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 49

Question 480

An app receives model responses from multiple regions and must determine whether one region is causing more safety refusals. What monitoring key helps?

  1. Label refusal metrics by deployment region, model version and safety-policy version.
  2. Compare regional refusal counts without accounting for model and policy versions.
  3. Monitor only global refusal totals and hide the deployment region label.
  4. Track server uptime alone and infer that safety policies are identical.

Explanation

Correct Answer

A:

A single global refusal total obscures regional configuration and version-specific behavior.

Incorrect Answers

B:

A higher regional total may reflect different traffic or settings rather than location-specific behavior.

C:

Aggregation removes the dimension needed to diagnose one region's higher rejection rate.

D:

Availability does not reveal refusal distribution or safety-filter configuration drift.

Question 481

A financial agent can approve its own high-value transfer after generating a recommendation. What control closes the gap?

  1. Let the financial agent temporarily grant itself payment-approval permission after recommending a transfer.
  2. Provide a policy-controlled kill switch that blocks actions but retains traces
  3. Use a human-in-the-loop draft-and-approve workflow for writes
  4. Require independent human or policy approval before the payment tool executes

Explanation

Correct Answer

D:

Separating recommendation from authorization prevents an unchecked agent from approving its own financial side effects. The practical requirement is to prevent self-authorized high-value transfers; a solution built to prevent data-driven privilege escalation would not meet that same need.

Incorrect Answers

A:

Self-elevation defeats separation of duties. High-value payments must be authorized by an independent human or enforceable policy approval gate before the tool executes.

B:

Stopping tool writes while preserving evidence supports safe incident response to problematic autonomy. This helps when a team needs to stop further agent side effects without losing evidence; it does not establish how to prevent self-authorized high-value transfers.

C:

Approval gates preserve human control over irreversible or regulated output. The function serves to stage side effects until review approval, whereas the actual requirement is to prevent self-authorized high-value transfers.

Question 482

A code-generating agent self-critiques confidently but sometimes ships invalid code and fabricated citations. Which TWO independent validations are necessary? Choose TWO.

  1. Cap critique-and-revision passes using quality improvement thresholds
  2. Check cited statements against the retrieved source passages
  3. Accept a generated retry routine as correct when the same model says it looks robust, without executing any tests.
  4. Accept a second ungrounded self-critique as proof that the citations are accurate.
  5. Run deterministic compilation and tests and feed failures into bounded repair

Explanation

Correct Answers

B:

Validating each cited claim against the actual source passage detects fabricated or unsupported citations independently of the model’s confidence. Compiler/tests and citation-to-source checking provide observable evidence beyond model self-assessment. A second opinion is not enough when both share the same fabricated evidence.

E:

Executable tests provide stronger correction signals than asking the model whether code works.

Incorrect Answers

A:

Bounded reflection prevents token waste when repeated attempts no longer improve output. Here, however, the stated objective is to validate code execution and cited evidence; this option focuses on stop unproductive self-critique loops.

C:

A self-declared robust retry routine is not independent validation. Compile and exercise failure paths in a deterministic test harness.

D:

Repeating the same model’s unsupported judgment does not independently establish citation validity.

Question 483

A supervisor wants an agent to draft case notes but never commit them until reviewed. Which operating mode fits?

  1. Give the agent permanent write permission and treat its generated case-note draft as automatically approved.
  2. Capture source document IDs, versions and retrieval timestamps in provenance
  3. Use a human-in-the-loop draft-and-approve workflow for writes
  4. Enforce tool allowlists and privilege changes through trusted policy control

Explanation

Correct Answer

C:

Approval gates preserve human control over irreversible or regulated output. Here the relevant capability must let the team stage side effects until review approval. An alternative targeting stop further agent side effects without losing evidence solves another problem.

Incorrect Answers

A:

Drafting and committing are separate permissions. Use a draft-and-approve workflow in which an authorized reviewer explicitly authorizes the write.

B:

Document version lineage allows later assessment of whether a decision relied on outdated or manipulated evidence. The function serves to record knowledge provenance for replay, whereas the actual requirement is to stage side effects until review approval.

D:

Untrusted retrieved text must not elevate tool rights or authorize new sensitive capabilities. Here the needed outcome is to stage side effects until review approval; an option designed to prevent data-driven privilege escalation would not directly provide it.

Question 484

An expense approval workflow has fixed policy thresholds but uses an LLM to decide whether a transfer exceeds $10,000. What architecture is preferable?

  1. Use distributed trace spans as the authoritative rule for determining whether the expense exceeds the allowed dollar threshold.
  2. Measure repeat-query frequency, source staleness tolerance and cache-hit correctness
  3. Enforce numerical approval thresholds with deterministic application rules
  4. Aggregate input/output token usage and prices by deployment version and workload

Explanation

Correct Answer

C:

Money limits are precisely evaluable and should not depend on probabilistic language generation. This resolves the specific need to apply fixed financial thresholds deterministically, while a measure that helps attribute end-to-end latency by stage addresses a different requirement.

Incorrect Answers

A:

Tracing is observational and should not decide a fixed financial eligibility rule. Validate the amount and apply the threshold deterministically before permitting a transfer.

B:

Caching decisions depend on how often requests recur and whether underlying facts can safely become stale. That is relevant to evaluate caching based on repetition and freshness, but the present scenario requires an ability to apply fixed financial thresholds deterministically.

D:

Cost comparison needs both token volume and the model-specific billing rate. The function serves to compare model serving cost correctly, whereas the actual requirement is to apply fixed financial thresholds deterministically.

Question 485

The audit log stores only agent output text, not which document version it used. What should be added?

  1. Retain only the final model's answer and erase the document version IDs used during retrieval.
  2. Capture source document IDs, versions and retrieval timestamps in provenance
  3. Record reviewer identity, decision, timestamp and affected task/version identifiers
  4. Record the request trace ID alongside authenticated actor and tool-write event

Explanation

Correct Answer

B:

Document version lineage allows later assessment of whether a decision relied on outdated or manipulated evidence. This directly supports the need to record knowledge provenance for replay, unlike a mechanism primarily intended to stage side effects until review approval.

Incorrect Answers

A:

The final answer alone cannot identify which source revision justified the conclusion. Preserve authenticated document IDs, version markers and retrieval provenance.

C:

Attributed, versioned review records support accountability and retrospective audit. Here the needed outcome is to record knowledge provenance for replay; an option designed to link human override to the exact agent task would not directly provide it.

D:

Correlated IDs and identities let reviewers reconstruct which specific request produced an action. Its primary purpose is to attribute a tool side effect to a request and actor; the stated constraint instead demands a way to record knowledge provenance for replay.

Question 486

A generative pipeline contains a deterministic tax formula and a model-based narrative summary. Which implementation choice reduces calculation errors?

  1. Ask the model to estimate tax mentally, then publish the estimate as authoritative.
  2. Round model-generated tax totals without recalculating from validated inputs.
  3. Replace the tax formula with prompt examples instead of executing code.
  4. Compute tax using validated code and use the model only to explain the verified result.

Explanation

Correct Answer

D:

LLMs are useful for language but a deterministic financial calculation should remain in a trusted implementation.

Incorrect Answers

A:

Language generation can introduce arithmetic errors even if the explanation sounds convincing.

B:

Rounding an unchecked result does not establish the correct deterministic tax amount.

C:

Examples cannot reliably enforce an exact financial calculation across cases.

Question 487

An agent requests more tool privileges after reading a document that says an emergency override is needed. What must decide privilege expansion?

  1. Capture source document IDs, versions and retrieval timestamps in provenance
  2. Enforce tool allowlists and privilege changes through trusted policy control
  3. Require independent human or policy approval before the payment tool executes
  4. Record the request trace ID alongside authenticated actor and tool-write event

Explanation

Correct Answer

B:

Untrusted retrieved text must not elevate tool rights or authorize new sensitive capabilities. The decisive difference is whether the design can prevent data-driven privilege escalation; the competing function serves to record knowledge provenance for replay.

Incorrect Answers

A:

Document version lineage allows later assessment of whether a decision relied on outdated or manipulated evidence. That capability is appropriate for record knowledge provenance for replay, not this separate task of finding how to prevent data-driven privilege escalation.

C:

Separating recommendation from authorization prevents an unchecked agent from approving its own financial side effects. Its primary purpose is to prevent self-authorized high-value transfers; the stated constraint instead demands a way to prevent data-driven privilege escalation.

D:

Correlated IDs and identities let reviewers reconstruct which specific request produced an action. That is relevant to attribute a tool side effect to a request and actor, but the present scenario requires an ability to prevent data-driven privilege escalation.

Question 488

A hybrid customer agent must summarize exceptions in free text but enforce a strict JSON contract for downstream systems. How should work be divided?

  1. Use LLM intent parsing followed by deterministic eligibility and authorization checks
  2. Measure repeat-query frequency, source staleness tolerance and cache-hit correctness
  3. Use telemetry trace spans in place of a structured output schema for the downstream invoice data.
  4. Let the model generate explanatory text and a schema validator enforce machine output

Explanation

Correct Answer

D:

Generative interpretation and deterministic schema checks serve different roles in robust workflows. The practical requirement is to combine flexible reasoning with strict output validation; a solution built to combine language understanding with rule-enforced booking would not meet that same need.

Incorrect Answers

A:

The model interprets natural language while trusted code enforces immutable business constraints. That is relevant to combine language understanding with rule-enforced booking, but the present scenario requires an ability to combine flexible reasoning with strict output validation.

B:

Caching decisions depend on how often requests recur and whether underlying facts can safely become stale. This helps when a team needs to evaluate caching based on repetition and freshness; it does not establish how to combine flexible reasoning with strict output validation.

C:

Trace spans record operations and timing, not a contract for business data. The application should enforce the required JSON schema independently of free-form summarization.

Question 489

A red-team evaluation discovers an agent can invoke an unauthorized export endpoint through a tool alias. What is the correct containment?

  1. Trust the tool's friendly display name even when it resolves to an unapproved endpoint.
  2. Apply an enforceable tool allowlist and server-side authorization independent of model-generated tool names.
  3. Add a prompt instruction saying the agent should not use export tools.
  4. Log export calls after they execute but leave the endpoint unrestricted.

Explanation

Correct Answer

B:

Tool aliases or planner choices cannot substitute for backend authorization. Restrict the permitted operations and reject unknown endpoints.

Incorrect Answers

A:

Aliases can conceal capabilities; backend authorization must bind actual operations and destinations.

C:

Prompt text is not an enforceable restriction on a reachable unauthorized endpoint.

D:

Logging a forbidden export after the fact does not stop the sensitive data from leaving the system. Enforce an allowlist at the tool gateway and block unauthorized tool aliases and endpoints before execution.

Question 490

An invoice pipeline invokes two language models for simple arithmetic instead of checking numeric fields. What should change?

  1. Let the model generate explanatory text and a schema validator enforce machine output
  2. Enforce numerical approval thresholds with deterministic application rules
  3. Compute amounts with trusted code and reserve the model for ambiguous language
  4. Add latency tracing but continue asking another language model to compute every invoice subtotal.

Explanation

Correct Answer

C:

Deterministic numeric computation is auditable and usually less costly and error-prone. Here the relevant capability must let the team use code for exact arithmetic. An alternative targeting combine flexible reasoning with strict output validation solves another problem.

Incorrect Answers

A:

Generative interpretation and deterministic schema checks serve different roles in robust workflows. This helps when a team needs to combine flexible reasoning with strict output validation; it does not establish how to use code for exact arithmetic.

B:

Money limits are precisely evaluable and should not depend on probabilistic language generation. The function serves to apply fixed financial thresholds deterministically, whereas the actual requirement is to use code for exact arithmetic.

D:

Tracing helps identify slow steps but does not reduce unnecessary model arithmetic. Use deterministic numeric checks for extracted amounts and calculations.

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