Microsoft AB-100 (Agentic AI Business Solutions Architect) Practice Test 67

Question 660

Operations has not established the cause of this defect: Half-complete transactions leave operational state inconsistent. The associated setting is: An agent creates an order but reservation of inventory fails. Which inspection should precede remediation?

  1. Verify virtual entity enablement and row-level access
  2. Compare cross-service commit histories and compensation events
  3. Compare advertised tools with backend permissions
  4. Trace agent plan for extra tool invocations

Explanation

Correct Answer

B:

Where half-complete transactions leave operational state inconsistent, the useful evidence comes from comparing cross-service commit histories and compensation events; its focus matches the architectural mechanism that compensating operations reconcile earlier successful steps in the process where an agent creates an order but reservation of inventory fails.

Incorrect Answers

A:

Finance users need current operational data in contextual help; its matching diagnostic check is reasonable for a static knowledge upload gives stale purchase-order states; the present investigation needs evidence for the different failure where half-complete transactions leave operational state inconsistent.

C:

Each agent builds a different unaudited HTTP wrapper; examining that other path does not determine whether compensating operations reconcile earlier successful steps, which is what distinguishes the issue in the stated use case.

D:

The proposed probe concerns an employee requests a standard equipment return initiation, where single-purpose tools and typed parameters bound the task; this incident instead requires evidence about compensating operations reconcile earlier successful steps and the failure that half-complete transactions leave operational state inconsistent.

Question 661

The acceptance board asks how to verify a corrected design. The setting is: Business applications depend on the agent endpoint continuously. Last time, a model upgrade breaks downstream response formatting. Which scenario should the team execute?

  1. Shift limited traffic to a candidate and monitor regressions
  2. Deploy a test agent and ensure discovery and ownership tracking
  3. Promote to staging and confirm correct connection targets
  4. Attempt inference from an unauthorized identity or network

Explanation

Correct Answer

A:

Shifting limited traffic to a candidate and monitor regressions challenges the path where canary traffic and versioned rollback reduce release risk for the business setting business applications depend on the agent endpoint continuously; the specific defect to exclude is that a model upgrade breaks downstream response formatting.

Incorrect Answers

B:

This check exercises a different system behavior: inventory metadata records owner, risk tier, tools and lifecycle state; a pass there cannot establish that a model upgrade breaks downstream response formatting no longer occurs in the required workflow.

C:

This check exercises a different system behavior: environment configuration separates deployable logic from endpoint bindings; a pass there cannot establish that a model upgrade breaks downstream response formatting no longer occurs in the required workflow.

D:

Although the exercise might catch untrusted clients submit requests and exfiltrate outputs, it would not directly validate the mechanism that canary traffic and versioned rollback reduce release risk in the target implementation.

Question 662

A release must prove that unvetted labeled data poisons a retrained classifier will not recur. The application is used in this situation: Business analysts can submit examples but not change production behavior. Which acceptance test most directly checks that failure mode?

  1. Test extraction and secret-disclosure prompts with synthetic fixtures
  2. Reconstruct one released model's source and threshold decisions
  3. Submit a malicious mislabeled training record and inspect review
  4. Promote a candidate in staging and inspect generated audit entries

Explanation

Correct Answer

C:

The scenario is that business analysts can submit examples but not change production behavior; the defect to rule out is that unvetted labeled data poisons a retrained classifier; testing via submitting a malicious mislabeled training record and inspecting review challenges the design dependency where separated data-curation and release permissions preserve integrity.

Incorrect Answers

A:

The suggested test serves an internal model sometimes returns confidential text verbatim, in which a client can request arbitrary portions of training examples; the present scenario instead concerns the reported problem where unvetted labeled data poisons a retrained classifier.

B:

Reconstructing one released model's source and threshold decisions verifies an adjacent objective where data scientists cannot explain why a newer classifier behaves differently; it leaves untested the causal path in which separated data-curation and release permissions preserve integrity for this case.

D:

In the alternate setting where a model is replaced without a traceable change ticket, the test exercises deployment audit links actor, approved artifact and release time; even a pass would not show that unvetted labeled data poisons a retrained classifier has been prevented.

Question 663

An operational risk could arise because a malicious source document might grant fictitious permissions to an agent. The AI system supports the business need that external knowledge contains embedded role-change commands. Which guardrail is designed for this threat?

  1. Forbid untrusted content from granting new capabilities
  2. Restore the previous prompt if critical intent selection regresses
  3. Do not allow model confidence prose to override missing evidence
  4. Block publishing when required evidence cannot be recovered

Explanation

Correct Answer

A:

Because a malicious source document might grant fictitious permissions to an agent, a guardrail must constrain the actual system behavior: trust hierarchy separates lower-trust content from agent instructions; protecting a response alone cannot enforce that boundary.

Incorrect Answers

B:

The proposed control targets a prompt revision might break critical intent handling without a rollback, while the given risk scenario is a malicious source document might grant fictitious permissions to an agent; these exposures arise in different operational contexts.

C:

The alternate guard addresses confident model language might override an explicit evidence requirement for policy answers must be tied to a current source; the present threat, a malicious source document might grant fictitious permissions to an agent, needs a different enforcement boundary.

D:

For the use case where an answer contradicts the handbook despite high model fluency, the control addresses the threat that the team might publish a model that never retrieves the evidence it requires; the specified agent must be protected against a malicious source document might grant fictitious permissions to an agent.

Question 664

The rollout supports this workflow: Staff reject suggestions because approval controls are unclear. Leaders want a metric that responds to the reported shortcoming that users cannot tell whether a value was generated or saved. Which measure best reflects success?

  1. Core form completion rate during AI service incidents
  2. Agent recommendations with recorded outcome and owner
  3. Correct acceptance and correction rate in assisted forms
  4. Case-workspace completion without navigation detours

Explanation

Correct Answer

C:

The observed shortcoming is that users cannot tell whether a value was generated or saved; measuring correct acceptance and correction rate in assisted forms gives business owners a directly relevant indicator for the process where staff reject suggestions because approval controls are unclear.

Incorrect Answers

A:

The proposed metric concerns field staff must keep working during a model outage, whereas the failure in this case is that users cannot tell whether a value was generated or saved; the indicator needed here is correct acceptance and correction rate in assisted forms.

B:

Operations teams review multiple proposed next actions daily; its intended measurement is agent recommendations with recorded outcome and owner; substituting that result here would not capture whether users cannot tell whether a value was generated or saved has been corrected.

D:

Staff need a contextual workspace assembled around one customer case; its intended measurement is case-workspace completion without navigation detours; substituting that result here would not capture whether users cannot tell whether a value was generated or saved has been corrected.

Question 665

The acceptance board asks how to verify a corrected design. The setting is: Agents regularly exceed an interactive response-time objective. Last time, engineers blame generation although the backend is slow. Which scenario should the team execute?

  1. Inject a tool timeout and locate the failing span
  2. Attempt inference from an unauthorized identity or network
  3. Delay an approver past the defined execution deadline
  4. Inject delay into one dependency and locate the timing change

Explanation

Correct Answer

D:

Engineers blame generation although the backend is slow; a release test must establish that this no longer occurs; injecting delay into one dependency and locating the timing change measures the specific behavior for which per-span latency isolates model work, network waits and tool execution.

Incorrect Answers

A:

The tested behavior is that agent trace spans provide component-level execution evidence through connected telemetry; the necessary evidence here concerns per-span latency isolates model work, network waits and tool execution and the previously observed engineers blame generation although the backend is slow.

B:

In a solution where a sensitive model endpoint is reachable using a broadly shared key, attempting inference from an unauthorized identity or network could catch the defect that untrusted clients submit requests and exfiltrate outputs; the current release gate must challenge the failure where engineers blame generation although the backend is slow.

C:

The suggested test serves an autonomous agent relies on slow approvals and delayed tool results, in which timeout creates an abandoned workflow with no owner; the present scenario instead concerns the reported problem where engineers blame generation although the backend is slow.

Question 666

Release reviewers will not accept a generic success message. The business scenario is that employees express several intents and need interactive clarification, and the historical failure is: A rigid menu fails whenever a request contains two intents. Which test is appropriately targeted?

  1. Call a read-only tool with valid and invalid identity contexts
  2. Submit paraphrases and multi-intent requests and verify correct tool selection
  3. Verify task handoff and result exchange across two supported runtimes
  4. Test adversarial lookalike cases not present in prompt examples
  5. Test mixed-domain requests and validate specialist selection

Explanation

Correct Answer

B:

Submitting paraphrases and multi-intent requests and verifying correct tool selection is aligned with the failure that a rigid menu fails whenever a request contains two intents; its causal relevance is that generative orchestration interprets intent and can combine authorized tools and knowledge for the situation where employees express several intents and need interactive clarification.

Incorrect Answers

A:

In a solution where a Copilot Studio agent needs a reusable interface to governed operations, calling a read-only tool with valid and invalid identity contexts could catch the defect that every agent reimplements the same proprietary tool signatures; the current release gate must challenge the failure where a rigid menu fails whenever a request contains two intents.

C:

The suggested test serves independently hosted agents must exchange tasks and results, in which cross-team agents rely on hard-coded, vendor-specific chat formats; the present scenario instead concerns the reported problem where a rigid menu fails whenever a request contains two intents.

D:

In a solution where the classifier confuses complaints with refund requests, testing adversarial lookalike cases not present in prompt examples could catch the defect that a generic instruction cannot express subtle category boundaries; the current release gate must challenge the failure where a rigid menu fails whenever a request contains two intents.

E:

Testing mixed-domain requests and validating specialist selection verifies an adjacent objective where a request spans procurement, finance and policy expertise; it leaves untested the causal path in which generative orchestration interprets intent and can combine authorized tools and knowledge for this case.

Question 667

One enterprise rollout has 2 separately testable requirements. (1) Business use: An answer contradicts the handbook despite high model fluency. Known defect: Engineers tune prompts without checking whether evidence was retrieved. (2) Business use: Average accuracy hides failures on complex finance questions. Known defect: Lower-cost routing degrades a business-critical subgroup. Which 2 diagnostic procedures directly satisfy these stated cases? Choose TWO.

  1. Compare quality by task complexity and selected model
  2. Compare retrieved document IDs for multiple test principals
  3. Review retrieved snippets alongside final claims
  4. Inspect cache keys and authorization on reused entries

Explanation

Correct Answers

A:

Comparing quality by task complexity and selected model can reveal why lower-cost routing degrades a business-critical subgroup when average accuracy hides failures on complex finance questions; that finding is needed alongside the independent diagnosis for an answer contradicts the handbook despite high model fluency.

C:

Reviewing retrieved snippets alongside final claims can reveal why engineers tune prompts without checking whether evidence was retrieved when an answer contradicts the handbook despite high model fluency; that finding is needed alongside the independent diagnosis for average accuracy hides failures on complex finance questions.

Incorrect Answers

B:

This diagnosis is associated with the different failure where agent cites a private HR document to unauthorized staff in a workflow where users have different rights to the same knowledge corpus; it does not directly solve the listed conditions: engineers tune prompts without checking whether evidence was retrieved; lower-cost routing degrades a business-critical subgroup.

D:

This diagnosis is associated with the different failure where one user's cached confidential answer appears for another user in a workflow where shared result caching makes answers fast across roles; it does not directly solve the listed conditions: engineers tune prompts without checking whether evidence was retrieved; lower-cost routing degrades a business-critical subgroup.

Question 668

The architecture board has evidence that AI recommends an action but downstream case data is incomplete. The affected situation is: A won opportunity should create a properly owned onboarding case. What change most directly establishes the missing behavior?

  1. Preserve customer communication approvals across solution releases
  2. Test Sales-to-Service customer context and case handoff
  3. Test audit completeness for AI-assisted customer concessions
  4. Validate finance and operations virtual-entity mappings after change

Explanation

Correct Answer

B:

The architecture must reconcile two facts: a won opportunity should create a properly owned onboarding case, yet AI recommends an action but downstream case data is incomplete; the selected pattern does so through the mechanism that end-to-end tests verify identity, mapping and assignment across apps.

Incorrect Answers

A:

The technical behavior of this alternative is that business approval checks remain independent of agent version; the user case being assessed is a won opportunity should create a properly owned onboarding case with a different failure: AI recommends an action but downstream case data is incomplete.

C:

Transaction-path tests confirm recommendation, review and posting records; its main purpose is service recommends a price adjustment requiring finance approval, not correction of the reported defect where AI recommends an action but downstream case data is incomplete in the target process.

D:

For AI assistance reads operational records through Dataverse, this other approach relies on the fact that integration tests protect virtual-entity schema and source semantics; the issue here is AI recommends an action but downstream case data is incomplete; that mechanism does not establish the required end-to-end tests verify identity, mapping and assignment across apps.

Question 669

A business stakeholder describes this operating situation: The same agent must move from development to test and production. Testing instead shows that manual recreation causes mismatched topics and actions. What design approach is most appropriate?

  1. Use separate service principals and least-privilege access per environment
  2. Package Copilot Studio agent components in a Power Platform solution
  3. Run connector contract tests before every promotion
  4. Validate agent publishing and channel configuration after promotion

Explanation

Correct Answer

B:

The same agent must move from development to test and production; its failing behavior is that manual recreation causes mismatched topics and actions; applying the underlying design principle—solution-based ALM transports supported components under controlled releases—connects the requirement to an implementable control.

Incorrect Answers

A:

The technical behavior of this alternative is that environment-scoped identities isolate access and simplify incident containment; the user case being assessed is the same agent must move from development to test and production with a different failure: manual recreation causes mismatched topics and actions.

C:

When an external API changes a required field used by an agent, versioned API contracts and consumer checks detect breaking changes; in the stated workflow, however, manual recreation causes mismatched topics and actions; the relevant architectural behavior is that solution-based ALM transports supported components under controlled releases.

D:

The alternative addresses deployment checks stop at package import success by a design in which post-import validation includes publishing and channel behavior; it does not supply the missing relationship for the same agent must move from development to test and production, where manual recreation causes mismatched topics and actions.

Question 670

A policy review covers an agent for which the requirement is common and product coverage is adequate. Its most relevant exposure is that a managed service might fail a required data-processing or exit condition. What safeguard should be required before release?

  1. Confirm tenant policy and data access before enabling adoption
  2. Review tenancy, data terms, exit and support commitments
  3. Disable autonomous regulated decisions under low evidence
  4. Narrow allowed tool capabilities and verify server-side authorization

Explanation

Correct Answer

B:

The requirement is common and product coverage is adequate; its risk scenario is that a managed service might fail a required data-processing or exit condition; the safeguard fits because managed functionality shifts routine platform operations to the service and the restriction addresses that specific exposure.

Incorrect Answers

A:

Users need ordinary meeting summaries and follow-up drafts; its guardrail addresses the possibility that a standard agent might violate tenant access policy when activated for many users; the present architecture must constrain the different threat that a managed service might fail a required data-processing or exit condition.

C:

This restriction is associated with the system cannot always find evidence for regulated guidance and addresses the exposure that a low-evidence agent might autonomously decide a regulated outcome; it does not specifically protect the target workflow against a managed service might fail a required data-processing or exit condition.

D:

The safeguard for a Copilot Studio agent needs a reusable interface to governed operations deals with a different threat: a shared MCP server might honor an unauthorized tool invocation; even after it is enforced, the exposure that a managed service might fail a required data-processing or exit condition remains possible in the target workflow where managed functionality shifts routine platform operations to the service.

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