Microsoft AB-100 Practice Test 3

Question 20

A shared department-policy index still returns restricted HR documents to staff after they transfer to another department. The chat user interface shows their new department correctly, but the indexed results are cached under a privileged service credential. What must be corrected to stop this specific leak?

Hide the restricted search hits after model generation using the department displayed by the browser.
Keep the privileged index results but instruct the model to avoid citing policies from the employee’s former department.
Enforce current employee ACLs before retrieval reaches the model, invalidate stale cached authorizations, and run role-change denial tests.
Give every employee the same service index credential so results are consistent across departments.

Explanation

Correct Answer

The defect is an outdated authorization boundary at query time. Revoking cached rights and testing the transferred employee ensures retired entitlements are actually denied.

Incorrect Answers

UI-level filtering is too late: privileged cached passages could already have entered the model context, logs or generated answer.

A prompt cannot revoke an employee’s former department rights or prevent sensitive passages from being retrieved under an overprivileged identity.

A shared privileged credential reproduces the root cause by bypassing the requesting user’s current department membership.

Question 21

A finance assistant must produce an audit-ready explanation of a cashflow forecast assembled from several source tables. What provenance is most important?

Use the most recent filename as proof that its values passed reconciliation.
Cite only the final chart image without a record of data sources.
Keep lineage from each derived measure to its input table version, time range and transformation rule.
Give the model permission to invent missing transactions so totals reconcile.

Explanation

Correct Answer

Reproducible lineage lets reviewers check the calculations and identify stale or conflicting sources.

Incorrect Answers

Naming conventions are not independent assurance of data quality.

A rendered chart without source lineage does not permit verification.

Fabricated records conceal errors and invalidate the forecast.

Question 22

A service agent answers policy questions from documents owned by Legal, HR and Sales. No team accepts responsibility for policy corrections. What should the program establish?

Assign each document family a steward with publication rights, review interval and escalation path.
Make every employee a global knowledge editor to distribute responsibility.
Assign ownership to the team that uploaded the most recent document in each policy family.
Treat the newest retrieved document as authoritative even without a controlling owner.

Explanation

Correct Answer

Stewardship sets responsibility for correctness and revision authority across the knowledge lifecycle.

Incorrect Answers

Unrestricted editing makes policy provenance and signoff harder.

Uploader identity is not a formally delegated policy steward; a cross-department correction needs named authority and a documented escalation chain.

Recency without approval is not enough to establish binding policy.

Question 23

A warranty service process uses fixed product eligibility rules plus subjective photographs of damage. Which design balances automation with exceptions?

Run rule-based checks first, use AI to summarize photo evidence, and route uncertain assessments for authorized human review.
Let visual confidence alone override expired warranty rules.
Reject all requests automatically when a customer uploads more than one photo.
Use one model output to alter policy and issue an immediate refund.

Explanation

Correct Answer

Separating unambiguous eligibility from visual judgment preserves policy control while improving reviewer throughput.

Incorrect Answers

An image-confidence score does not determine legal warranty entitlement.

Photo quantity is not a legitimate proxy for evidence quality.

Agent actions must not both redefine and execute policy without approval.

Question 24

An agent prototype for a procurement desk completed ten happy-path cases but failed whenever the supplier ERP was unavailable. What evidence should define readiness?

Test bounded ERP outages, safe retries, partial workflows and escalation outcomes against measurable acceptance criteria.
Use only aggregate response-time metrics from successful ERP calls and omit error and retry outcomes from pilot acceptance.
Increase the client retry count without an idempotency key or upper bound whenever the supplier ERP connection fails.
Exclude outage cases from testing because they are inconvenient for a demo.

Explanation

Correct Answer

The business workflow must behave safely under realistic failure modes; error cases are part of production readiness.

Incorrect Answers

A fast happy path does not prove safe partial-failure handling or truthful status reporting during a dependency outage.

Unbounded retries can amplify supplier load and may duplicate transactions after an ambiguous network timeout.

Ignoring known failures would overstate readiness and leave an operational gap.

Question 25

A bank proposes an agent to process account changes. Each request must be reversible and the originator accountable. Which step is essential before delegating execution?

Allow the agent to update any account because the workflow sounds routine.
Map each change to verified customer identity, explicit authorization and an auditable reversal process.
Use a free-text customer name as a substitute for account ownership verification.
Discard actor identity after a successful tool call to protect the prompt.

Explanation

Correct Answer

Account changes require identity, permission and compensating controls before execution is granted.

Incorrect Answers

Routine appearance does not confer authority to alter financial records.

Unverified names can be duplicated and cannot establish account ownership.

Discarding actor provenance damages audit and incident response.

Question 26

A strategy team uses an AI assistant to predict churn, but its data warehouse joins the latest billing snapshot to outdated support cases. What analytic correction is needed?

Discard support records entirely because a join is more work than using billing alone.
Evaluate the churn classifier using only how accurate its emailed retention summaries appear to reviewers rather than churn outcomes.
Mix subsequent renewal outcomes into features during historical training without temporal controls.
Align feature timestamps to the prediction event and verify no future or stale records leak into evaluation.

Explanation

Correct Answer

Valid prediction requires point-in-time aligned source data and controls against leakage and stale context.

Incorrect Answers

Dropping useful features is not an evidence-based fix to poor time alignment.

A plausible explanation has no guarantee of future churn prediction accuracy or temporal separation of features.

Future outcome leakage produces misleadingly optimistic model scores.

Question 27

A regulated loan process permits agents to draft recommendations but reserves lending approval for an accredited officer. What control should the design include?

Use the model’s predicted customer approval probability as the officer signoff required by the lending transaction policy.
An enforceable approval task bound to the officer’s identity before the lending service commits a decision.
Grant the agent full credit-limit authority during periods of high workload.
Ask the model to write “approved by a human” after the record is saved.

Explanation

Correct Answer

The independent authorization boundary preserves legal accountability while allowing AI to assist with document review.

Incorrect Answers

Prediction probability is not a verified credit officer identity or an approval recorded in the financial system.

Operational load cannot waive an explicit human signoff requirement.

A claim of approval is not a verified decision from the responsible actor.

Question 28

An agent suggests credit limits from a customer master that contains duplicate records with different risk scores. What should the architect require before deployment?

Average risk scores automatically without confirming they use the same metric.
Resolve identities and conflicting risk features at their authoritative source and add a data-reconciliation rejection rule.
Hide both records in the UI after feeding them to the agent for inference.
Select whichever duplicate record returns first in vector search.

Explanation

Correct Answer

Conflicting identity features can produce unsafe financial decisions and require source reconciliation before grounding.

Incorrect Answers

Scores using different vintages or definitions cannot be averaged safely.

Late UI hiding does not prevent incorrect model reasoning.

Retrieval order is not a business rule for assigning risk.

Question 29

A sales agent recommends discounts based on pricing updates that may arrive hourly. Which data strategy controls stale-price errors?

Use a discount suggestion if the agent quotes a price from a recent-looking indexed promotion, without rechecking the transaction-effective rate.
Use a time-bounded price feed and validate the effective price at the transaction service before quoting a binding offer.
Assume prices from yesterday’s semantic index are still valid for every customer.
Let the model invent a discount if the API is delayed.

Explanation

Correct Answer

A binding quote needs current authoritative pricing and a final transaction-time validation.

Incorrect Answers

An index may lag live promotion changes; the payment or pricing service must validate the effective discount at commit.

A relevant indexed quote can be stale once price lists change.

Invented discounts can violate margin and approval policy.

Question 30

A support assistant analyzes scanned forms with some unreadable signatures. The capture pipeline emits a blank signature field. What data-quality action is appropriate?

Mark the field as missing with source confidence and require verification before using it for an authorized action.
Treat blank as affirmative consent because the scan was processed successfully.
Use the agent’s fluent summary as proof the physical signature exists.
Populate an unreadable signature field with the value from a previous document by the same customer without verifying consent.

Explanation

Correct Answer

Missing evidence must not be converted into a legal assertion; verification and explicit missing-value handling are needed.

Incorrect Answers

Successful extraction does not establish positive consent.

Generated prose is not direct evidence of a signature.

Historical signatures cannot be transplanted into a new legally meaningful approval event; missing evidence requires verification.

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

Microsoft AB-100 Practice Test 1 - 10 Questions (Last Reviewed 13-Nov-2026)
Microsoft AB-100 Practice Test 2 - 10 Questions (Last Reviewed 12-Nov-2026)
Microsoft AB-100 Practice Test 3 - 10 Questions (Last Reviewed 11-Nov-2026)
Microsoft AB-100 Practice Test 4 - 10 Questions (Last Reviewed 10-Nov-2026)
Microsoft AB-100 Practice Test 5 - 10 Questions (Last Reviewed 09-Nov-2026)

Pass Your Microsoft AB-100 Exam Easy!

Question 20

Explanation

Correct Answer

Incorrect Answers

Question 21

Explanation

Correct Answer

Incorrect Answers

Question 22

Explanation

Correct Answer

Incorrect Answers

Question 23

Explanation

Correct Answer

Incorrect Answers

Question 24

Explanation

Correct Answer

Incorrect Answers

Question 25

Explanation

Correct Answer

Incorrect Answers

Question 26

Explanation

Correct Answer

Incorrect Answers

Question 27

Explanation

Correct Answer

Incorrect Answers

Question 28

Explanation

Correct Answer

Incorrect Answers

Question 29

Explanation

Correct Answer

Incorrect Answers

Question 30

Explanation

Correct Answer

Incorrect Answers

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