Microsoft AB-100 (Agentic AI Business Solutions Architect) Practice Test 5

Question 40

A proposed deployment supports this business case: A unique domain vocabulary produces persistent measured errors. Reviewers also warn that customization might promote a model trained on data without a privacy review. Which restriction mitigates that concrete risk?

  1. Keep proprietary calculations and secrets on the service side
  2. Budget for independent hosting, safety testing and monitoring
  3. Avoid publishing benefits without attribution and uncertainty bounds
  4. Apply privacy review and model-version rollback before release

Explanation

Correct Answer

D:

Because customization might promote a model trained on data without a privacy review, a guardrail must constrain the actual system behavior: a measured capability gap, appropriate dataset and MLOps plan justify customization; protecting a response alone cannot enforce that boundary.

Incorrect Answers

A:

The alternate guard addresses a third-party conversational host might receive proprietary scoring logic or secrets for standard conversation needs a proprietary scoring service; the present threat, customization might promote a model trained on data without a privacy review, needs a different enforcement boundary.

B:

The alternate guard addresses a bespoke runtime might be shipped without an accountable security or support team for a workflow needs proprietary scheduling, model choice and runtime control; the present threat, customization might promote a model trained on data without a privacy review, needs a different enforcement boundary.

C:

The alternate guard addresses seasonal demand could be misrepresented as attributable AI benefit for management attributes a sales lift to Copilot without a comparison; the present threat, customization might promote a model trained on data without a privacy review, needs a different enforcement boundary.

Question 41

An agent-assisted service does not behave as specified: A single case exhausts model and tool quotas. It is used in this situation: A research agent repeatedly queries sources on difficult cases. Which observable diagnostic procedure is best?

  1. Inspect trace depth and token consumption on stuck requests
  2. Record recall at K and citation-grounding errors separately
  3. Compare competing topic traces for representative phrasing
  4. Compare tool inventory with legitimate workflow steps

Explanation

Correct Answer

A:

A single case exhausts model and tool quotas; to attribute the defect rather than guess at a fix, inspecting trace depth and token consumption on stuck requests examines the dependency where time, step and spend budgets terminate unproductive loops.

Incorrect Answers

B:

The proposed inspection is associated with stakeholders report hallucinations but cannot isolate search errors, where labeled queries and expected evidence distinguish retrieval from synthesis; it would not expose the failure mode that a single case exhausts model and tool quotas with the required specificity.

C:

Two topics claim nearly identical scheduling intents; its matching diagnostic check is reasonable for agent invokes the wrong booking action for a reschedule; the present investigation needs evidence for the different failure where a single case exhausts model and tool quotas.

D:

For the failure that it autonomously contacts a supplier without approval, comparing tool inventory with legitimate workflow steps is relevant; this question instead reports that a single case exhausts model and tool quotas, so different evidence is needed.

Question 42

In a business situation where an updated agent performs differently months after release, a previous build showed this defect: Old scores cannot be reproduced because prompts and tests changed. Which reproducible test belongs in the release gate?

  1. Rebuild a staged agent from recorded versions and run smoke tests
  2. Re-run a sampled historical release on its frozen corpus
  3. Introduce a known distribution shift and verify review triggers
  4. Rebuild a test index and switch traffic under observation

Explanation

Correct Answer

B:

For the required workflow where an updated agent performs differently months after release, the test needs to reproduce or prevent the failure where old scores cannot be reproduced because prompts and tests changed; re-runing a sampled historical release on its frozen corpus exercises the path where versioned evaluation sets, model IDs and thresholds document release claims.

Incorrect Answers

A:

In the alternate setting where model and tool changes are promoted independently, the test exercises a complete agent release manifest pins related runtime artifacts; even a pass would not show that old scores cannot be reproduced because prompts and tests changed has been prevented.

C:

The tested behavior is that drift monitoring and outcomes justify retraining when performance shifts; the necessary evidence here concerns versioned evaluation sets, model IDs and thresholds document release claims and the previously observed old scores cannot be reproduced because prompts and tests changed.

D:

The suggested test serves a search index changes embedding model and chunking approach, in which existing vectors are queried with incompatible assumptions; the present scenario instead concerns the reported problem where old scores cannot be reproduced because prompts and tests changed.

Question 43

Risk reviewers need a control, not merely a monitoring dashboard. The use case is: Customers believe generated advice is a human decision. Potential harm: Customers might be unable to contest an AI-influenced high-impact recommendation. Which protection fits?

  1. Block release if privileged writes bypass approvals
  2. Require human appeal routes for contested high-impact outcomes
  3. Do not allow autonomous purchasing without approved limits
  4. Use redacted samples and check for misleading cluster labels

Explanation

Correct Answer

B:

Customers believe generated advice is a human decision; its risk scenario is that customers might be unable to contest an AI-influenced high-impact recommendation; the safeguard fits because transparency cues distinguish AI suggestions and human approvals and the restriction addresses that specific exposure.

Incorrect Answers

A:

The alternate guard addresses an updated connector might gain a direct privileged posting route for agents suggest payment adjustments and inventory holds; the present threat, customers might be unable to contest an AI-influenced high-impact recommendation, needs a different enforcement boundary.

C:

For the use case where an AI-driven replenishment proposal influences procurement, the control addresses the threat that autonomous replenishment proposals might place purchases beyond approved limits; the specified agent must be protected against customers might be unable to contest an AI-influenced high-impact recommendation.

D:

The safeguard for thousands of transcripts contain recurring misunderstandings deals with a different threat: a semantic clustering exercise might expose transcript data or invent misleading themes; even after it is enforced, the exposure that customers might be unable to contest an AI-influenced high-impact recommendation remains possible in the target workflow where transparency cues distinguish AI suggestions and human approvals.

Question 44

Before introducing a fix, the team must explain why the agent obeys text retrieved from a supplier document. The affected setting is: External knowledge contains embedded role-change commands. What should be inspected?

  1. Compare answers with retrieved and deliberately missing documents
  2. Compare retrieved document IDs for multiple test principals
  3. Review retrieved snippets alongside final claims
  4. Trace source text influence on tool choices

Explanation

Correct Answer

D:

In the setting where external knowledge contains embedded role-change commands, tracing source text influence on tool choices can isolate the failure in which the agent obeys text retrieved from a supplier document, because trust hierarchy separates lower-trust content from agent instructions.

Incorrect Answers

A:

In the separate scenario where policy answers must be tied to a current source, the same investigation could explain the agent confidently answers when retrieval fails; it does not isolate the reported cause in a system where external knowledge contains embedded role-change commands.

B:

In the process where users have different rights to the same knowledge corpus, the diagnostic procedure could explain agent cites a private HR document to unauthorized staff; this particular incident instead turns on whether trust hierarchy separates lower-trust content from agent instructions and why the agent obeys text retrieved from a supplier document.

C:

For the failure that engineers tune prompts without checking whether evidence was retrieved, reviewing retrieved snippets alongside final claims is relevant; this question instead reports that the agent obeys text retrieved from a supplier document, so different evidence is needed.

Question 45

An investment review asks what measurable result would justify continuing an AI solution where a newly deployed Foundry agent has no detailed diagnostics. The original issue is: The on-call team cannot explain failed tool calls. Which indicator should be tracked?

  1. Release candidates satisfying critical evaluation thresholds
  2. Verified task resolution rate
  3. Time needed to recover a known-good model/data pair
  4. Ingested documents passing quality and access gates
  5. Deployments with complete monitored trace coverage

Explanation

Correct Answer

E:

Because connected observability supplies traces and operational metrics, the outcome worth following is deployments with complete monitored trace coverage; the metric is tied to the reported issue—the on-call team cannot explain failed tool calls—for the scenario where a newly deployed Foundry agent has no detailed diagnostics.

Incorrect Answers

A:

A change in release candidates satisfying critical evaluation thresholds is informative for a hosted agent gains an additional research tool; it does not establish whether the current process achieves deployments with complete monitored trace coverage.

B:

For the workflow where a service bot handles many chats but few cases are resolved, verified task resolution rate is an actionable measure; it does not establish whether the on-call team cannot explain failed tool calls in the situation under review.

C:

Although time needed to recover a known-good model/data pair is observable, the affected process involves a newly deployed Foundry agent has no detailed diagnostics; its acceptance outcome is deployments with complete monitored trace coverage, not the alternative measurement.

D:

For the situation where new documents enter a production source automatically, ingested documents passing quality and access gates is a sensible indicator; the problem that the on-call team cannot explain failed tool calls requires measurement of deployments with complete monitored trace coverage.

Question 46

The architect wants evidence before altering the deployment. In a workflow where agent experience must appear in selected Microsoft 365 contexts, an agent package is assumed to work in every host automatically. What should the team examine first?

  1. Review target host support and publishing settings
  2. Trace latency distributions and retries for connector calls
  3. Review action arguments and created escalation ownership
  4. Compare output format across supported channels

Explanation

Correct Answer

A:

An agent package is assumed to work in every host automatically; to attribute the defect rather than guess at a fix, reviewing target host support and publishing settings examines the dependency where host-specific support, packaging and admin approval shape deployment.

Incorrect Answers

B:

A sales workflow hangs awaiting optional enrichment; examining that other path does not determine whether host-specific support, packaging and admin approval shape deployment, which is what distinguishes the issue in the stated use case.

C:

For the failure that AI suggestions are never converted into traceable escalation records, reviewing action arguments and created escalation ownership is relevant; this question instead reports that an agent package is assumed to work in every host automatically, so different evidence is needed.

D:

The same agent is offered in voice and asynchronous chat; its matching diagnostic check is reasonable for voice responses contain unreadable links and lengthy tables; the present investigation needs evidence for the different failure where an agent package is assumed to work in every host automatically.

Question 47

A test plan focuses on a previously missed failure: Generated output occasionally omits required keys. It occurred when a Copilot prompt feeds records into downstream automation. Which check directly validates that behavior?

  1. Query the same protected record from authorized and denied identities
  2. Test unseen invoices with ambiguous tax fields
  3. Embed hostile instructions in a test knowledge file
  4. Test null, long and unusual user inputs

Explanation

Correct Answer

D:

Testing null, long and unusual user inputs challenges the path where schema validation catches responses that violate integration contracts for the business setting a Copilot prompt feeds records into downstream automation; the specific defect to exclude is that generated output occasionally omits required keys.

Incorrect Answers

A:

In a solution where agent looks up records on behalf of employees with different rights, querying the same protected record from authorized and denied identities could catch the defect that shared admin credentials return restricted cases to ordinary users; the current release gate must challenge the failure where generated output occasionally omits required keys.

B:

Testing unseen invoices with ambiguous tax fields verifies an adjacent objective where a document model performs well overall but fails on tax identifiers; it leaves untested the causal path in which schema validation catches responses that violate integration contracts for this case.

C:

Embedding hostile instructions in a test knowledge file verifies an adjacent objective where external knowledge contains embedded role-change commands; it leaves untested the causal path in which schema validation catches responses that violate integration contracts for this case.

Question 48

An agent-assisted service does not behave as specified: One general agent conflates expertise and may call the wrong tools. It is used in this situation: A request spans procurement, finance and policy expertise. Which observable diagnostic procedure is best?

  1. Audit decision state across sales and finance transactions
  2. Identify workflow steps with fixed rules versus variable language
  3. Observe navigation abandonment and contextual data transfer
  4. Inspect routing decisions and specialist task descriptions

Explanation

Correct Answer

D:

One general agent conflates expertise and may call the wrong tools; to attribute the defect rather than guess at a fix, inspecting routing decisions and specialist task descriptions examines the dependency where a coordinating agent selects bounded specialists and merges their outputs.

Incorrect Answers

A:

The proposed inspection is associated with service, sales and finance jointly decide high-value concessions, where cross-app workflow routes the recommendation to the accountable approver; it would not expose the failure mode that one general agent conflates expertise and may call the wrong tools with the required specificity.

B:

A business process mixes deterministic calculations and open-ended judgments; its matching diagnostic check is reasonable for a single agent unpredictably handles policy rules and narrative summarization; the present investigation needs evidence for the different failure where one general agent conflates expertise and may call the wrong tools.

C:

The proposed probe concerns users must remain in a proprietary work-order application, where app-hosted integration brings relevant AI actions into an owned user interface; this incident instead requires evidence about a coordinating agent selects bounded specialists and merges their outputs and the failure that one general agent conflates expertise and may call the wrong tools.

Question 49

A risk analysis highlights a threat in this workflow: A prompt injection requests data export to an outside address. Exposure: An injected tool instruction might export customer records externally. Which enforceable safeguard addresses it at the relevant boundary?

  1. Redact secrets and restrict access to captured traces
  2. Deny external sends unless independently approved
  3. Reject invalid outputs before they reach mutation endpoints
  4. Impose bounded timeouts and safe partial response rules

Explanation

Correct Answer

B:

Avoiding the consequence where an injected tool instruction might export customer records externally requires a control for a prompt injection requests data export to an outside address; in this architecture, execution-time destination and permission checks block exfiltration, which is the point at which the boundary can be applied.

Incorrect Answers

A:

The alternate guard addresses captured agent traces might store credentials or be visible to unauthorized analysts for operations need to correlate prompts, model calls and tools in a Foundry workflow; the present threat, an injected tool instruction might export customer records externally, needs a different enforcement boundary.

C:

Another process could need this guardrail because a malformed JSON response might reach a write endpoint and corrupt records; the design being assessed supports a prompt injection requests data export to an outside address and has the explicit exposure that an injected tool instruction might export customer records externally.

D:

The alternate boundary mitigates slow dependencies might leave an interaction hanging without bounded failure handling, which is relevant to agents regularly exceed an interactive response-time objective; it leaves the current workflow vulnerable to an injected tool instruction might export customer records externally.

Question 50

A pilot scorecard needs evidence of business improvement rather than only invocation counts. The use case is: A user asks the agent to read an unrelated executive record. Known defect: The model may try a tool outside the user's rights. Which KPI should be selected?

  1. Unauthorized write operations executed by read-only agents
  2. Safe failure responses under dependency loss
  3. Unauthorized retrievals completed in red-team cases
  4. Unapproved external destinations reached through agent tools

Explanation

Correct Answer

C:

For the business use where a user asks the agent to read an unrelated executive record, the result to quantify is unauthorized retrievals completed in red-team cases; it connects measurement to the mechanism where authorization evaluations verify server-side deny behavior, not merely to agent traffic.

Incorrect Answers

A:

For the situation where an agent should summarize account data but not alter it, unauthorized write operations executed by read-only agents is a sensible indicator; the problem that the model may try a tool outside the user's rights requires measurement of unauthorized retrievals completed in red-team cases.

B:

The premise of this indicator is failure injection checks controlled degradation and abstention; the architecture being evaluated instead serves a user asks the agent to read an unrelated executive record, and its success is measured by unauthorized retrievals completed in red-team cases.

D:

A change in unapproved external destinations reached through agent tools is informative for a prompt injection requests data export to an outside address; it does not establish whether the current process achieves unauthorized retrievals completed in red-team cases.

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