Microsoft AB-100 Practice Test 61

Question 600

A Canvas app stores a private escalation note in a client-visible variable and passes it to an agent, although the user is not entitled to see the note. What architectural correction is required?

Hide the note control using a conditional Visible property only.
Fetch authorized escalation context server-side and return only fields permitted for the current app user.
Encode the note with Base64 before storing it in the app variable.
Ask the agent to ignore restricted notes if the user requests them.

Explanation

Correct Answer

Client variables can be inspected and are not a secure data-access boundary for restricted notes.

Incorrect Answers

A hidden control does not establish server-side permission filtering or prevent data extraction.

Encoding does not provide authorization or confidentiality to someone who can read client state.

The restricted content should not have reached the user-accessible app context in the first place.

Question 601

A generated Canvas page shows a list of expense requests with Approve and Reject actions. After a phone-layout change, Reject disappears but Approve remains. What is the correct acceptance gate?

Treat successful page rendering as proof that all business actions are present.
Allow an agent to guess whether missing Reject meant a silent refusal.
Require responsive usability tests that verify both actions, permission checks and state transitions on mobile and desktop.
Remove Reject from the desktop layout too so both platforms behave consistently.

Explanation

Correct Answer

Layout optimizations must not silently remove the authorized user’s ability to refuse an expense.

Incorrect Answers

Visual load success does not exercise or verify each mandatory action and its result.

A generated interpretation cannot substitute for an explicit recorded decision by the reviewer.

Consistent removal still violates the required business decision workflow and role controls.

Question 602

A sales application calls a hosted AI service to summarize leads. Its core CRM create/update actions must still work if inference is unavailable. What is the appropriate reliability design?

Make every CRM Save wait until the summary model recovers.
Accept an empty successful summary response as proof the model evaluated the lead.
Increase the model temperature when a network timeout occurs.
Isolate summarization behind a bounded timeout and graceful fallback so CRM transactions remain independent.

Explanation

Correct Answer

An optional AI read operation must not become a single point of failure for core customer-record writes.

Incorrect Answers

Unbounded inference dependency would turn a secondary capability outage into a full transactional outage.

A fabricated success obscures the service failure and may mislead the sales team.

Generation sampling controls cannot repair unavailable network connectivity or service dependency.

Question 603

A Power Platform application adds an AI action used by 800 agents during a 10-minute morning peak. Pilot latency measurements come from two users. Which performance-sizing input is necessary?

Load-test expected concurrency, API quotas and downstream throttling under the real burst pattern.
Use only average daily calls because peak traffic is negligible to capacity.
Increase every action timeout to 30 minutes instead of testing throughput.
Assume environment licensing automatically reserves unlimited model throughput.

Explanation

Correct Answer

Two isolated response times do not reveal queueing behavior or rate-limit failures at production concurrency.

Incorrect Answers

Average volume can conceal short bursts that exceed connector or inference limits.

Long waits do not add capacity and can amplify resource exhaustion.

Licensing and actual service concurrency limits are separate capacity considerations.

Question 604

An invoice system performs a model call even when the incoming payload is byte-identical to a previously validated invoice version. Which optimization preserves correctness?

Always switch to the cheapest model without evaluating the extraction-error impact.
Use a versioned, permission-scoped cache for safe read-only extraction and rerun when source bytes or policy change.
Cache all payment approvals globally and replay them for any customer with a similar invoice.
Disable validation on cache hits to reduce runtime cost.

Explanation

Correct Answer

Identical approved inputs can reuse validated extraction results only while freshness and authorization constraints hold.

Incorrect Answers

Lower call price can increase costly manual corrections and downstream accounting risk.

Approval is transactional and identity-bound, not a reusable inference result.

Cache reuse still requires verifying record identity, source version and eligible caller scope.

Question 605

A customer-onboarding AI check can take two minutes; the Canvas app action normally times out before completion. What architecture reduces client failures?

Rerun the entire onboarding job after each client timeout with a new ID.
Have the app claim completion once the queue accepts a payload, before the job is processed.
Queue the work with a durable job ID and let the app poll or receive an authorized completion notification.
Increase model randomness to speed up the background workflow.

Explanation

Correct Answer

Long processing should not depend on one synchronous mobile HTTP session remaining open.

Incorrect Answers

Uncoordinated retries can duplicate provisioning and lose prior completion state.

Queue acceptance is not the final business outcome; downstream state requires separate confirmation.

Temperature affects generation behavior and does not provide durable asynchronous job execution.

Question 606

A Well-Architected review finds the AI service produces appealing summaries, but it leaks staff notes through overprivileged connectors and has no failure alerts. Which work should precede release?

Improve the summary’s tone and launch without adjusting privileges.
Give the application global access to avoid connector errors during demonstrations.
Suppress telemetry to reduce storage cost, then wait for user complaints.
Enforce least-privilege data access, define monitored failure signals and exercise recovery paths.

Explanation

Correct Answer

Security and reliability are production quality attributes independent of whether generated text reads well.

Incorrect Answers

Friendly prose cannot mitigate source confidentiality exposure or loss of operational monitoring.

Broad credentials worsen the identified disclosure rather than fixing caller entitlement enforcement.

Without diagnostic signals the team cannot detect outages or prove the intended reliability targets.

Question 607

A repair center receives “reschedule ticket T-882 for Friday.” Its action contract has fixed intent types and typed ticket/date fields. What design best controls the request?

Extract the supported reschedule intent and validated ticket/date entities, then call the authorized action.
Ask a general generator to invent a repair date when a phrase seems plausible.
Use semantic search for repair manuals as the only processing step.
Immediately update every ticket mentioning Friday in its description.

Explanation

Correct Answer

A bounded utterance can be routed to a deterministic operation without allowing an unconstrained model to invent commands.

Incorrect Answers

A guessed date does not establish the requested calendar value or ticket owner authorization.

Retrieval can answer informational questions but cannot reliably execute a typed ticket update.

Keyword matching without a unique ticket identifier may mutate unrelated customer jobs.

Question 608

A customer asks, “Compare my last three electricity invoices, identify likely causes of the increase and draft a dispute.” No direct intent maps to this composite task. Which orchestration is justified?

Compare invoice file sizes instead of the posted amounts because file metadata is easier to access.
Plan a bounded sequence of authorized invoice reads, verifiable comparison and draft creation, reserving submission for explicit approval.
Route the whole message to the one-field invoice-status topic and declare the analysis complete.
Grant the model automatic authority to submit a formal dispute after summarizing charges.

Explanation

Correct Answer

The question combines retrieval, numeric analysis and document drafting; a structured plan can coordinate those separate steps.

Incorrect Answers

Attachment size is not evidence of electricity usage, billed rate or invoice change.

A status lookup does not calculate differences or create the requested evidence-backed draft.

A requested draft is not permission for an irreversible external submission.

Question 609

A benefits portal asks whether an employee is currently eligible for a statutory leave category. The authoritative rules engine exposes an eligibility Boolean plus rule version. Which conversational approach is preferable?

Calculate eligibility from the employee’s department display name alone.
Suppress the engine version and answer from an old policy summary.
Use the rule result as the decision and let the agent explain the rule with provenance without modifying its outcome.
Ask the LLM to vote against the rules engine whenever its sentiment differs.

Explanation

Correct Answer

A model can describe eligibility, but the verified deterministic engine controls the binding yes/no classification.

Incorrect Answers

Display names cannot prove all leave conditions, effective dates or service history.

Stale policy material can contradict the current version that determined eligibility.

Language sentiment cannot substitute for legal criteria evaluated in the authoritative service.

Question 610

An HR assistant receives “send the file to Sam.” Two authenticated employees named Sam are in different departments and only one is permitted to receive the performance report. What should routing do?

Pick the first Sam in the directory and send without confirmation.
Send to both Sams to improve delivery success.
Ask a language model to invent a recipient email address from prior message patterns.
Clarify the recipient identity and enforce file-level share authorization before any transfer.

Explanation

Correct Answer

Named-entity recognition alone does not disambiguate a recipient or authorize sharing personnel records.

Incorrect Answers

Directory rank has no relation to the user’s intended recipient or allowed sharing scope.

Distributing confidential material to two people increases exposure and violates least privilege.

Generated contact information is not a verified identity or permitted destination.

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

Microsoft AB-100 Practice Test 59 - 10 Questions (Last Reviewed 12-Nov-2026)
Microsoft AB-100 Practice Test 60 - 10 Questions (Last Reviewed 11-Nov-2026)
Microsoft AB-100 Practice Test 61 - 10 Questions (Last Reviewed 10-Nov-2026)
Microsoft AB-100 Practice Test 62 - 10 Questions (Last Reviewed 09-Nov-2026)
Microsoft AB-100 Practice Test 63 - 10 Questions (Last Reviewed 08-Nov-2026)

Pass Your Microsoft AB-100 Exam Easy!

Question 600

Explanation

Correct Answer

Incorrect Answers

Question 601

Explanation

Correct Answer

Incorrect Answers

Question 602

Explanation

Correct Answer

Incorrect Answers

Question 603

Explanation

Correct Answer

Incorrect Answers

Question 604

Explanation

Correct Answer

Incorrect Answers

Question 605

Explanation

Correct Answer

Incorrect Answers

Question 606

Explanation

Correct Answer

Incorrect Answers

Question 607

Explanation

Correct Answer

Incorrect Answers

Question 608

Explanation

Correct Answer

Incorrect Answers

Question 609

Explanation

Correct Answer

Incorrect Answers

Question 610

Explanation

Correct Answer

Incorrect Answers

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