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 28

Question 270

A tool returns two independent objects with the same tool-call identifier and agent output becomes confused. What must the controller preserve?

  1. Treat tool-call IDs as interchangeable because the response payloads have similar shapes
  2. Define strict identifier fields and resolve identities server-side before writing
  3. Mark one of the two tool results as an access denial instead of detecting their duplicated tool-call identifier.
  4. Correlate each returned result to the original unique call identifier

Explanation

Correct Answer

D:

The model execution protocol relies on stable identity for pending and completed tool operations. The practical requirement is to match tool results to the right call IDs; a solution built to enforce business rules for a valid but high-risk tool call would not meet that same need.

Incorrect Answers

A:

Result correlation depends on exact call identity. Reusing an identifier for independent tool calls risks assigning evidence to the wrong action.

B:

Unambiguous IDs and server checks prevent display-name collisions from changing the wrong record. This helps when a team needs to bind tool operations to verified identity identifiers; it does not establish how to match tool results to the right call IDs.

C:

A duplicated correlation ID is a protocol integrity fault, not an authorization decision. Enforce one response per expected call identifier and handle duplicates explicitly.

Question 271

A finance report shows unexpectedly high generated-token charges for simple category labels. Which generation control is most relevant?

  1. Raise the generation maximum to several thousand tokens for every single-label classification request.
  2. Break down end-to-end latency by model generation and individual tool calls
  3. Apply per-client rate limiting and bounded input-token admission rules
  4. Limit output length and require a compact fixed response schema

Explanation

Correct Answer

D:

A fixed bounded answer discourages unnecessary verbose generation and reduces billable output tokens. This directly supports the need to bound output token usage, unlike a mechanism primarily intended to diagnose serving quota exhaustion under bursts.

Incorrect Answers

A:

A long maximum output budget does not enforce concise labels and may increase waste. Use a compact response schema and output limit appropriate to the fixed set of categories.

B:

Per-stage timing separates model serving delay from slow external dependencies. That capability is appropriate for find slow external tool stages, not this separate task of finding how to bound output token usage.

C:

Tenant-level limits prevent one caller from exhausting shared serving capacity. Here the needed outcome is to bound output token usage; an option designed to enforce fair use for shared AI serving would not directly provide it.

Question 272

An agent tool returns instructions to reveal a hidden system prompt along with otherwise useful search results. How should the controller process this content?

  1. Promote the tool's instruction text into the assistant's system instruction context.
  2. Reveal the hidden prompt to demonstrate that the tool result was processed.
  3. Run every command described by the retrieved result before returning the search excerpts.
  4. Separate untrusted tool data from higher-priority instructions and ignore the malicious instruction text.

Explanation

Correct Answer

D:

Tool output is evidence, not policy. The response may use the factual data while rejecting instruction attempts.

Incorrect Answers

A:

Tool output is lower-trust data and must not override trusted application directives.

B:

Disclosing internal instructions would satisfy the malicious instruction rather than isolate it.

C:

Untrusted retrieval text is evidence, not permission to execute commands or expose secrets.

Question 273

A small fraction of users create very long requests that monopolize shared deployment capacity. What control protects fairness?

  1. Apply per-client rate limiting and bounded input-token admission rules
  2. Break down end-to-end latency by model generation and individual tool calls
  3. Validate regional model quota and request capacity changes before cutover
  4. Limit generated answer size while leaving unbounded input prompts free to consume the shared request budget.

Explanation

Correct Answer

A:

Tenant-level limits prevent one caller from exhausting shared serving capacity. The decisive difference is whether the design can enforce fair use for shared AI serving; the competing function serves to find slow external tool stages.

Incorrect Answers

B:

Per-stage timing separates model serving delay from slow external dependencies. Here the needed outcome is to enforce fair use for shared AI serving; an option designed to find slow external tool stages would not directly provide it.

C:

Quotas are resource and deployment constraints; capacity should be confirmed before directing production traffic. Its primary purpose is to prepare regional serving capacity; the stated constraint instead demands a way to enforce fair use for shared AI serving.

D:

Very long incoming prompts can monopolize prefill, context and token-rate resources even when outputs are short. Apply input-size admission controls and fair per-customer concurrency or rate limits.

Question 274

A tool's network request returns HTTP 403 and the agent confidently says there were zero customer records. How should the result be represented?

  1. Convert every authorization-denied tool response into an empty successful customer-record array.
  2. Validate submitted arguments against the function schema and reject the invalid call
  3. Design side-effecting operations with idempotency and outcome reconciliation
  4. Pass a structured authorization failure rather than a fabricated empty successful payload

Explanation

Correct Answer

D:

Authorization denial does not establish the absence of data and must be handled as failure. Here the relevant capability must let the team separate forbidden access from empty results. An alternative targeting bind tool operations to verified identity identifiers solves another problem.

Incorrect Answers

A:

HTTP 403 means the caller was denied, not that the requested records do not exist. Preserve a structured authorization error and stop downstream reasoning from interpreting it as valid empty data.

B:

Schema validation prevents nonsensical payloads reaching financial side effects. The function serves to reject tool argument type mismatches, whereas the actual requirement is to separate forbidden access from empty results.

C:

Safe retry requires recognizing requests already executed and not duplicating effects. That capability is appropriate for avoid repeat writes after uncertain network outcomes, not this separate task of finding how to separate forbidden access from empty results.

Question 275

Users report only total response duration, but the team needs to locate slow retrieval versus slow model generation. Which telemetry granularity helps?

  1. Track only total HTTP request duration across all services.
  2. Measure server CPU utilization and assume it equals model inference latency.
  3. Log only the timestamp when the final answer reaches the client.
  4. Instrument correlated retrieval, tool-call, first-token, and total-generation latency spans.

Explanation

Correct Answer

D:

Segmented traces identify where time is spent; aggregate end-to-end duration cannot isolate search, tool execution and output generation.

Incorrect Answers

A:

An aggregate number cannot distinguish slow search, waiting on a tool, or generation time.

B:

Frontend CPU does not reveal per-stage delays in remote model or retrieval calls.

C:

A final timestamp alone cannot attribute tail latency to the correct dependency.

Question 276

A model proposes a valid refund request, but company policy requires manager approval over a threshold. Where should this be enforced?

  1. Reject every refund above the company threshold as a permanent permission error without evaluating required manager approval.
  2. Correlate each returned result to the original unique call identifier
  3. Check approval and value limits inside the application tool boundary before execution
  4. Design side-effecting operations with idempotency and outcome reconciliation

Explanation

Correct Answer

C:

Policy cannot depend on model compliance with prompt instructions when real money moves. This directly supports the need to enforce business rules for a valid but high-risk tool call, unlike a mechanism primarily intended to separate forbidden access from empty results.

Incorrect Answers

A:

The stated threshold may permit a refund after approved escalation. Check a trusted approval artifact and prevent the external write until that policy condition is met.

B:

The model execution protocol relies on stable identity for pending and completed tool operations. That capability is appropriate for match tool results to the right call IDs, not this separate task of finding how to enforce business rules for a valid but high-risk tool call.

D:

Safe retry requires recognizing requests already executed and not duplicating effects. Its primary purpose is to avoid repeat writes after uncertain network outcomes; the stated constraint instead demands a way to enforce business rules for a valid but high-risk tool call.

Question 277

A support service faces burst throttling while paying excessively for lengthy repeated prompts. Which TWO actions address both quota pressure and avoidable token usage? Choose TWO.

  1. Minimize repeated context and use supported caching or retrieval for reusable evidence
  2. Measure response latency without recording token volumes, throttling, admission limits or repeated prompt-context costs.
  3. View the current quota dashboard but leave burst admission and concurrency unchanged.
  4. Send every throttled request again immediately with no admission limit, while shortening only the response text.
  5. Apply per-client rate limiting and bounded input-token admission rules

Explanation

Correct Answers

A:

Avoiding redundant prompt tokens reduces cost and can improve latency without changing the required answer. Minimizing repeated context cuts token waste and reduces request load. Per-client rate limits and bounded input admission mitigate bursts before provider throttling.

E:

Admission control and bounded request sizes prevent sudden bursts from consuming the complete available serving quota.

Incorrect Answers

B:

Latency alone does not manage burst admission or remove unnecessary repeated input tokens. Rate limiting and smaller reusable context address the two named causes.

C:

Visibility into quota does not itself reduce request bursts or enforce per-client backpressure.

D:

Immediate unbounded retries worsen burst throttling. Address both provider backoff/concurrency and avoidable repeated input content; output-length trimming alone does not satisfy the two distinct constraints.

Question 278

A function may be retried after a transport timeout even though the first attempt succeeded. What interface characteristic matters?

  1. Treat a transport timeout as access denied and immediately submit an unrelated new payment request without idempotency tracking.
  2. Expose only explicitly allowed callable functions in the runtime tool registry
  3. Check approval and value limits inside the application tool boundary before execution
  4. Design side-effecting operations with idempotency and outcome reconciliation

Explanation

Correct Answer

D:

Safe retry requires recognizing requests already executed and not duplicating effects. The decisive difference is whether the design can avoid repeat writes after uncertain network outcomes; the competing function serves to separate forbidden access from empty results.

Incorrect Answers

A:

A timeout does not prove the first side effect failed or was unauthorized. Use a stable idempotency key and inspect operation status before retrying writes.

B:

A textual description must not grant capabilities outside server-controlled tool permissions. Here the needed outcome is to avoid repeat writes after uncertain network outcomes; an option designed to restrict executable tools with an allowlist would not directly provide it.

C:

Policy cannot depend on model compliance with prompt instructions when real money moves. Its primary purpose is to enforce business rules for a valid but high-risk tool call; the stated constraint instead demands a way to avoid repeat writes after uncertain network outcomes.

Question 279

A summarizer is still fast but its factual accuracy drops after product policy changes. Which monitoring process reveals the problem?

  1. Measure citation coverage and grounded-answer relevance on a labeled test set
  2. Evaluate grounding faithfulness against newly curated reference examples
  3. Track safety-filter outcomes and policy-violation rates by use case
  4. Measure response latency only and assume unchanged speed proves the policy-grounded answers remain accurate.

Explanation

Correct Answer

B:

Quality checks must detect whether answers remain supported by changing authoritative evidence. This resolves the specific need to monitor grounded factual correctness, while a measure that helps monitor evidence use in generated answers addresses a different requirement.

Incorrect Answers

A:

Grounding-quality evaluation detects failures between successful retrieval and supported generation. Its primary purpose is to monitor evidence use in generated answers; the stated constraint instead demands a way to monitor grounded factual correctness.

C:

Safety measurements reveal harmful output patterns that general request success metrics miss. This helps when a team needs to observe content-policy violations; it does not establish how to monitor grounded factual correctness.

D:

An assistant can stay fast while citing outdated policy. Reevaluate answer faithfulness against updated authoritative test cases.

Question 280

A generative model uses a high temperature for a regulated classification field and outputs unstable labels. What inference adjustment fits?

  1. Use lower-variance sampling and constrain labels through a supported schema or enum.
  2. Increase temperature further to encourage diverse category spellings.
  3. Remove the allowed-label enum and allow unrestricted explanatory prose.
  4. Change only the embedding index dimension, leaving generation settings unchanged.

Explanation

Correct Answer

A:

Regulated classification benefits from repeatable outputs and explicit allowed labels instead of unbounded free-form generation.

Incorrect Answers

B:

More randomness makes a regulated closed-label field less stable and harder to validate.

C:

Removing the closed list of classification labels allows arbitrary text that downstream systems cannot validate. Keep an enum or equivalent schema constraint and test deterministic decoding settings on representative regulated examples.

D:

Embedding dimensions control retrieval compatibility rather than classification output variance.

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