Microsoft AB-100 Practice Test 25

Question 240

An AI-led service process completes 1,000 requests, but 150 are later reversed for wrong eligibility. Which metric is needed for management?

Initial request count only, because reversals happen later.
Optimize average inference cost per initial request while excluding the cost and count of eligibility reversals.
Treat the ERP’s initial acceptance response as final eligibility success without reconciling later reversals.
Correct sustained completion after reversals, with error-rework and exception costs.

Explanation

Correct Answer

Business value depends on durable correct outcomes, not temporary successful-looking transitions.

Incorrect Answers

Ignoring reversals inflates completion quality.

Token efficiency is an operating-cost component but cannot substitute for sustained correct outcomes and remediation cost.

A provisional submission or acceptance can be reversed; sustained correct fulfillment and rework cost determine success.

Question 241

An agent’s fraud detection accuracy changes abruptly after a new data connector goes live. What should monitoring compare?

Compare only the agent’s response latency before and after the connector release, without examining values supplied to the fraud model.
Versioned connector inputs, feature distributions and error labels before versus after rollout.
Compare connector uptime alone and treat healthy HTTP responses as proof that fraud features have not changed.
Compare the number of agent messages before and after release, without retaining connector input schemas or labels.

Explanation

Correct Answer

Input provenance and error segmentation help isolate connector-induced drift.

Incorrect Answers

A connector can change input semantics while latency stays constant. Fraud-detection drift requires feature and labeled-outcome comparisons.

A healthy transport connection can still return altered or incorrectly mapped features. Input distributions and labeled errors must be examined.

Conversation volume does not explain a shift in classification correctness; versioned data and segmented errors are decisive.

Question 242

A model router’s monthly expense falls 40%, but premium customer complaint rates double. What should the architect conclude?

Lower cost automatically proves overall success.
Stop collecting quality signals once savings are recorded.
Accept elevated premium-customer complaints as an unavoidable tradeoff because the overall model bill fell, without impact analysis.
The routing change is not proven successful; evaluate risk-adjusted business quality and restore safer routing where needed.

Explanation

Correct Answer

Cost optimization must be constrained by outcome and customer-segment quality requirements.

Incorrect Answers

Expenses are only one component of value.

Removing quality telemetry conceals harm.

Cost savings cannot establish acceptable service quality when the high-value segment shows substantially worse outcomes.

Question 243

A retail agent performs retrieval, a model call and a refund tool call. Errors occur only when a specific connector version is used. What trace attribute is most useful?

Record the exact connector/tool version alongside correlated request and error information.
Log only the connector display name without recording its deployed version or resolved tool binding.
Omit tool versions because all connectors behave identically.
Record the model version and temperature but omit the invoked refund-connector version from correlated traces.

Explanation

Correct Answer

Version-aware traces allow a technical regression to be linked to the changed dependency.

Incorrect Answers

A friendly name can remain constant across releases, preventing attribution to the version correlated with the observed errors.

Tool behavior can change with releases.

Model settings cannot distinguish a connector-specific regression if the exact tool release is absent from the failing request trace.

Question 244

A Copilot agent receives negative feedback from one multilingual region, but total satisfaction stays high. How should the team prioritize investigation?

Assume every locale uses the same language-specific entity recognition settings.
Translate all feedback into one language and aggregate it without keeping the affected region or intent identifiers.
Segment feedback by locale and intent, verifying translation and grounding errors with affected users.
Ignore regional feedback because the enterprise average is positive.

Explanation

Correct Answer

Segmented incident evidence identifies a localized quality problem hidden by global aggregates.

Incorrect Answers

Language and cultural differences can affect recognition accuracy.

Aggregation can erase the locale-specific retrieval or translation failure that the complaint trend reveals.

A high average can mask material group harm.

Question 245

A service agent’s response time rises after a new CRM lookup is introduced. Which telemetry comparison identifies the change?

Compare correlated tool-call latency and timeout rates before and after the connector rollout.
Compare only the total chat duration without separate timing for the added CRM lookup.
Ignore dependency timing and attribute the delay to the model automatically.
Increase the model deployment capacity without checking whether the added CRM lookup has become the dominant wait.

Explanation

Correct Answer

Stage-by-stage latency and error signals can isolate the new lookup path.

Incorrect Answers

An end-to-end measure confirms slower service but cannot attribute delay to the new connector stage without correlated per-tool timing.

Agent pipelines may spend most of their time outside inference.

A model capacity change may leave an external CRM bottleneck unchanged; traces should isolate the stage that regressed.

Question 246

A company uses an agent to draft procurement orders. Of 500 proposals, 430 pass staff review, 40 require corrections and 30 are rejected. What outcome measure supports improvement?

Claim 100% success because proposals were generated.
Optimize token cost per proposed order while omitting staff rejection and correction categories.
Report the 86% first-pass acceptance rate plus correction and rejection reasons.
Ignore corrections because the final orders may eventually be approved.

Explanation

Correct Answer

First-pass business acceptance and error categories identify real workflow quality.

Incorrect Answers

Generated proposals are not accepted transactions.

Cheap proposal generation can still create rework; first-pass acceptance and correction causes are needed to improve purchasing value.

Corrections consume staff effort and affect value.

Question 247

A customer support agent’s hallucination rate increases only when warranty PDFs exceed a new length threshold. What should the tuning team inspect first?

Increase the model context limit first without checking whether long warranty sections are excluded by chunking or retrieval.
Reprocess long PDFs at higher OCR resolution even though the failure begins at a document-length threshold and section coverage is unverified.
Retrieval chunking, truncation and source coverage across document sizes before retraining.
Fine-tune the response model on long-document summaries before verifying the cited warranty clauses survive retrieval.

Explanation

Correct Answer

Long sources can expose context-window and retrieval boundary defects needing source-level diagnosis.

Incorrect Answers

A larger context window cannot restore passages never retrieved or truncated during preprocessing; first establish evidence coverage.

Image quality improvements target recognition errors, not missing passages from chunk limits or retrieval truncation.

Tuning may mask missing context but cannot validate source coverage. Inspect retrieval boundaries and citations before model changes.

Question 248

A cheaper model handles 90% of incoming requests but fails more on the 10% that need financial calculations. What should the router do?

Hide calculation failures from the quality dashboard.
Choose the larger model only when the user asks for a shorter answer.
Keep routing all requests cheaply because the average cost target was reached.
Route calculation-heavy cases based on measured error thresholds and verify numeric outputs independently.

Explanation

Correct Answer

Cost and correctness should be jointly optimized by risk class; arithmetic requires grounded independent validation.

Incorrect Answers

Concealing failures prevents safe routing changes.

Desired response length is not a reliable risk detector.

Average savings can be outweighed by financial errors.

Question 249

An AI agent logs tool results but cannot connect them to individual business requests. What tracing improvement is needed?

Propagate a stable request correlation ID through model, connector, tool and ERP transaction spans.
Use a different random identifier at every step with no linkage.
Store only the last model-generated answer.
Turn off tool logs to reduce storage costs.

Explanation

Correct Answer

A common trace identity allows incident reconstruction and partial workflow diagnosis.

Incorrect Answers

Unlinked identifiers prevent end-to-end analysis.

Final prose omits intermediate actions.

Disabling logs hides the root cause.

Question 250

A knowledge agent receives positive feedback on article summaries but repeated reports of missing citations. How should product management interpret that?

Close all citation issues because users like the writing.
Record citation availability as a separate quality defect rather than allowing satisfaction to mask traceability gaps.
Ask the model to fabricate a plausible URL for every answer.
Display article titles as citations without checking that each title resolves to the paragraph supporting the answer.

Explanation

Correct Answer

A useful summary still needs independently checkable evidence for factual and compliance-sensitive claims.

Incorrect Answers

Good sentiment does not ensure trustworthy attribution.

Invented URLs mislead reviewers.

Decorative or unresolved titles are not traceable evidence, even when the answer is otherwise well-liked.

Agentic AI Business Solutions Architect (AB-100) Practice Tests

Microsoft AB-100 Practice Test 23 - 10 Questions (Last Reviewed 22-Oct-2026)
Microsoft AB-100 Practice Test 24 - 10 Questions (Last Reviewed 21-Oct-2026)
Microsoft AB-100 Practice Test 25 - 10 Questions (Last Reviewed 20-Oct-2026)
Microsoft AB-100 Practice Test 26 - 10 Questions (Last Reviewed 19-Oct-2026)
Microsoft AB-100 Practice Test 27 - 10 Questions (Last Reviewed 18-Oct-2026)

Pass Your Microsoft AB-100 Exam Easy!

Question 240

Explanation

Correct Answer

Incorrect Answers

Question 241

Explanation

Correct Answer

Incorrect Answers

Question 242

Explanation

Correct Answer

Incorrect Answers

Question 243

Explanation

Correct Answer

Incorrect Answers

Question 244

Explanation

Correct Answer

Incorrect Answers

Question 245

Explanation

Correct Answer

Incorrect Answers

Question 246

Explanation

Correct Answer

Incorrect Answers

Question 247

Explanation

Correct Answer

Incorrect Answers

Question 248

Explanation

Correct Answer

Incorrect Answers

Question 249

Explanation

Correct Answer

Incorrect Answers

Question 250

Explanation

Correct Answer

Incorrect Answers

Agentic AI Business Solutions Architect (AB-100) Practice Tests