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 50

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.

Question 491

Investigators cannot connect a reviewer override to the affected agent execution. What durable data is needed?

  1. Provide a policy-controlled kill switch that blocks actions but retains traces
  2. Use a human-in-the-loop draft-and-approve workflow for writes
  3. Record reviewer identity, decision, timestamp and affected task/version identifiers
  4. Enforce tool allowlists and privilege changes through trusted policy control

Explanation

Correct Answer

C:

Attributed, versioned review records support accountability and retrospective audit. In this scenario the team must link human override to the exact agent task. A feature used to stop further agent side effects without losing evidence does not substitute for it.

Incorrect Answers

A:

Stopping tool writes while preserving evidence supports safe incident response to problematic autonomy. Here the needed outcome is to link human override to the exact agent task; an option designed to stop further agent side effects without losing evidence would not directly provide it.

B:

Approval gates preserve human control over irreversible or regulated output. Its primary purpose is to stage side effects until review approval; the stated constraint instead demands a way to link human override to the exact agent task.

D:

Untrusted retrieved text must not elevate tool rights or authorize new sensitive capabilities. This helps when a team needs to prevent data-driven privilege escalation; it does not establish how to link human override to the exact agent task.

Question 492

A production agent's latency is high but monitoring records only one total request duration. What is the next instrumentation step?

  1. Store only total HTTP response time and omit timing of retrieval, tool execution and separate model calls.
  2. Use LLM intent parsing followed by deterministic eligibility and authorization checks
  3. Compute amounts with trusted code and reserve the model for ambiguous language
  4. Collect spans for retrieval, each model call, tool execution and waiting time

Explanation

Correct Answer

D:

Fine-grained tracing reveals which dependency dominates response time and where to optimize. This directly supports the need to attribute end-to-end latency by stage, unlike a mechanism primarily intended to combine flexible reasoning with strict output validation.

Incorrect Answers

A:

An aggregate duration cannot identify the slow stage in a tool-using agent. Add correlated spans and timing breakdowns across each component of the workflow.

B:

The model interprets natural language while trusted code enforces immutable business constraints. That capability is appropriate for combine language understanding with rule-enforced booking, not this separate task of finding how to attribute end-to-end latency by stage.

C:

Deterministic numeric computation is auditable and usually less costly and error-prone. Here the needed outcome is to attribute end-to-end latency by stage; an option designed to use code for exact arithmetic would not directly provide it.

Question 493

A risk owner needs to halt autonomous actions after repeated safety violations, without terminating observability. What controller should exist?

  1. Record reviewer identity, decision, timestamp and affected task/version identifiers
  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. Provide a policy-controlled kill switch that blocks actions but retains traces

Explanation

Correct Answer

D:

Stopping tool writes while preserving evidence supports safe incident response to problematic autonomy. This resolves the specific need to stop further agent side effects without losing evidence, while a measure that helps link human override to the exact agent task addresses a different requirement.

Incorrect Answers

A:

Attributed, versioned review records support accountability and retrospective audit. Its primary purpose is to link human override to the exact agent task; the stated constraint instead demands a way to stop further agent side effects without losing evidence.

B:

Document version lineage allows later assessment of whether a decision relied on outdated or manipulated evidence. That is relevant to record knowledge provenance for replay, but the present scenario requires an ability to stop further agent side effects without losing evidence.

C:

Approval gates preserve human control over irreversible or regulated output. This helps when a team needs to stage side effects until review approval; it does not establish how to stop further agent side effects without losing evidence.

Question 494

A tool-enabled agent must stop when no tool is allowed to satisfy the user's request. What is a safe planner outcome?

  1. Pretend the requested tool ran successfully and provide a fabricated confirmation.
  2. Choose the closest available tool even if it lacks the required permissions.
  3. Return an explicit inability or approval-required response without inventing tool access.
  4. Silently ignore the request while claiming that all objectives were completed.

Explanation

Correct Answer

C:

A model should not hallucinate the effects of tools it cannot call or bypass the authorization boundary.

Incorrect Answers

A:

Inventing external effects misleads users and bypasses the declared permission limits.

B:

An unrelated permitted tool cannot lawfully substitute for a disallowed action.

D:

A safe refusal or approval requirement must be explicit rather than misrepresenting completion.

Question 495

A procurement agent recommends purchases but must not self-approve them, and reviewers need replayable evidence. Which THREE controls address approval and auditability? Choose THREE.

  1. Record the request trace ID alongside authenticated actor and tool-write event
  2. Record only an unauthenticated “approved” flag without linking it to the actor or transaction.
  3. Require independent human or policy approval before the payment tool executes
  4. Allow the recommending agent to raise its own tool permissions when it judges a purchase time-sensitive.
  5. Permit the recommending agent to sign and submit the purchase without any independent review.
  6. Capture source document IDs, versions and retrieval timestamps in provenance

Explanation

Correct Answers

A:

Correlated IDs and identities let reviewers reconstruct which specific request produced an action. Request tracing, independent approval and source provenance provide attributable controls over decisions and external writes.

C:

Separating recommendation from authorization prevents an unchecked agent from approving its own financial side effects.

F:

Document version lineage allows later assessment of whether a decision relied on outdated or manipulated evidence.

Incorrect Answers

B:

A bare approval flag is not sufficient for replayable, attributable approval evidence.

D:

An agent-controlled privilege escalation breaks independent approval. A trusted authorizer must approve payment, and the evidence trail must tie it to a real actor.

E:

The same actor unilaterally approving its proposal defeats the required separation of duties.

Question 496

A developer wants to compare monthly token cost between two candidate models. What telemetry must be grouped?

  1. Aggregate input/output token usage and prices by deployment version and workload
  2. Aggregate all monthly token counts into one unlabeled figure without identifying the deployed models.
  3. Use LLM intent parsing followed by deterministic eligibility and authorization checks
  4. Let the model generate explanatory text and a schema validator enforce machine output

Explanation

Correct Answer

A:

Cost comparison needs both token volume and the model-specific billing rate. The decisive difference is whether the design can compare model serving cost correctly; the competing function serves to attribute end-to-end latency by stage.

Incorrect Answers

B:

A total without model identifiers cannot compare per-model pricing or consumption. Attribute token usage to the actual deployment/model and time period.

C:

The model interprets natural language while trusted code enforces immutable business constraints. Its primary purpose is to combine language understanding with rule-enforced booking; the stated constraint instead demands a way to compare model serving cost correctly.

D:

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 compare model serving cost correctly.

Question 497

A travel booking agent can understand requests but booking eligibility follows strict published rules. What design fits?

  1. Use LLM intent parsing followed by deterministic eligibility and authorization checks
  2. Rely on trace collection to decide whether booking requests meet the published eligibility rules.
  3. Aggregate input/output token usage and prices by deployment version and workload
  4. Enforce numerical approval thresholds with deterministic application rules

Explanation

Correct Answer

A:

The model interprets natural language while trusted code enforces immutable business constraints. In this scenario the team must combine language understanding with rule-enforced booking. A feature used to attribute end-to-end latency by stage does not substitute for it.

Incorrect Answers

B:

Tracing reports execution paths but cannot enforce eligibility. A deterministic policy check should decide whether a booking operation is allowed.

C:

Cost comparison needs both token volume and the model-specific billing rate. That is relevant to compare model serving cost correctly, but the present scenario requires an ability to combine language understanding with rule-enforced booking.

D:

Money limits are precisely evaluable and should not depend on probabilistic language generation. This helps when a team needs to apply fixed financial thresholds deterministically; it does not establish how to combine language understanding with rule-enforced booking.

Question 498

A team wants to choose whether to cache RAG responses or recompute them for each request. What evidence should it gather?

  1. Aggregate input/output token usage and prices by deployment version and workload
  2. Measure repeat-query frequency, source staleness tolerance and cache-hit correctness
  3. Compute amounts with trusted code and reserve the model for ambiguous language
  4. Enforce numerical approval thresholds with deterministic application rules

Explanation

Correct Answer

B:

Caching decisions depend on how often requests recur and whether underlying facts can safely become stale. This resolves the specific need to evaluate caching based on repetition and freshness, while a measure that helps compare model serving cost correctly addresses a different requirement.

Incorrect Answers

A:

Cost comparison needs both token volume and the model-specific billing rate. Its primary purpose is to compare model serving cost correctly; the stated constraint instead demands a way to evaluate caching based on repetition and freshness.

C:

Deterministic numeric computation is auditable and usually less costly and error-prone. This helps when a team needs to use code for exact arithmetic; it does not establish how to evaluate caching based on repetition and freshness.

D:

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 evaluate caching based on repetition and freshness.

Question 499

An agent uses a backend function that produces side effects, and the client drops connection after submission. What request record allows recovery?

  1. Generate a new operation ID and repeat the side effect after every client reconnect.
  2. Assume loss of the client connection means the backend rolled back all changes.
  3. Persist a durable operation ID with status polling and idempotent reattachment.
  4. Store only the final response text in the browser's local memory.

Explanation

Correct Answer

C:

Client disconnects do not imply server failure. Durable operation identity allows safe continuation without duplicate actions.

Incorrect Answers

A:

A new ID may create duplicate writes instead of reattaching to the existing operation.

B:

A dropped client connection does not prove the server canceled a submitted write. Record an idempotency key, operation ID and observable final status so reconnecting clients can recover the result without duplicating work.

D:

Client-side text cannot reliably track authoritative backend completion and deduplicate retries.

Question 500

A support workflow needs free-form intent understanding but deterministic monetary limits and detailed latency diagnosis. Which THREE mechanisms belong in the design? Choose THREE.

  1. Use LLM intent parsing followed by deterministic eligibility and authorization checks
  2. Enforce numerical approval thresholds with deterministic application rules
  3. Collect spans for retrieval, each model call, tool execution and waiting time
  4. Let the model generate explanatory text and a schema validator enforce machine output
  5. Let the language model approximate monetary totals without any deterministic calculation or validation.
  6. Measure repeat-query frequency, source staleness tolerance and cache-hit correctness

Explanation

Correct Answers

A:

The model interprets natural language while trusted code enforces immutable business constraints. Hard-coded threshold rules, stage tracing and model-plus-rule separation prevent unsafe spending decisions and expose bottlenecks.

B:

Money limits are precisely evaluable and should not depend on probabilistic language generation.

C:

Fine-grained tracing reveals which dependency dominates response time and where to optimize.

Incorrect Answers

D:

Generative interpretation and deterministic schema checks serve different roles in robust workflows. This addresses combine flexible reasoning with strict output validation. The scenario instead requires a way to enforce exact limits, trace latency and combine language with rules.

E:

Model-estimated amounts without trusted arithmetic and approval checks violate the deterministic monetary constraints.

F:

Caching decisions depend on how often requests recur and whether underlying facts can safely become stale. It would help when attempting to evaluate caching based on repetition and freshness; the current failure mode is different: enforce exact limits, trace latency and combine language with rules.

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