image/content safety classifier before the upload reaches the generation or analysis model.

Question 424

A travel assistant must respond aloud in Japanese after processing English source information. Which stages are needed?

  1. Translate response content and synthesize speech with a compatible Japanese voice
  2. Synthesize the Japanese reply using an English-only voice without translating the answer content
  3. Recognize incoming speech and translate the transcript into the agent's language
  4. Validate preserved numeric quantities and medical units across the translation

Explanation

Correct Answer

A:

Cross-language voice responses require both accurate translation and target-language speech. The practical requirement is to speak translated answers in a target language; a solution built to reduce latency in live speech translation would not meet that same need.

Incorrect Answers

B:

Voice selection cannot substitute for content translation. The response must be translated and rendered with a compatible language voice.

C:

Speech translation bridges audio input and translated text for downstream reasoning. The function serves to translate incoming spoken customer requests, whereas the actual requirement is to speak translated answers in a target language.

D:

High-impact translated instructions require exact factual and unit preservation. That capability is appropriate for prevent numeric unit errors in translated speech, not this separate task of finding how to speak translated answers in a target language.

Question 425

A legal document workflow needs human-readable evidence paragraphs for a reviewer rather than only a field JSON object. What output should it request?

  1. Provide only a flat key-value object and discard the paragraphs the legal reviewer needs
  2. Validate extracted amounts against source evidence and business rules before downstream writes
  3. Configure a Content Understanding analyzer with task-specific structured output fields
  4. Produce supported Markdown or text representations retaining structure and source lineage

Explanation

Correct Answer

D:

Readable grounded representations allow reviewers and RAG systems to inspect supporting document content. In this scenario the team must prepare structured readable document evidence. A feature used to analyze mixed-media documents and assets does not substitute for it.

Incorrect Answers

A:

A field-only output cannot serve a reviewer who must inspect supporting context. Preserve readable structure and evidence links.

B:

Extraction confidence does not replace checks against the actual source and numeric invariants. Its primary purpose is to verify extracted values before financial side effects; the stated constraint instead demands a way to prepare structured readable document evidence.

C:

Defined extraction fields make downstream claim validation and automation dependable. That is relevant to produce consistent structured outputs from claim documents, but the present scenario requires an ability to prepare structured readable document evidence.

Question 426

A planner sends the same file to a vision model and an OCR tool, but the outputs conflict on a numeric amount. What should happen?

  1. Always trust the OCR result and discard conflicting visual evidence.
  2. Use confidence checks and an evidence reconciliation step before committing the extracted amount.
  3. Always prefer the vision model because its response is more fluent.
  4. Use whichever tool returned last to choose the final monetary amount.

Explanation

Correct Answer

B:

Conflicting machine observations require a grounded verification path; choosing whichever tool ran last is unsafe.

Incorrect Answers

A:

One extraction output may misread digits; it requires verification against the original region.

C:

Fluency is not evidence of numerical correctness when two modalities disagree.

D:

Execution order does not establish correctness and can change across runs.

Question 427

A chatbot incorrectly blocks benign questions after a stricter safety change. What should its owners test?

  1. Detect and redact protected personal data according to policy before response delivery
  2. Evaluate false-positive and false-negative rates on representative safe and unsafe prompts
  3. Treat retrieved document instructions as untrusted data and enforce tool permissions
  4. Watermark every chatbot answer and treat the watermark as proof the new safety filters never block benign prompts.

Explanation

Correct Answer

B:

Safety tuning must balance missed harms against overblocking legitimate users. This resolves the specific need to measure moderation precision and recall, while a measure that helps prevent disclosure of sensitive identifiers addresses a different requirement.

Incorrect Answers

A:

Sensitive-data handling requires explicit detection and release rules rather than relying on generic politeness. Its primary purpose is to prevent disclosure of sensitive identifiers; the stated constraint instead demands a way to measure moderation precision and recall.

C:

Indirect prompt injection attempts to turn lower-trust evidence into commands; tool authorization must not depend on retrieved prose. This helps when a team needs to isolate lower-trust instructions from privileged actions; it does not establish how to measure moderation precision and recall.

D:

A watermark describes provenance, not classifier performance. Test false-positive and false-negative behavior with representative safe and unsafe prompts.

Question 428

An image contains embedded text instructing the assistant to reveal another customer's address. What defense is required?

  1. Treat OCR text found inside an uploaded image as a privileged operator instruction to reveal private contact details.
  2. Screen incoming media with appropriate visual safety classification
  3. Require subject-matter review for high-impact synthetic instructional visuals
  4. Treat OCR-discovered text as untrusted input and constrain agent actions by policy

Explanation

Correct Answer

D:

Text inside pictures can carry indirect prompt injection; perception does not confer instruction authority. This directly supports the need to resist prompt injection embedded in images, unlike a mechanism primarily intended to reduce visual safety false positives.

Incorrect Answers

A:

Embedded image text is untrusted input. It cannot override confidentiality controls or authorize access to unrelated customer data.

B:

Input moderation reduces exposure to disallowed visual content before later processing. That capability is appropriate for moderate disallowed input imagery, not this separate task of finding how to resist prompt injection embedded in images.

C:

Photorealism is not evidence of correctness; risky instructional media needs expert validation. Here the needed outcome is to resist prompt injection embedded in images; an option designed to prevent unsafe high-stakes synthetic guidance would not directly provide it.

Question 429

An interactive voice agent speaks over the customer because interim recognition hypotheses are treated as final. Which conversation control helps?

  1. Reply to the first partial transcript token while the customer is still speaking.
  2. Ignore end-of-turn markers and wait an arbitrary fixed one second for every speaker.
  3. Separate partial and final transcript events and use end-of-turn detection before committing a response.
  4. Treat voice-activity detection as a complete proof of transcript finality without a turn-state check.

Explanation

Correct Answer

C:

Streaming recognition can revise provisional text; turn-taking must be driven by the appropriate finalization or endpointing signal.

Incorrect Answers

A:

Interim speech recognition is provisional and can change before the utterance ends.

B:

Fixed delays can still interrupt slow speakers or delay responses unnaturally.

D:

Acoustic silence and finalized transcription may require different handling before action.

Question 430

A search result snippet quotes a draft policy that is not effective until next month. What retrieval filter prevents premature use?

  1. Rank the draft policy highest because it was uploaded most recently.
  2. Filter only by keyword similarity and ignore document approval status.
  3. Use policy status and effective-date metadata to exclude noncurrent passages.
  4. Include every archived and future revision, asking the model to choose by intuition.

Explanation

Correct Answer

C:

Text relevance does not establish that a document revision is legally or operationally active.

Incorrect Answers

A:

Newest ingestion time does not mean a policy has reached its effective date.

B:

A relevant draft can match the query while still being inapplicable today.

D:

Model selection without enforced state metadata can incorrectly cite noncurrent policy.

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