Microsoft AB-100 Practice Test 72

Question 710

A legal policy agent refuses to answer unsupported clauses and routes them to counsel. A dashboard labels every abstention as a task failure. What correction is appropriate?

Tell the agent never to abstain regardless of evidence.
Count every escalation as a completed legal approval.
Exclude all abstentions from evaluation denominators without documenting them.
Separate justified abstentions from avoidable inability to answer, using policy support and escalation results.

Explanation

Correct Answer

An evidence-based refusal can be a correct safe behavior when authoritative guidance is missing.

Incorrect Answers

Forced answers can convert an incomplete policy library into confident misinformation.

Routing a case to counsel is not equivalent to an authorized decision.

Removing cases obscures coverage gaps and prevents transparent risk analysis.

Question 711

A CRM deduplication agent reports success for each merge it proposes. Some merge attempts are rejected because the records belong to different customers. What should monitoring distinguish?

Proposed merges, validated executed merges, rejected unsafe merges and post-merge reversals.
Record every proposal as a successful deduplication.
Suppress rejected attempts because they reduce success percentage.
Use the model’s confidence as proof two identities were merged.

Explanation

Correct Answer

Only committed, approved outcomes should count as successful record corrections.

Incorrect Answers

An unexecuted suggestion has not changed the customer data.

Rejections can demonstrate safety and reveal model or matching weaknesses.

Confidence is not an authoritative CRM result or identity reconciliation.

Question 712

A service organization claims savings from an agent but ticket volumes differ greatly between peak and off-peak months. Which monitoring design supports ROI attribution?

Assume every recommended response removes one full human case.
Compare normalized case cohorts, verified handling time and actual review costs against the appropriate baseline.
Compare absolute labor costs in two unrelated months only.
Count agent sign-ins as hours saved.

Explanation

Correct Answer

Volume and seasonality must be considered to avoid attributing external shifts to the agent.

Incorrect Answers

Agents may still require review, correction and escalation costs.

Staffing and seasonal demand may explain the difference without any AI benefit.

Question 713

A claims extraction model was validated on paper forms, but incoming mobile photographs now dominate. What drift signal should trigger review?

Monitor only model endpoint uptime.
Disable source-channel labels to simplify the dashboard.
Compare input-format mix, extraction errors and confidence calibration against the approved baseline by channel.
Assume a fixed model is accurate on every future input modality.

Explanation

Correct Answer

A distribution shift can degrade field reliability without changing model deployment code.

Incorrect Answers

Availability can be perfect while extracted values become wrong.

Without segmented provenance the team cannot identify which input population changed.

Validation on paper scans does not prove robustness to angled or low-light phone photos.

Question 714

A policy agent’s answers change after both its prompt and knowledge library were updated. No release identifiers were logged. What would enable cause attribution?

Update both again until the response sounds better.
Assume the source changed because the prompt title stayed constant.
Measure only the number of users chatting after the release.
Version prompts and source snapshots independently and run baseline questions across controlled combinations.

Explanation

Correct Answer

Separable versions allow testing whether the behavior change came from instructions or grounding.

Incorrect Answers

Uncontrolled changes make root-cause identification harder and can introduce new defects.

Titles do not prove underlying instruction bytes or index contents were unchanged.

Adoption does not isolate the factual source of changed answers.

Question 715

A model upgrade leads an agent to answer previously escalated legal questions without citations. What regression check is necessary?

Track unsupported-answer and justified-abstention rates on labeled high-risk legal cases before promotion.
Optimize only for the number of direct answers.
Use the model vendor’s generic benchmark without company policy cases.
Suppress escalation events from monitoring.

Explanation

Correct Answer

Lower refusal can look like better completion while increasing unjustified legal claims.

Incorrect Answers

Forced coverage can reward unsupported confident statements.

Public benchmarks may not evaluate the specific source authority and escalation requirement.

Removing visibility prevents detection of changed safety decisions.

Question 716

After a model deployment change, an agent emits dates as `MM/DD/YYYY` while the service expects ISO `YYYY-MM-DD`. Which tuning signal is needed?

Remove schema checks so more actions appear to succeed.
Measure schema-valid action payload rate and reject incompatible dates with explicit field-level errors.
Let the service guess ambiguous date formats.
Count model completions as successful actions even if payload validation fails.

Explanation

Correct Answer

Tool correctness includes format and meaning; a valid-looking date may be misinterpreted across locales.

Incorrect Answers

Bypassing validation spreads errors into downstream systems.

An ambiguous date can schedule the wrong task or reverse day and month.

Generated arguments are not equivalent to accepted business transactions.

Question 717

An agent’s answers degrade gradually over six weekly releases. The knowledge documents and model routes change frequently. Which monitoring approach catches regression reliably?

Replace all test questions each week without overlapping baselines.
Ask the agent to declare itself improved after every release.
Run a versioned representative evaluation set on each release and retain task-level trend data with configuration provenance.
Rely on aggregate weekly message count to infer answer quality.

Explanation

Correct Answer

Consistent benchmark cases show which releases introduced specific factual or action errors.

Incorrect Answers

Constantly changing samples prevent direct longitudinal comparisons.

Self-reported progress is not independent verification of actual task outcomes.

Usage growth can occur despite declining correctness.

Question 718

A payment-validation agent has only a 0.2% error rate, but one error type could pay the wrong beneficiary. What governance should apply?

Ignore the error because global accuracy exceeds 99%.
Apply the same tolerance to wrong beneficiary and punctuation errors.
Stop monitoring after the model first reaches its overall target.
Set a stricter threshold and immediate incident route for wrong-beneficiary outcomes, separate from benign formatting errors.

Explanation

Correct Answer

Risk severity rather than overall average accuracy determines the response to sensitive financial failures.

Incorrect Answers

Aggregate accuracy cannot justify material unauthorized transfers.

The financial harm and remediation urgency of these failures differ markedly.

New conditions and deployment drift can introduce serious failures after launch.

Question 719

A product-answer agent starts failing every question about one catalog after a nightly ingestion job errors. Model predictions elsewhere are stable. What should monitoring flag?

A source-ingestion freshness/coverage failure with the affected product-catalog partition.
Immediately retrain the model as the only fix.
Assume a burst of model hallucinations with no source diagnosis.
Use unrelated catalog entries to fill missing product prices.

Explanation

Correct Answer

Missing or stale source data can degrade grounded answers without changes to underlying model weights.

Incorrect Answers

Retraining cannot recover authoritative records that the index failed to load.

The shared catalog outage points to a distinct dependency failure.

Other products cannot establish the specific missing item’s current terms.

Question 720

A service assistant encounters new repair codes and proposes tuning itself on all recent conversations, including restricted customer notes. How should adaptation be governed?

Disable model-version records after tuning to reduce administrative load.
Route novel-code errors into an approved data-review and evaluation cycle with consent and scoped training access.
Automatically train on every conversation after each misclassification.
Ignore the new repair codes while reporting no model errors.

Explanation

Correct Answer

Model improvement must preserve data rights, reproducible tests and promotion approval.

Incorrect Answers

Without version provenance future regressions cannot be traced or rolled back.

Unreviewed online training can ingest protected data and cause uncontrolled behavior drift.

Silent coverage gaps can lead to incorrect service guidance.

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

Microsoft AB-100 Practice Test 70 - 10 Questions (Last Reviewed 01-Nov-2026)
Microsoft AB-100 Practice Test 71 - 10 Questions (Last Reviewed 31-Oct-2026)
Microsoft AB-100 Practice Test 72 - 10 Questions (Last Reviewed 30-Oct-2026)
Microsoft AB-100 Practice Test 73 - 10 Questions (Last Reviewed 29-Oct-2026)
Microsoft AB-100 Practice Test 74 - 10 Questions (Last Reviewed 28-Oct-2026)

Pass Your Microsoft AB-100 Exam Easy!

Question 710

Explanation

Correct Answer

Incorrect Answers

Question 711

Explanation

Correct Answer

Incorrect Answers

Question 712

Explanation

Correct Answer

Incorrect Answers

Question 713

Explanation

Correct Answer

Incorrect Answers

Question 714

Explanation

Correct Answer

Incorrect Answers

Question 715

Explanation

Correct Answer

Incorrect Answers

Question 716

Explanation

Correct Answer

Incorrect Answers

Question 717

Explanation

Correct Answer

Incorrect Answers

Question 718

Explanation

Correct Answer

Incorrect Answers

Question 719

Explanation

Correct Answer

Incorrect Answers

Question 720

Explanation

Correct Answer

Incorrect Answers

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