Microsoft AB-100 Practice Test 18

Question 170

A Copilot Studio support topic gets two equally plausible intent matches. Which fallback behavior should be used?

Execute both tool actions and keep whichever returns first.
Ask the user to choose the intended topic or escalate without executing either risky action.
Randomly select one action because the model must always proceed.
Execute the first matching intent by topic registration order even when two high-impact actions score equally.

Explanation

Correct Answer

A clarification step prevents ambiguity from causing unintended service changes.

Incorrect Answers

Parallel side effects can produce multiple transactions.

Random selection can trigger the wrong workflow.

Topic ordering is not user confirmation and can run the wrong transaction; clarification must precede side effects.

Question 171

A product agent uses an index whose source ACLs changed after ingestion. What is required before granting a user search results?

Remove all document ownership metadata to speed search.
Re-evaluate effective document permissions using current caller identity and source access policy.
Ask the model to suppress titles but still retrieve full restricted content.
Assume permissions captured months ago remain valid indefinitely.

Explanation

Correct Answer

Permission changes must be honored at the retrieval authorization boundary.

Incorrect Answers

Removing metadata weakens access enforcement.

Late hiding does not prevent model exposure.

Stale ACLs can reveal revoked content.

Question 172

A canvas app that schedules service visits includes an AI time estimate. What should happen before creating the actual booking?

Use a chat transcript instead of obtaining booking API confirmation.
Write a calendar entry without checking conflicts or identity.
Validate technician availability and customer consent in the scheduling service when the booking is committed.
Treat the AI estimate as a guaranteed booked appointment.

Explanation

Correct Answer

A prediction assists planning but cannot replace authoritative scheduling validation.

Incorrect Answers

A conversation is not the system-of-record confirmation.

Unvalidated writes may double-book resources.

Estimates are not reservations.

Question 173

An intelligent Power Platform solution depends on a connector with low availability. What reliability control should be documented?

Increase the connector request timeout indefinitely so an unavailable dependency is never reported as a failure.
Assume a successful pilot guarantees future availability.
Hide dependency errors from monitoring dashboards.
Define circuit-breaker or bounded-retry behavior, meaningful fallback and recovery testing.

Explanation

Correct Answer

Resilience requires explicit behavior when an external component is unavailable.

Incorrect Answers

A longer unbounded wait can hang cases and hide outages; recovery requires bounded timeouts, a truthful fallback and detection.

One pilot cannot establish steady-state reliability.

Hidden errors prevent diagnosis and recovery.

Question 174

An assistant routes IT tickets to one of six fixed teams based on the user description. Which evaluation best fits a standard intent-classifier design?

Route each incident to the team whose description has the closest embedding match without measuring errors against labeled examples.
Accept unconstrained free-text team names as valid without routing checks.
Evaluate only the classifier’s average confidence instead of its precision and recall for each assigned support team.
Test labeled utterances and measure per-team confusion, recall and precision.

Explanation

Correct Answer

The bounded labels permit measurable classification quality before automated assignment.

Incorrect Answers

Similarity can be a classifier feature, but fixed-team routing needs validated labels and confusion analysis to detect wrong assignments.

Unrecognized team names can break deterministic routing.

Average confidence can be uncalibrated and conceal misrouting of uncommon or high-priority teams.

Question 175

An agent flow completes a data update but its webhook callback is delivered twice. What design prevents duplicate downstream notifications?

Change correlation IDs on each callback to hide duplicates.
Delete the callback log after the first attempt.
Send both notifications and assume duplicates are harmless.
Deduplicate by a stable event ID and make the notification step idempotent.

Explanation

Correct Answer

Duplicate deliveries are normal in many integrations; idempotency prevents repeated side effects.

Incorrect Answers

New IDs prevent effective deduplication.

Logs are needed for event processing evidence.

Users can be misled by duplicate actions.

Question 176

A Copilot Studio prompt action produces a support response and structured case priority. Security detects HTML script strings inside a free-text field. Which validation is needed?

Treat generated text as untrusted content, escape it in the UI and validate the typed priority separately.
Use a valid priority enum as proof that other fields are harmless.
Allow free-text fields to change case permissions.
Execute returned HTML because it comes from the agent.

Explanation

Correct Answer

Output sanitization and field-specific checks prevent untrusted model text becoming active code or authority.

Incorrect Answers

One valid field does not validate the entire payload.

Content fields should not grant authorization changes.

Generated text is not trusted executable content.

Question 177

A global service Copilot needs to interpret “service agreement” and “support contract” as synonyms while keeping separate contract tiers. What mapping approach fits?

Change the underlying CRM tier field based on wording alone.
Treat the two phrases as permission to ignore contract tier restrictions.
Merge synonyms using fuzzy search and use the highest-scoring matching contract record as the new contract identifier.
Map approved synonyms to one business object but preserve tier and entitlement fields for the final answer.

Explanation

Correct Answer

Vocabulary normalization can improve understanding without changing factual account entitlements.

Incorrect Answers

Language should not mutate business records.

Synonyms do not remove service-tier conditions.

Text similarity can resolve vocabulary but cannot establish a unique account contract; the mapped object must keep authoritative IDs, tiers and entitlements.

Question 178

A manufacturer wants a custom Foundry model to detect a proprietary fault code. An existing model works on most categories, but fails one class. What evidence should precede training?

A claim that proprietary names always require a new foundation model.
Unapproved production data copied into a shared training folder.
A representative held-out failure analysis showing the class gap and adequate governed labeled samples.
Use a few favorable fault-code demonstrations as a training trigger without measuring precision or recall on a labeled held-out class.

Explanation

Correct Answer

Model customization needs measurable benefit, controlled training data and lifecycle governance.

Incorrect Answers

Vocabulary novelty alone does not prove base models are unsuitable.

Unconsented sensitive data may violate policy.

A demonstration does not quantify the isolated failure or whether governed custom training will outperform the current model.

Question 179

A firm wants a Microsoft 365 Copilot extension to retrieve project records from its controlled ticket database. What identity behavior must be maintained?

Retrieve the entire ticket table and hide restricted rows after generation.
Grant the extension an all-ticket administrator session for every user.
Assume Microsoft 365 sign-in automatically grants external database rights.
Resolve caller identity and filter ticket records according to the effective user permissions.

Explanation

Correct Answer

Integration must preserve both host identity and external data authorization.

Incorrect Answers

Late hiding does not prevent model exposure to forbidden content.

A shared admin token can leak other teams’ records.

Host authentication is not blanket external resource permission.

Question 180

A Copilot Studio agent must invoke an external financial validation API. What extensibility contract should be designed?

An action that accepts customer IDs without verifying record-level access.
An authenticated, typed action with explicit input checks, errors and least-privilege permissions.
A connector that treats any HTTP 200 body as confirmed financial validation.
A free-text URL assembled by the model with a tenant administrator secret.

Explanation

Correct Answer

A governed tool requires reliable schema, authorization, response interpretation and safe failure handling.

Incorrect Answers

Unverified IDs enable cross-customer retrieval or mutation.

Transport success is not substantive business approval.

Open-ended generated URLs can redirect requests or expose secrets.

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

Microsoft AB-100 Practice Test 16 - 10 Questions (Last Reviewed 29-Oct-2026)
Microsoft AB-100 Practice Test 17 - 10 Questions (Last Reviewed 28-Oct-2026)
Microsoft AB-100 Practice Test 18 - 10 Questions (Last Reviewed 27-Oct-2026)
Microsoft AB-100 Practice Test 19 - 10 Questions (Last Reviewed 26-Oct-2026)
Microsoft AB-100 Practice Test 20 - 10 Questions (Last Reviewed 25-Oct-2026)

Pass Your Microsoft AB-100 Exam Easy!

Question 170

Explanation

Correct Answer

Incorrect Answers

Question 171

Explanation

Correct Answer

Incorrect Answers

Question 172

Explanation

Correct Answer

Incorrect Answers

Question 173

Explanation

Correct Answer

Incorrect Answers

Question 174

Explanation

Correct Answer

Incorrect Answers

Question 175

Explanation

Correct Answer

Incorrect Answers

Question 176

Explanation

Correct Answer

Incorrect Answers

Question 177

Explanation

Correct Answer

Incorrect Answers

Question 178

Explanation

Correct Answer

Incorrect Answers

Question 179

Explanation

Correct Answer

Incorrect Answers

Question 180

Explanation

Correct Answer

Incorrect Answers

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