Microsoft AB-100 (Agentic AI Business Solutions Architect) Practice Test 43

Question 420

A design adds autonomous behavior to a process where proprietary records are copied to an open test workspace. This creates a documented risk: Restricted fine-tuning documents might be readable by another project. Which enforced control is the best fit?

  1. Require data-use approval and encrypted governed storage
  2. Require finance process-owner approval for production changes
  3. Do not share private cached outputs across principals
  4. Reject cached results that omit original security context

Explanation

Correct Answer

A:

Because restricted fine-tuning documents might be readable by another project, a guardrail must constrain the actual system behavior: segregated datasets and access controls protect sensitive data; protecting a response alone cannot enforce that boundary.

Incorrect Answers

B:

The safeguard for invoice posting agents affect critical close processes deals with a different threat: a finance Copilot prompt might change during a protected close window; even after it is enforced, the exposure that restricted fine-tuning documents might be readable by another project remains possible in the target workflow where segregated datasets and access controls protect sensitive data.

C:

Where shared result caching makes answers fast across roles, the guard can address a user might receive another principal's cached protected answer; the exposure that restricted fine-tuning documents might be readable by another project remains relevant to the present assessment.

D:

That protection is tied to the mechanism that retrieval applies effective source entitlements to candidate documents; it does not prevent the distinct risk where restricted fine-tuning documents might be readable by another project in the system under review.

Question 421

An enterprise team is making a platform design decision. Business context: Agent handles threshold-based return decisions. Present failure: Human testers overlook values just above approval limits. Which approach resolves the mismatch?

  1. Tune model routing based on segmented evaluation outcomes
  2. Test multi-turn context retention and clarification
  3. Review AI-generated test oracles with independent experts
  4. Generate boundary cases from approved acceptance requirements

Explanation

Correct Answer

D:

The architecture must reconcile two facts: agent handles threshold-based return decisions, yet human testers overlook values just above approval limits; the selected pattern does so through the mechanism that assisted test generation explores stated boundary conditions.

Incorrect Answers

A:

Stratified evaluation guides thresholds for difficult requests; its main purpose is average accuracy hides failures on complex finance questions, not correction of the reported defect where human testers overlook values just above approval limits in the target process.

B:

For agent must remember order context while asking for missing details, this other approach relies on the fact that multi-turn tests assess state tracking, entity separation and follow-up questions; the issue here is human testers overlook values just above approval limits; that mechanism does not establish the required assisted test generation explores stated boundary conditions.

C:

When Copilot writes a test expecting a plausible but wrong answer, independent acceptance references prevent self-confirming evaluation; in the stated workflow, however, human testers overlook values just above approval limits; the relevant architectural behavior is that assisted test generation explores stated boundary conditions.

Question 422

A policy review covers an agent for which a bot only needs to check warranty and prepare a draft. Its most relevant exposure is that a drafting tool might send an external supplier message without authority. What safeguard should be required before release?

  1. Keep business-critical formulas out of free-form generation
  2. Reject unvalidated model output before ticket creation
  3. Separate draft generation from external communication authority
  4. Require human sign-off when regulated language is modified

Explanation

Correct Answer

C:

When a drafting tool might send an external supplier message without authority, the use case a bot only needs to check warranty and prepare a draft requires a guardrail aligned with the behavior that limited intent and tool scope prevent unnecessary actions.

Incorrect Answers

A:

That protection is tied to the mechanism that deterministic application code computes numeric outputs from authorized data; it does not prevent the distinct risk where a drafting tool might send an external supplier message without authority in the system under review.

B:

This restriction is associated with a summarizer sends output into a ticket ingestion service and addresses the exposure that a malformed structured summary might create corrupt service tickets; it does not specifically protect the target workflow against a drafting tool might send an external supplier message without authority.

D:

The alternate guard addresses a model might alter mandatory legal wording in customer correspondence for compliance letters must preserve approved language and disclaimers; the present threat, a drafting tool might send an external supplier message without authority, needs a different enforcement boundary.

Question 423

An enterprise team is making a platform design decision. Business context: Finance users need current operational data in contextual help. Present failure: A static knowledge upload gives stale purchase-order states. Which approach resolves the mismatch?

  1. Use finance and supply-chain in-app agent knowledge through Dataverse virtual entities
  2. Use an AI hub-governed reusable prompt for Power Platform makers
  3. Apply security and privacy controls to app-held AI context
  4. Add a plugin action to a supported declarative agent
  5. Apply experience optimization to human-in-the-loop decisions

Explanation

Correct Answer

A:

Without the intended behavior, a static knowledge upload gives stale purchase-order states; the proposed architecture establishes the specific design principle that authorized virtual entities can surface current finance and operations data to supported agents for the process where finance users need current operational data in contextual help.

Incorrect Answers

B:

The technical behavior of this alternative is that centralized reusable AI capabilities promote consistency and controlled updates; the user case being assessed is finance users need current operational data in contextual help with a different failure: a static knowledge upload gives stale purchase-order states.

C:

This design would make sense if a canvas app passes customer records to a model action, particularly where hidden fields from the UI appear in the prompt payload; the requested architecture instead needs the dependency that authorized virtual entities can surface current finance and operations data to supported agents because a static knowledge upload gives stale purchase-order states.

D:

This design would make sense if a Teams user should query and update a supported external service, particularly where read-only knowledge alone cannot create a service request; the requested architecture instead needs the dependency that authorized virtual entities can surface current finance and operations data to supported agents because a static knowledge upload gives stale purchase-order states.

E:

For another type of architecture, clear review interfaces surface provenance and editable proposals is useful; this design must instead support finance users need current operational data in contextual help despite the problem where a static knowledge upload gives stale purchase-order states.

Question 424

The fault affects an AI-powered workflow in which two customers store similarly named support incidents. Specifically, assistant retrieves another tenant's incident detail. Which diagnostic investigation best distinguishes the relevant failure path?

  1. Inspect source identifiers and ingestion timestamps in index records
  2. Inspect third-party manifests and requested access scopes
  3. Inspect vector-query filters and tenant IDs of returned hits
  4. Audit tokens and connection bindings for each stage

Explanation

Correct Answer

C:

The reported fault is that assistant retrieves another tenant's incident detail in a system where two customers store similarly named support incidents; evidence from inspecting vector-query filters and tenant IDs of returned hits can show whether partitioning and authorized filters constrain relevant candidates.

Incorrect Answers

A:

Inspecting source identifiers and ingestion timestamps in index records explores a separate connection, where content lineage records update, transformation and index revision; the diagnosed requirement is two customers store similarly named support incidents, and the actual defect is assistant retrieves another tenant's incident detail.

B:

Inspecting third-party manifests and requested access scopes explores a separate connection, where procurement and security review evaluate permissions and data flows; the diagnosed requirement is two customers store similarly named support incidents, and the actual defect is assistant retrieves another tenant's incident detail.

D:

That probe evaluates environment-scoped identities isolate access and simplify incident containment in a workflow where test automation must not reach live customer records; the reported failure concerns the path for which partitioning and authorized filters constrain relevant candidates instead.

Question 425

In a business situation where copilot, Dataverse, Dynamics and custom APIs share one process, a previous build showed this defect: Failures cannot be attributed to any single integration owner. Which reproducible test belongs in the release gate?

  1. Inject a dependency timeout and verify accountable error handling
  2. Send invalid IDs, oversized values and out-of-range quantities
  3. Run a pilot with a comparable holdout group
  4. Emit duplicate and delayed events and verify idempotent processing

Explanation

Correct Answer

A:

The historical failure is that failures cannot be attributed to any single integration owner; for the setting where copilot, Dataverse, Dynamics and custom APIs share one process, injecting a dependency timeout and verifying accountable error handling verifies the dependency where contracts and ownership boundaries identify responsibility for each state change.

Incorrect Answers

B:

In the alternate setting where the model submits malformed identifiers to an operational API, the test exercises typed inputs and server validation reject malformed or out-of-scope calls; even a pass would not show that failures cannot be attributed to any single integration owner has been prevented.

C:

For management attributes a sales lift to Copilot without a comparison, the alternate test is about whether matched comparisons separate solution impact from unrelated shifts; it cannot prove the outcome needed after the reported regression where failures cannot be attributed to any single integration owner.

D:

This check exercises a different system behavior: business events and durable orchestration propagate changes asynchronously; a pass there cannot establish that failures cannot be attributed to any single integration owner no longer occurs in the required workflow.

Question 426

A policy review covers an agent for which an agent will evolve continuously after its initial project delivery. Its most relevant exposure is that production agents could launch without an incident owner, runbook or operations funding. What safeguard should be required before release?

  1. Validate residency, safety and output contract on fallback models
  2. Gate launch on ownership, runbooks and support budgets
  3. Require idempotency at cross-service mutation boundaries
  4. Require exception ownership and policy-version review

Explanation

Correct Answer

B:

Because production agents could launch without an incident owner, runbook or operations funding, a guardrail must constrain the actual system behavior: long-lived product ownership funds ongoing evaluation and operations; protecting a response alone cannot enforce that boundary.

Incorrect Answers

A:

Where peak load can saturate a primary model endpoint, the guard can address fallback routing might send sensitive workloads to a disallowed model location; the exposure that production agents could launch without an incident owner, runbook or operations funding remains relevant to the present assessment.

C:

Another process could need this guardrail because network retries could create duplicated writes across independent systems; the design being assessed supports an agent will evolve continuously after its initial project delivery and has the explicit exposure that production agents could launch without an incident owner, runbook or operations funding.

D:

That protection is tied to the mechanism that a CoE publishes patterns, reusable controls, standards and escalation paths; it does not prevent the distinct risk where production agents could launch without an incident owner, runbook or operations funding in the system under review.

Question 427

The fault affects an AI-powered workflow in which business units create agents with no central visibility. Specifically, security cannot identify which agents use protected connectors. Which diagnostic investigation best distinguishes the relevant failure path?

  1. Reconcile tenant deployments with the approved registry
  2. Verify spans and errors arrive in the expected telemetry resource
  3. Match release manifest to transformation and model checksums
  4. Audit quarantined document reasons and promotion events
  5. Inspect artifact signatures and output schema tests
  6. Trace token acquisition and authorization results by environment

Explanation

Correct Answer

A:

For the case where business units create agents with no central visibility, the diagnostic path must explain the failure that security cannot identify which agents use protected connectors; reconciling tenant deployments with the approved registry targets the mechanism by which inventory metadata records owner, risk tier, tools and lifecycle state.

Incorrect Answers

B:

For the failure that the on-call team cannot explain failed tool calls, verifying spans and errors arrive in the expected telemetry resource is relevant; this question instead reports that security cannot identify which agents use protected connectors, so different evidence is needed.

C:

The evidence this method produces applies to a workflow where a preprocessing release changes category labels incorrectly; it cannot by itself establish why security cannot identify which agents use protected connectors in the current scenario.

D:

In the process where new documents enter a production source automatically, the diagnostic procedure could explain unreviewed confidential drafts appear in agent answers; this particular incident instead turns on whether inventory metadata records owner, risk tier, tools and lifecycle state and why security cannot identify which agents use protected connectors.

E:

The evidence this method produces applies to a workflow where a retraining release changes the tokenizer and label order; it cannot by itself establish why security cannot identify which agents use protected connectors in the current scenario.

F:

For the failure that users receive permission errors after successful package import, tracing token acquisition and authorization results by environment is relevant; this question instead reports that security cannot identify which agents use protected connectors, so different evidence is needed.

Question 428

A product owner must measure whether an AI-enabled change actually improves this process: An internal helper produces attractive but unsupported policy answers. The earlier problem was: Satisfaction scores hide factual mistakes. Which outcome indicator is most relevant?

  1. Unauthorized retrievals completed in red-team cases
  2. Median turns to resolution and escalation rate by channel
  3. Deployments with complete monitored trace coverage
  4. Material claim accuracy with valid citations

Explanation

Correct Answer

D:

For the business use where an internal helper produces attractive but unsupported policy answers, the result to quantify is material claim accuracy with valid citations; it connects measurement to the mechanism where expert-checked evaluation examples quantify evidence fidelity, not merely to agent traffic.

Incorrect Answers

A:

The alternative tracks unauthorized retrievals completed in red-team cases, arising from the mechanism where authorization evaluations verify server-side deny behavior; its movement alone says little about the failure that satisfaction scores hide factual mistakes.

B:

The premise of this indicator is channel-specific analytics reveal where interactions fail; the architecture being evaluated instead serves an internal helper produces attractive but unsupported policy answers, and its success is measured by material claim accuracy with valid citations.

C:

The proposed metric concerns a newly deployed Foundry agent has no detailed diagnostics, whereas the failure in this case is that satisfaction scores hide factual mistakes; the indicator needed here is material claim accuracy with valid citations.

Question 429

An incident review has one confirmed symptom: A generic instruction cannot express subtle category boundaries. For the use case where the classifier confuses complaints with refund requests, which targeted diagnostic check would reveal the most pertinent evidence?

  1. Analyze confusion matrix at neighboring categories
  2. Inspect citations and human review notes for unsupported conclusions
  3. Inspect approvals against the approved decision table
  4. Compare errors for historical and newly effective rules

Explanation

Correct Answer

A:

The investigation directly tests the causal connection: representative examples teach decision boundaries without changing underlying facts; it collects evidence for a workflow where the classifier confuses complaints with refund requests but the outcome is instead that a generic instruction cannot express subtle category boundaries.

Incorrect Answers

B:

In the process where analysts compare unstructured reports but must retain human interpretation, the diagnostic procedure could explain narratives present speculation as if it were validated insight; this particular incident instead turns on whether representative examples teach decision boundaries without changing underlying facts and why a generic instruction cannot express subtle category boundaries.

C:

The proposed probe concerns the task follows fixed approval branches with no interpretation needed, where explicit conditions execute known business rules predictably and leave an auditable branch history; this incident instead requires evidence about representative examples teach decision boundaries without changing underlying facts and the failure that a generic instruction cannot express subtle category boundaries.

D:

For the failure that a trained model remembers obsolete guidance after source changes, comparing errors for historical and newly effective rules is relevant; this question instead reports that a generic instruction cannot express subtle category boundaries, so different evidence is needed.

Question 430

A solution architecture is being reviewed for this business situation: A sensitive model endpoint is reachable using a broadly shared key. The observed defect is that untrusted clients submit requests and exfiltrate outputs. Which architectural decision addresses the underlying requirement?

  1. Use identity-aware caching for grounded agent responses
  2. Require independent backend authorization for generated tool calls
  3. Isolate model evaluation and fine-tuning data by classification
  4. Protect inference endpoints with network and identity boundaries

Explanation

Correct Answer

D:

The defect is that untrusted clients submit requests and exfiltrate outputs; the expected business case is that a sensitive model endpoint is reachable using a broadly shared key; the design fits because identity-based access and supported network isolation reduce exposure.

Incorrect Answers

A:

The alternative addresses one user's cached confidential answer appears for another user by a design in which cache keys and access validation prevent security-context leakage; it does not supply the missing relationship for a sensitive model endpoint is reachable using a broadly shared key, where untrusted clients submit requests and exfiltrate outputs.

B:

Server-side access checks enforce real actor entitlements; its main purpose is prompt wording persuades the agent to submit a privileged operation, not correction of the reported defect where untrusted clients submit requests and exfiltrate outputs in the target process.

C:

For proprietary records are copied to an open test workspace, this other approach relies on the fact that segregated datasets and access controls protect sensitive data; the issue here is untrusted clients submit requests and exfiltrate outputs; that mechanism does not establish the required identity-based access and supported network isolation reduce exposure.

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