Microsoft AB-100 (Agentic AI Business Solutions Architect) Practice Test 28

Question 270

A planned AI solution must support this situation: Two customers store similarly named support incidents. Its prototype fails because assistant retrieves another tenant's incident detail. Which design best fits the constraint?

  1. Define policies for third-party agent onboarding
  2. Use separate service principals and least-privilege access per environment
  3. Use identity-aware caching for grounded agent responses
  4. Track grounding-source updates and index rebuild provenance
  5. Enforce tenant and entity boundaries before vector search results are used

Explanation

Correct Answer

E:

Two customers store similarly named support incidents; its failing behavior is that assistant retrieves another tenant's incident detail; applying the underlying design principle—partitioning and authorized filters constrain relevant candidates—connects the requirement to an implementable control.

Incorrect Answers

A:

In the case where employees install external agents with unknown data pathways, the alternate design depends on procurement and security review evaluate permissions and data flows; it does not correct the target failure, assistant retrieves another tenant's incident detail.

B:

This design would make sense if test automation must not reach live customer records, particularly where a shared connection identity grants sandbox agents production write access; the requested architecture instead needs the dependency that partitioning and authorized filters constrain relevant candidates because assistant retrieves another tenant's incident detail.

C:

Cache keys and access validation prevent security-context leakage; its main purpose is shared result caching makes answers fast across roles, not correction of the reported defect where assistant retrieves another tenant's incident detail in the target process.

D:

The technical behavior of this alternative is that content lineage records update, transformation and index revision; the user case being assessed is two customers store similarly named support incidents with a different failure: assistant retrieves another tenant's incident detail.

Question 271

Support engineers report this failure: Rate limiting causes intermittent failures. The business context is: A regional launch multiplies concurrent agent requests. What evidence should be collected to isolate the cause?

  1. Correlate request rates with throttles and queue depth
  2. Compare quality by task complexity and selected model
  3. Compare operation keys with final business transaction counts
  4. Compare abandonment and handoff reasons by channel

Explanation

Correct Answer

A:

Rate limiting causes intermittent failures; to attribute the defect rather than guess at a fix, correlating request rates with throttles and queue depth examines the dependency where capacity metrics and throttling errors show saturation risk.

Incorrect Answers

B:

The proposed inspection is associated with average accuracy hides failures on complex finance questions, where stratified evaluation guides thresholds for difficult requests; it would not expose the failure mode that rate limiting causes intermittent failures with the required specificity.

C:

Several events refer to one business order at once; its matching diagnostic check is reasonable for parallel agents each submit a costly replacement; the present investigation needs evidence for the different failure where rate limiting causes intermittent failures.

D:

Combined metrics hide voice users trapped in fallback loops; examining that other path does not determine whether capacity metrics and throttling errors show saturation risk, which is what distinguishes the issue in the stated use case.

Question 272

A risk analysis highlights a threat in this workflow: Anonymous callers can ask generic questions but not change orders. Exposure: An anonymous caller might execute an account mutation after providing only a guessed ID. Which enforceable safeguard addresses it at the relevant boundary?

  1. Apply tenant policies and scoped content permissions
  2. Exclude sensitive attributes not needed by the receiving agent
  3. Fail closed on unresolved caller identity
  4. Block unclassified high-risk fields at the connector boundary

Explanation

Correct Answer

C:

Anonymous callers can ask generic questions but not change orders; its risk scenario is that an anonymous caller might execute an account mutation after providing only a guessed ID; the safeguard fits because authentication and authorization gates separate general help from account control and the restriction addresses that specific exposure.

Incorrect Answers

A:

For the use case where employees want curated assistance without owning a separate LLM runtime, the control addresses the threat that a declarative agent might be published beyond permitted tenant groups; the specified agent must be protected against an anonymous caller might execute an account mutation after providing only a guessed ID.

B:

Another process could need this guardrail because a case handoff could include personal attributes irrelevant to the receiving team's work; the design being assessed supports anonymous callers can ask generic questions but not change orders and has the explicit exposure that an anonymous caller might execute an account mutation after providing only a guessed ID.

D:

The proposed control targets an ERP connector might expose protected margin columns in its response payload, while the given risk scenario is an anonymous caller might execute an account mutation after providing only a guessed ID; these exposures arise in different operational contexts.

Question 273

Before production promotion, the team needs evidence for this setting: Downstream research agents run longer than the interactive request. A prior regression caused the following: Orchestrator returns a partial result as though the workflow finished. Which validation is most persuasive?

  1. Demonstrate a custom multi-step workflow and failure recovery
  2. Propose a concession above limit and confirm approval enforcement
  3. Delay a specialist and confirm timeout and retry behavior
  4. Execute the target action without leaving the work-order screen

Explanation

Correct Answer

C:

For the required workflow where downstream research agents run longer than the interactive request, the test needs to reproduce or prevent the failure where orchestrator returns a partial result as though the workflow finished; delaying a specialist and confirming timeout and retry behavior exercises the path where task IDs, status transitions, deadlines and durable handoff state support coordination.

Incorrect Answers

A:

This check exercises a different system behavior: custom-engine architecture controls runtime, model calls and integrations; a pass there cannot establish that orchestrator returns a partial result as though the workflow finished no longer occurs in the required workflow.

B:

For service, sales and finance jointly decide high-value concessions, the alternate test is about whether cross-app workflow routes the recommendation to the accountable approver; it cannot prove the outcome needed after the reported regression where orchestrator returns a partial result as though the workflow finished.

D:

In a solution where users must remain in a proprietary work-order application, executing the target action without leaving the work-order screen could catch the defect that requiring a separate chat portal increases context switching; the current release gate must challenge the failure where orchestrator returns a partial result as though the workflow finished.

Question 274

An enterprise team is making a platform design decision. Business context: An agent creates an order but reservation of inventory fails. Present failure: Half-complete transactions leave operational state inconsistent. Which approach resolves the mismatch?

  1. Use MCP tools for a governed reusable integration surface
  2. Use finance and supply-chain in-app agent knowledge through Dataverse virtual entities
  3. Specify stop conditions and budgets for long-running agents
  4. Use compensating workflows for partially completed side effects

Explanation

Correct Answer

D:

The architecture must reconcile two facts: an agent creates an order but reservation of inventory fails, yet half-complete transactions leave operational state inconsistent; the selected pattern does so through the mechanism that compensating operations reconcile earlier successful steps.

Incorrect Answers

A:

For another type of architecture, a supported MCP server exposes typed tool capabilities to clients is useful; this design must instead support an agent creates an order but reservation of inventory fails despite the problem where half-complete transactions leave operational state inconsistent.

B:

For finance users need current operational data in contextual help, this other approach relies on the fact that authorized virtual entities can surface current finance and operations data to supported agents; the issue here is half-complete transactions leave operational state inconsistent; that mechanism does not establish the required compensating operations reconcile earlier successful steps.

C:

There is a valid place for the mechanism that time, step and spend budgets terminate unproductive loops, namely a research agent repeatedly queries sources on difficult cases; it does not remove the specific defect that half-complete transactions leave operational state inconsistent from this case.

Question 275

A policy review covers an agent for which shipment exceptions should initiate investigation without a chat message. Its most relevant exposure is that spoofed or repeated business events might initiate unauthorized duplicate workflows. What safeguard should be required before release?

  1. Do not expose fields omitted by server authorization
  2. Cap tool depth and require evidence for consequential conclusions
  3. Label unsupported explanations as hypotheses or abstain
  4. Authenticate event sources and suppress duplicates

Explanation

Correct Answer

D:

In the scenario where shipment exceptions should initiate investigation without a chat message, the explicit risk is that spoofed or repeated business events might initiate unauthorized duplicate workflows; enforcement must follow the mechanism by which an event trigger starts the bounded agent workflow on state change.

Incorrect Answers

A:

This restriction is associated with staff need a contextual workspace assembled around one customer case and addresses the exposure that a generated page might reveal data fields hidden from its user by backend permissions; it does not specifically protect the target workflow against spoofed or repeated business events might initiate unauthorized duplicate workflows.

B:

Where procurement exceptions require comparing several policies and records, the guard can address a reasoning agent might continue invoking tools without evidence or budget limits; the exposure that spoofed or repeated business events might initiate unauthorized duplicate workflows remains relevant to the present assessment.

C:

The alternate guard addresses an analytical response might present an unsupported hypothesis as a ledger fact for finance asks why a variance was flagged; the present threat, spoofed or repeated business events might initiate unauthorized duplicate workflows, needs a different enforcement boundary.

Question 276

Business leaders reject an activity-only metric for this process: An agent should summarize account data but not alter it. They need to detect the issue where retrieval configuration accidentally grants write capability. Which performance indicator fits?

  1. Unauthorized grounding documents included in answers
  2. Privileged mutations without verified principal authorization
  3. Unauthorized write operations executed by read-only agents
  4. Cross-role cache disclosure incidents

Explanation

Correct Answer

C:

Because distinct permission surfaces enforce least privilege by task, the outcome worth following is unauthorized write operations executed by read-only agents; the metric is tied to the reported issue—retrieval configuration accidentally grants write capability—for the scenario where an agent should summarize account data but not alter it.

Incorrect Answers

A:

The proposed metric concerns users have different rights to the same knowledge corpus, whereas the failure in this case is that retrieval configuration accidentally grants write capability; the indicator needed here is unauthorized write operations executed by read-only agents.

B:

Privileged mutations without verified principal authorization; that indicator describes a distinct workflow where prompt wording persuades the agent to submit a privileged operation; it does not quantify unauthorized write operations executed by read-only agents, which is needed for the stated case.

D:

The alternative tracks cross-role cache disclosure incidents, arising from the mechanism where cache keys and access validation prevent security-context leakage; its movement alone says little about the failure that retrieval configuration accidentally grants write capability.

Question 277

A quality engineer is writing a regression case after the incident where a new connector allows a service identity to post directly. The underlying business scenario is: Agents suggest payment adjustments and inventory holds. What should the test exercise?

  1. Interrupt processing after the first app commits
  2. Replace a candidate hash in staging and verify rejection
  3. Try posting below and above delegated approval limits
  4. Attempt inference from an unauthorized identity or network

Explanation

Correct Answer

C:

Trying posting below and above delegated approval limits challenges the path where authorization regression gates verify posting authority remains constrained for the business setting agents suggest payment adjustments and inventory holds; the specific defect to exclude is that a new connector allows a service identity to post directly.

Incorrect Answers

A:

For a blocked invoice triggers a hold on related procurement actions, the alternate test is about whether failure-injection tests verify retries and compensation between systems; it cannot prove the outcome needed after the reported regression where a new connector allows a service identity to post directly.

B:

Replacing a candidate hash in staging and verifying rejection verifies an adjacent objective where deployment automation pulls model artifacts from unverified origins; it leaves untested the causal path in which authorization regression gates verify posting authority remains constrained for this case.

D:

Although the exercise might catch untrusted clients submit requests and exfiltrate outputs, it would not directly validate the mechanism that authorization regression gates verify posting authority remains constrained in the target implementation.

Question 278

Operations has not established the cause of this defect: Agent-generated language promises the wrong financial remedy. The associated setting is: Refund, credit, reimbursement and adjustment have different accounting impacts. Which inspection should precede remediation?

  1. Compare outputs against approved language baselines
  2. Compare prompt versions used by dependent apps
  3. Compare generated recommendations with finance-approved term definitions
  4. Compare source timestamps and principal mapping on connector calls

Explanation

Correct Answer

C:

Where agent-generated language promises the wrong financial remedy, the useful evidence comes from comparing generated recommendations with finance-approved term definitions; its focus matches the architectural mechanism that controlled definitions tie each term to approved policies and authorities in the process where refund, credit, reimbursement and adjustment have different accounting impacts.

Incorrect Answers

A:

The proposed probe concerns compliance letters must preserve approved language and disclaimers, where templating and approved clauses limit generated variation; this incident instead requires evidence about controlled definitions tie each term to approved policies and authorities and the failure that agent-generated language promises the wrong financial remedy.

B:

In the process where multiple low-code apps need a common extraction prompt, the diagnostic procedure could explain each app drifts to a different unreviewed AI prompt; this particular incident instead turns on whether controlled definitions tie each term to approved policies and authorities and why agent-generated language promises the wrong financial remedy.

D:

In the separate scenario where sellers need current credit exposure held outside CRM, the same investigation could explain Copilot relies on last quarter's manually imported account report; it does not isolate the reported cause in a system where refund, credit, reimbursement and adjustment have different accounting impacts.

Question 279

The rollout supports this workflow: An assistant responds fluently but cites unrelated support articles. Leaders want a metric that responds to the reported shortcoming that surface quality scores appear healthy despite wrong documents. Which measure best reflects success?

  1. Total processing cost per correct verified extraction
  2. Unauthorized grounding documents included in answers
  3. Correct source passage retrieved within top K
  4. Rate of requests rejected by capacity constraints
  5. Plan selection agreement on a frozen regression set

Explanation

Correct Answer

C:

For the use case where an assistant responds fluently but cites unrelated support articles, the indicator correct source passage retrieved within top K is worth baselining against the earlier issue where surface quality scores appear healthy despite wrong documents.

Incorrect Answers

A:

The proposed metric concerns a new model reduces price but also degrades exact extractions, whereas the failure in this case is that surface quality scores appear healthy despite wrong documents; the indicator needed here is correct source passage retrieved within top K.

B:

Unauthorized grounding documents included in answers; that indicator describes a distinct workflow where users have different rights to the same knowledge corpus; it does not quantify correct source passage retrieved within top K, which is needed for the stated case.

D:

That scorecard concerns the activity in which a regional launch multiplies concurrent agent requests and its result of rate of requests rejected by capacity constraints; the reported shortcoming here is surface quality scores appear healthy despite wrong documents.

E:

A new topic description changes which agent is invoked; its intended measurement is plan selection agreement on a frozen regression set; substituting that result here would not capture whether surface quality scores appear healthy despite wrong documents has been corrected.

Question 280

Implementation alternatives must be compared for a workflow where operations must distinguish healthy traffic from quality incidents. The pilot shows this gap: Teams discover failures through complaints several days later. Which alternative should lead the solution?

  1. Inspect retrieval spans when cited answers omit evidence
  2. Test negative and adversarial cases alongside happy paths
  3. Create actionable SLOs and incident alerts for agent flows
  4. Connect Foundry agent tracing to Application Insights

Explanation

Correct Answer

C:

The mechanism that service-level objectives and alert thresholds turn telemetry into response signals fits the scenario where operations must distinguish healthy traffic from quality incidents; it addresses the defect that teams discover failures through complaints several days later.

Incorrect Answers

A:

For an assistant responds fluently but cites unrelated support articles, this other approach relies on the fact that retrieval telemetry distinguishes evidence search from answer generation; the issue here is teams discover failures through complaints several days later; that mechanism does not establish the required service-level objectives and alert thresholds turn telemetry into response signals.

B:

The alternative addresses a malicious instruction triggers a privileged tool by a design in which negative tests expose policy and validation boundaries; it does not supply the missing relationship for operations must distinguish healthy traffic from quality incidents, where teams discover failures through complaints several days later.

D:

Agent trace spans provide component-level execution evidence through connected telemetry; its main purpose is operations need to correlate prompts, model calls and tools in a Foundry workflow, not correction of the reported defect where teams discover failures through complaints several days later in the target process.

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