Microsoft AB-100 Practice Test 20

Question 190

A contact agent can reason in multiple steps and can receive audio. A business sponsor assumes these modes guarantee every answer is supported. What evaluation is required?

Verify factual grounding, tool permissions and modality-specific errors for the actual deployed configuration.
Treat reasoning mode as proof that source citations are correct.
Assume microphone capture guarantees that the customer was authenticated.
Skip scenario tests if the model speaks confidently.

Explanation

Correct Answer

A feature mode changes interaction behavior but does not certify evidence accuracy or authorization.

Incorrect Answers

Reasoning traces can still use unsupported facts.

Voice input is not an identity attestation.

Confidence does not replace test evidence.

Question 191

A SharePoint agent answers questions in Teams but sometimes misses a document update made moments earlier. What source behavior should the design specify?

Assume Teams membership synchronizes every SharePoint index instantly.
Ask the model to fabricate a newer paragraph when none is found.
A documented indexing freshness expectation with refresh or a clearly timestamped response.
Use older cached content as current without version metadata.

Explanation

Correct Answer

Search freshness can lag source changes and must be treated as a measurable dependency.

Incorrect Answers

Collaboration membership does not guarantee index currency.

Generated text cannot replace approved document revisions.

Stale content can give incorrect guidance.

Question 192

A Foundry solution customizes a language model for product descriptions. The evaluator sees lower training loss but worse held-out factual accuracy. What should the architect do?

Use the training-set loss curve as the promotion criterion and omit worse held-out accuracy from the release gate.
Increase response temperature to compensate for lower held-out factual accuracy after tuning.
Promote immediately because training loss is the only valid metric.
Block promotion and examine generalization and training-data provenance before release.

Explanation

Correct Answer

Training fit can improve while real-world accuracy degrades; held-out validation controls promotion.

Incorrect Answers

Training loss reflects fit to observed training data, whereas deterioration on independent product claims is direct evidence of generalization risk.

Sampling changes cannot repair unsupported facts or a failed holdout baseline and may introduce more variation.

Training loss alone is not proof of deployment readiness.

Question 193

An M365 Copilot agent needs to perform approved ServiceNow-like incident lookups, but users expect the result to appear within their existing Microsoft 365 workflow. What design evidence matters?

A requirement to replace Microsoft 365 with a new public website.
Supported extension surfaces, user context preservation, resource authorization and operational support.
Granting the extension all tenant permissions because it appears in Microsoft 365.
A model demo that invents incident IDs without tool calls.

Explanation

Correct Answer

The value of an extension is integrated access to verified workflows within governed identities.

Incorrect Answers

A separate public site adds unwanted integration and sign-in risks.

Host placement does not grant universal resource access.

A demonstration without a service query proves little.

Question 194

A Copilot Studio agent has an extension to create new customer cases. Case creation sometimes succeeds twice after retries. What change is needed?

Generate a new case ID on each retry by default.
Return “case created” without checking the service response.
Use idempotency keys and a stable client request ID with readback before retrying.
Ignore duplicate-case warnings and close one later.

Explanation

Correct Answer

Idempotent operations reduce accidental duplicate records during transient failures.

Incorrect Answers

New IDs turn one intention into multiple records.

Agent text is not a service acknowledgment.

Cleanup later is avoidable and costly.

Question 195

A sales agent discovers an MCP tool that exports every customer contract. What is the best tool-approval question?

Can the tool be reached by the model, irrespective of data permissions?
Can the agent call it quickly enough to improve response latency?
Approve contract export because its MCP catalog description declares it read-only, without verifying record-level access or data minimization.
Does this caller have a legitimate need and enforced contract-level authorization for that export?

Explanation

Correct Answer

Availability and purpose are insufficient without role and resource authorization.

Incorrect Answers

Reachability is not proof of permission.

Performance is not a replacement for access checks.

A read-only bulk export can still expose restricted contracts; the tool needs caller-specific need and enforceable contract entitlements.

Question 196

A customer agent uses Computer Use to move data from a legacy website to Finance. The form displays a hidden required checkbox before committing. What should the workflow do?

Validate the full form state and require confirmation of the resulting Finance record before claiming success.
Ignore unexpected validation prompts and keep clicking submit.
Assume text entry alone commits the transaction.
Assume a successful browser confirmation banner proves the corresponding Finance ledger transaction persisted.

Explanation

Correct Answer

UI automation must verify application outcomes and handle dynamic required controls.

Incorrect Answers

Blind repeated submission can duplicate actions.

Typing is not persistence.

A browser banner is weaker evidence than a backend transaction receipt; reconcile the authoritative Finance record and its identifiers after submission.

Question 197

A Copilot Studio voice agent must respond with a spoken translation of customer support instructions. Which error handling matters?

Detect speech recognition uncertainty and confirm critical details before executing any action.
Use a text-only tool as proof it received and processed audio.
Treat low-confidence transcripts as verified customer authorization.
Save all low-confidence speech transcripts for indefinite model retraining without consent or sensitivity classification.

Explanation

Correct Answer

Voice interfaces need modality-specific confidence and privacy controls beyond text generation.

Incorrect Answers

Text-only APIs do not imply audio processing.

Recognition errors can change account numbers or intents.

Indefinite transcript retention is not required for correct voice handling and may violate recording and privacy controls; confirm critical fields with minimal governed storage.

Question 198

A Teams agent summarizes a SharePoint project plan using a search index that includes confidential annexes. What must its context construction enforce?

Treat every file linked in Teams as public to all participants.
Use the meeting owner’s global privileges for all attendees.
Retrieve all annexes and suppress the confidential titles only in the displayed answer.
Security trimming and source/version selection for each attendee before document chunks enter the prompt.

Explanation

Correct Answer

Prompt context must be built only from sources the current caller can access.

Incorrect Answers

Sharing a link does not always grant document access.

Owner privileges may exceed each attendee’s rights.

Post-answer censorship cannot remove exposure through model context.

Question 199

A product catalog assistant needs a customized Foundry model, but new products arrive every week. Which approach controls the data lifecycle?

Evaluate whether retrieval updates suffice and version any required tuning with regular holdout tests.
Retrain blindly on every draft product description without review.
Freeze the model forever and assume new products are recognized.
Skip source freshness checks because the model has been customized.

Explanation

Correct Answer

Changing factual catalogs often favor updatable grounding over needless model retraining.

Incorrect Answers

Unapproved drafts can contaminate training.

Frozen models cannot know new product facts without sources.

Customization does not eliminate data currency needs.

Question 200

A company extends Microsoft 365 Copilot with a service-management agent that handles confidential incident reports. Select TWO requirements for safe extensibility.

Authorize incident retrieval by matching the user’s department name to keywords in the incident title.
Apply incident-level retrieval permissions and audit each tool access.
Give the extension unrestricted tenant administrator access because it is internal.
Copy all incident attachments into every chat prompt for better completeness.
Use an approved extension path integrated into Microsoft 365 user context.

Explanation

Correct Answers

Authorization must be enforced at source and tool boundaries for each request.

Supported integration lets the agent operate inside intended collaboration surfaces.

Incorrect Answers

A matching title is not an identity or incident-level entitlement; the backend must check the authenticated user and each record before retrieval.

Global privilege can leak unrelated incident evidence.

Mass copying attachments violates data-minimization controls.

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

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)
Microsoft AB-100 Practice Test 21 - 10 Questions (Last Reviewed 24-Oct-2026)
Microsoft AB-100 Practice Test 22 - 10 Questions (Last Reviewed 23-Oct-2026)

Pass Your Microsoft AB-100 Exam Easy!

Question 190

Explanation

Correct Answer

Incorrect Answers

Question 191

Explanation

Correct Answer

Incorrect Answers

Question 192

Explanation

Correct Answer

Incorrect Answers

Question 193

Explanation

Correct Answer

Incorrect Answers

Question 194

Explanation

Correct Answer

Incorrect Answers

Question 195

Explanation

Correct Answer

Incorrect Answers

Question 196

Explanation

Correct Answer

Incorrect Answers

Question 197

Explanation

Correct Answer

Incorrect Answers

Question 198

Explanation

Correct Answer

Incorrect Answers

Question 199

Explanation

Correct Answer

Incorrect Answers

Question 200

Explanation

Correct Answers

Incorrect Answers

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