Microsoft AB-100 Practice Test 75

Question 740

A policy assistant now abstains on 35% of questions, half because no approved source exists. How should coverage be evaluated?

Force answers for every case and measure only apparent completion.
Report supported-answer coverage separately from justified abstention and unnecessary refusal on answerable cases.
Count all abstentions as hallucinations.
Exclude abstentions entirely and report only accuracy among answered questions.

Explanation

Correct Answer

Safe abstention and capability failure are different outcomes requiring different responses.

Incorrect Answers

Suppressing refusal can increase false authoritative statements.

Withholding unsupported policy can be the correct behavior.

That conceals the service’s inability to address many user requests.

Question 741

A receipt model correctly detects merchants but confuses 8 and 3 in invoice totals. What evaluation should govern its use for reimbursement?

Rate extracted totals by visual plausibility rather than source arithmetic.
Accept all amounts within 100% of the source total.
Score exact monetary extraction and reconciled payment outcomes with thresholds for high-impact errors.
Use only merchant-name accuracy because it is easiest to measure.

Explanation

Correct Answer

Merchant recognition is not enough to establish that the financial amount was extracted safely.

Incorrect Answers

Plausibility cannot substitute for receipt evidence and reconciliation.

Such an unconstrained tolerance would permit material overpayment.

Correct supplier identity does not prove an accurate total or tax amount.

Question 742

A workflow schedules a technician, orders a part and tells the customer that service is confirmed. Some runs never create the part order. What should the evaluation assert?

Count technician scheduling alone as full task completion.
Use the final chatbot message as the only ground truth.
Ignore missing order IDs if the model reports high certainty.
Test each required durable state and the final customer claim against matching authoritative receipts.

Explanation

Correct Answer

End-to-end success requires all mandatory workflow commitments, not just one booking or a generated statement.

Incorrect Answers

The part order is a required independent state and may still be absent.

Language output cannot prove purchase or service records were committed.

A generated confidence score cannot replace a durable transaction receipt from the service.

Question 743

A custom classifier sends uncertain cases to reviewers based on scores, but scores of 0.9 are correct only 60% of the time. Which validation is needed?

Assess probability calibration by score band and adjust human-review thresholds using held-out outcomes.
Assume scores are calibrated because they lie between zero and one.
Raise every score by 0.2 so fewer cases reach review.
Ignore the holdout set because the model learned the training data well.

Explanation

Correct Answer

A numeric score must align with observed error frequency before it controls high-impact automation.

Incorrect Answers

The output range alone does not establish empirical reliability.

Artificially increasing confidence worsens unsafe acceptance of wrong classifications.

Training fit does not verify confidence calibration on unseen examples.

Question 744

A compliance copilot uses a new prompt that reads more clearly but sometimes cites a rescinded policy. What evaluation should decide promotion?

Increase output length so every answer seems more detailed.
Run effective-date and revision-validity tests alongside clarity scoring and reject stale authoritative claims.
Promote when readers prefer the new writing style.
Count only the number of citations per response.

Explanation

Correct Answer

Clear language is desirable but cannot outweigh the legal validity of its referenced policy.

Incorrect Answers

Detail does not correct a superseded governing clause.

Style preference cannot validate the current policy source.

More references can still all point to obsolete or unapproved sources.

Question 745

A company splits invoice data randomly by page, so pages from the same invoice appear in both training and test sets. What evaluation change is required?

Keep the current split because the overall accuracy is high.
Add invoice IDs as model inputs solely to improve test accuracy.
Split by invoice/document identity and time as appropriate so related records do not leak across the holdout boundary.
Use a larger random page sample with the same invoice overlap.

Explanation

Correct Answer

Document-level separation measures generalization to genuinely unseen invoices.

Incorrect Answers

Inflated results may not represent production performance on new invoices.

Memorized identifiers can amplify leakage rather than demonstrate extraction skill.

More leaked pages do not remove correlated training/test contamination.

Question 746

An equipment failure classifier is 98% accurate overall but misses most “fire risk” incidents, which represent 1% of cases. What holdout design is necessary?

Treat all remaining classes as equivalent because they are more frequent.
Choose a threshold that never predicts fire risk to maximize common-class accuracy.
Assess only training accuracy for the fire-risk examples.
Include sufficient labeled fire-risk cases and report per-class recall and false-negative severity.

Explanation

Correct Answer

Overall accuracy can mask catastrophic misses of a rare high-impact class.

Incorrect Answers

Common-class scores cannot prove detection of the rare dangerous condition.

That strategy will systematically miss the critical fire-risk incidents and endanger users.

Memorizing labeled training examples does not establish real-world generalization to new incidents.

Question 747

A custom service classifier was developed on English calls from one region. It is proposed for multilingual use across 14 countries. What must occur before rollout?

Evaluate language and region cohorts on representative held-out data with documented fairness and accuracy thresholds.
Assume translation eliminates all classification error.
Use only the original region’s validation score.
Deploy globally and review errors only after serious customer complaints.

Explanation

Correct Answer

The new input distribution may differ in terminology, accent and policy context.

Incorrect Answers

Translation introduces its own ambiguity and cannot validate local legal terms.

One market cannot represent all new languages or business contexts.

High-impact regional failures require proactive evidence before release.

Question 748

A custom text extraction model is trained on clean PDFs, but live scans contain skew, blur and handwriting. Which evaluation supports readiness?

Ignore handwriting by deleting all handwritten sections before review.
Benchmark representative image-quality strata and report extraction errors and escalation rates for each.
Run another clean-PDF test and assume noisy inputs behave identically.
Lower the confidence threshold to force automated acceptance.

Explanation

Correct Answer

Noise robustness is a production requirement distinct from performance on clean training PDFs.

Incorrect Answers

Discarding source evidence can remove required business terms and values.

Clean data do not exercise degradation under poor capture conditions.

More permissive thresholds can increase wrong values in operational records.

Question 749

A model predicts late customer payments. The team evaluates it on random records from the same accounting periods used to build the features. What test reduces forward-looking leakage?

Train and test on all customer invoices simultaneously.
Judge performance only by how persuasive the explanation sounds.
Use a point-in-time training cutoff and evaluate on later periods using only features available at each prediction date.
Use subsequent payment settlement as an input feature for earlier predictions.

Explanation

Correct Answer

A temporal holdout prevents future payment outcomes from contaminating supposed predictive evidence.

Incorrect Answers

No separation remains to estimate how the model performs on new payment periods.

Narrative confidence does not measure future-payment prediction error.

Information that did not yet exist would artificially improve the benchmark.

Question 750

A tuned model improves average language fluency but changes the classification of sensitive legal documents. How should reviewers assess the revision?

Use only the tuning loss metric as proof of business correctness.
Change the benchmark examples until the new model passes.
Ignore a few sensitive errors when average prose is better.
Run both model versions against a fixed labeled legal holdout with error severity and decision trace comparison.

Explanation

Correct Answer

Stable test material reveals whether improvements introduce new regulated-document misclassifications.

Incorrect Answers

Lower training loss may not imply better held-out legal decisions.

Adjusting test cases to favor a release destroys the independent acceptance gate.

High-impact category mistakes require targeted review even if aggregate quality rises.

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

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)
Microsoft AB-100 Practice Test 75 - 10 Questions (Last Reviewed 27-Oct-2026)
Microsoft AB-100 Practice Test 76 - 10 Questions (Last Reviewed 26-Oct-2026)
Microsoft AB-100 Practice Test 77 - 10 Questions (Last Reviewed 25-Oct-2026)

Pass Your Microsoft AB-100 Exam Easy!

Question 740

Explanation

Correct Answer

Incorrect Answers

Question 741

Explanation

Correct Answer

Incorrect Answers

Question 742

Explanation

Correct Answer

Incorrect Answers

Question 743

Explanation

Correct Answer

Incorrect Answers

Question 744

Explanation

Correct Answer

Incorrect Answers

Question 745

Explanation

Correct Answer

Incorrect Answers

Question 746

Explanation

Correct Answer

Incorrect Answers

Question 747

Explanation

Correct Answer

Incorrect Answers

Question 748

Explanation

Correct Answer

Incorrect Answers

Question 749

Explanation

Correct Answer

Incorrect Answers

Question 750

Explanation

Correct Answer

Incorrect Answers

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