Question 500
The pilot needs a durable correction: Host-managed orchestration cannot satisfy a unique execution plan. This affects the workflow in which the solution demands its own model selection and hosting policies. Which design choice should the architect approve?
- Extend a prebuilt agent with an approved business action
- Create a custom-engine agent for bespoke model orchestration
- Use a custom model only when general models fail a defined task
- Use Microsoft 365 Copilot connectors for external knowledge
Explanation
Correct Answer
B:The solution demands its own model selection and hosting policies; its failing behavior is that host-managed orchestration cannot satisfy a unique execution plan; applying the underlying design principle—custom-engine architecture controls runtime, model calls and integrations—connects the requirement to an implementable control.
Incorrect Answers
A:The technical behavior of this alternative is that supported extensibility adds an action while retaining the host's orchestration; the user case being assessed is the solution demands its own model selection and hosting policies with a different failure: host-managed orchestration cannot satisfy a unique execution plan.
C:There is a valid place for the mechanism that a measured capability gap, appropriate dataset and MLOps plan justify customization, namely a unique domain vocabulary produces persistent measured errors; it does not remove the specific defect that host-managed orchestration cannot satisfy a unique execution plan from this case.
D:In the case where teams want approved external records searchable in Copilot, the alternate design depends on connectors expose external enterprise content to supported Microsoft search/Copilot experiences; it does not correct the target failure, host-managed orchestration cannot satisfy a unique execution plan.
Question 501
Risk reviewers need a control, not merely a monitoring dashboard. The use case is: Customer terminology changes over time. Potential harm: Retraining might incorporate sensitive examples without renewed data rights review. Which protection fits?
- Restrict modifications and protect audit storage from tampering
- Follow privacy and role controls on analytics and transcripts
- Route low-confidence tax values for human validation
- Require new data rights and labeling checks before training
Explanation
Correct Answer
D:Customer terminology changes over time; its risk scenario is that retraining might incorporate sensitive examples without renewed data rights review; the safeguard fits because drift monitoring and outcomes justify retraining when performance shifts and the restriction addresses that specific exposure.
Incorrect Answers
A:The safeguard for a model is replaced without a traceable change ticket deals with a different threat: audit logs might be altered to obscure who deployed a model; even after it is enforced, the exposure that retraining might incorporate sensitive examples without renewed data rights review remains possible in the target workflow where drift monitoring and outcomes justify retraining when performance shifts.
B:The proposed control targets conversation analytics might expose personal transcript material beyond an appropriate role, while the given risk scenario is retraining might incorporate sensitive examples without renewed data rights review; these exposures arise in different operational contexts.
C:Another process could need this guardrail because an uncertain extracted tax ID might flow directly into a posted finance record; the design being assessed supports customer terminology changes over time and has the explicit exposure that retraining might incorporate sensitive examples without renewed data rights review.
Question 502
The operating team must monitor whether an architectural correction works. Business context: A small portion of calls time out during peak demand. Historical symptom: Mean response time looks acceptable. Which outcome measure is the best match?
- Ninety-fifth percentile end-to-end response latency
- Total processing cost per correct verified extraction
- Wrong-tool invocation rate on representative intents
- Regional resolution and escalation change after deployment
Explanation
Correct Answer
A:Because tail percentiles surface extreme experiences affecting users, the outcome worth following is ninety-fifth percentile end-to-end response latency; the metric is tied to the reported issue—mean response time looks acceptable—for the scenario where a small portion of calls time out during peak demand.
Incorrect Answers
B:Although total processing cost per correct verified extraction is observable, the affected process involves a small portion of calls time out during peak demand; its acceptance outcome is ninety-fifth percentile end-to-end response latency, not the alternative measurement.
C:For the situation where planner calls the wrong tool among overlapping capabilities, wrong-tool invocation rate on representative intents is a sensible indicator; the problem that mean response time looks acceptable requires measurement of ninety-fifth percentile end-to-end response latency.
D:Regional resolution and escalation change after deployment; that indicator describes a distinct workflow where a conversational change affects only one support geography; it does not quantify ninety-fifth percentile end-to-end response latency, which is needed for the stated case.
Question 503
A design adds autonomous behavior to a process where knowledge retrieval is temporarily unavailable. This creates a documented risk: During a grounding outage the agent might fabricate sensitive guidance. Which enforced control is the best fit?
- Require explicit approval when critical thresholds fail
- Prevent actions before identity and missing fields are verified
- Fail closed for unsupported sensitive answers
- Escalate unresolved high-impact requests instead of guessing
Explanation
Correct Answer
D:In the scenario where knowledge retrieval is temporarily unavailable, the explicit risk is that during a grounding outage the agent might fabricate sensitive guidance; enforcement must follow the mechanism by which failure injection checks controlled degradation and abstention.
Incorrect Answers
A:For the use case where prompt revisions cause intermittent regressions, the control addresses the threat that a maker might publish a candidate despite a failed critical evaluation gate; the specified agent must be protected against during a grounding outage the agent might fabricate sensitive guidance.
B:The alternate guard addresses an agent might act before collecting mandatory identity and form fields for agent must remember order context while asking for missing details; the present threat, during a grounding outage the agent might fabricate sensitive guidance, needs a different enforcement boundary.
C:The proposed control targets missing retrieval evidence might still generate definitive sensitive advice, while the given risk scenario is during a grounding outage the agent might fabricate sensitive guidance; these exposures arise in different operational contexts.
Question 504
An enterprise team is making a platform design decision. Business context: Teams rebuild similar assistants without seeing existing assets. Present failure: Multiple agents offer overlapping tools with inconsistent permissions. Which approach resolves the mismatch?
- Use MCP for an agent's access to external tools and data
- Maintain a reusable approved-agent and prompt catalogue
- Use a controlled baseline to measure causal impact
- Use Dataverse as the governed handoff for customer cases
- Establish a cross-functional AI Center of Excellence playbook
Explanation
Correct Answer
B:Because discovery and ownership metadata support reuse and controlled deprecation, this approach corrects the reported breakdown where multiple agents offer overlapping tools with inconsistent permissions; the affected situation is that teams rebuild similar assistants without seeing existing assets.
Incorrect Answers
A:For a Copilot Studio agent needs a reusable interface to governed operations, this other approach relies on the fact that MCP presents supported tools and resources through a common integration contract; the issue here is multiple agents offer overlapping tools with inconsistent permissions; that mechanism does not establish the required discovery and ownership metadata support reuse and controlled deprecation.
C:There is a valid place for the mechanism that matched comparisons separate solution impact from unrelated shifts, namely management attributes a sales lift to Copilot without a comparison; it does not remove the specific defect that multiple agents offer overlapping tools with inconsistent permissions from this case.
D:Shared authorized records support coordinated business context; its main purpose is sales and service agents need the same customer and case context, not correction of the reported defect where multiple agents offer overlapping tools with inconsistent permissions in the target process.
E:In the case where independent teams choose inconsistent evaluation and security methods, the alternate design depends on a CoE publishes patterns, reusable controls, standards and escalation paths; it does not correct the target failure, multiple agents offer overlapping tools with inconsistent permissions.
Question 505
One enterprise rollout has 3 separately testable requirements. (1) Business use: Hostile instructions arrive both as user input and retrieved documents. Existing shortcoming: The agent is exposed to repeated jailbreak patterns. (2) Business use: External knowledge contains embedded role-change commands. Existing shortcoming: The agent obeys text retrieved from a supplier document. (3) Business use: A prompt injection requests data export to an outside address. Existing shortcoming: Model-level refusal alone may not prevent a tool call. Which 3 success metrics directly satisfy these stated cases? Choose THREE.
- Unapproved external destinations reached through agent tools
- Responses satisfying the agreed output schema
- Detected attacks blocked before unsafe execution
- Successful indirect prompt-injection attacks
Explanation
Correct Answers
A:The outcome unapproved external destinations reached through agent tools measures whether the workflow where a prompt injection requests data export to an outside address is improving; the scorecard separately needs detected attacks blocked before unsafe execution; successful indirect prompt-injection attacks.
C:The outcome detected attacks blocked before unsafe execution measures whether the workflow where hostile instructions arrive both as user input and retrieved documents is improving; the scorecard separately needs successful indirect prompt-injection attacks; unapproved external destinations reached through agent tools.
D:The outcome successful indirect prompt-injection attacks measures whether the workflow where external knowledge contains embedded role-change commands is improving; the scorecard separately needs detected attacks blocked before unsafe execution; unapproved external destinations reached through agent tools.
Incorrect Answers
B:The alternative measures responses satisfying the agreed output schema for the situation where a Copilot prompt feeds records into downstream automation; it cannot substitute for performance outcomes in the separate named cases: hostile instructions arrive both as user input and retrieved documents; external knowledge contains embedded role-change commands; a prompt injection requests data export to an outside address.
Question 506
A release must prove that average handling-time savings disappear during rework will not recur. The application is used in this situation: AI drafts requests quickly but specialists must correct some outputs. Which acceptance test most directly checks that failure mode?
- Compare matched assisted and unassisted case cohorts
- Attempt concurrent updates through both app interfaces
- Re-evaluate the router on production-like stratified examples
- Attempt publication with and without reviewer sign-off
Explanation
Correct Answer
A:The historical failure is that average handling-time savings disappear during rework; for the setting where AI drafts requests quickly but specialists must correct some outputs, comparing matched assisted and unassisted case cohorts verifies the dependency where end-to-end unit cost includes review, escalation and retry effort.
Incorrect Answers
B:The tested behavior is that field ownership and integration contracts prevent competing writes; the necessary evidence here concerns end-to-end unit cost includes review, escalation and retry effort and the previously observed average handling-time savings disappear during rework.
C:Re-evaluating the router on production-like stratified examples verifies an adjacent objective where real task mix differs from prelaunch performance tests; it leaves untested the causal path in which end-to-end unit cost includes review, escalation and retry effort for this case.
D:For standard capabilities draft customer communications but policy requires approval, the alternate test is about whether a review step preserves accountability without discarding assisted drafting; it cannot prove the outcome needed after the reported regression where average handling-time savings disappear during rework.
Question 507
A release must prove that every issue is routed to prompt engineers despite diverse causes will not recur. The application is used in this situation: Complaints use vague labels such as agent is wrong. Which acceptance test most directly checks that failure mode?
- Simulate elevated error rates and verify alert routing
- Run a frozen policy question set after every model change
- Seed paraphrased failures and verify useful grouping
- Classify a mixed set of complaints and check routing consistency
Explanation
Correct Answer
D:Classifying a mixed set of complaints and checking routing consistency is aligned with the failure that every issue is routed to prompt engineers despite diverse causes; its causal relevance is that structured defect taxonomy assigns the right remediation owner for the situation where complaints use vague labels such as agent is wrong.
Incorrect Answers
A:This check exercises a different system behavior: service-level objectives and alert thresholds turn telemetry into response signals; a pass there cannot establish that every issue is routed to prompt engineers despite diverse causes no longer occurs in the required workflow.
B:Running a frozen policy question set after every model change verifies an adjacent objective where an internal helper produces attractive but unsupported policy answers; it leaves untested the causal path in which structured defect taxonomy assigns the right remediation owner for this case.
C:A pass on that verification would speak to thousands of transcripts contain recurring misunderstandings; it would not show that structured defect taxonomy assigns the right remediation owner, which is needed to avoid the observed defect every issue is routed to prompt engineers despite diverse causes.
Question 508
During integration testing, happy-path fixtures omit unauthorized and invalid inputs. The associated business context is: Agents can call APIs with customer IDs and amount values. Which architectural remedy is the strongest match?
- Test audit completeness for AI-assisted customer concessions
- Generate boundary cases from approved acceptance requirements
- Generate negative cases from tool and identity specifications
- Build risk-based test coverage for critical user journeys
Explanation
Correct Answer
C:Because specification-guided generation covers missing fields and denied roles, this approach corrects the reported breakdown where happy-path fixtures omit unauthorized and invalid inputs; the affected situation is that agents can call APIs with customer IDs and amount values.
Incorrect Answers
A:When service recommends a price adjustment requiring finance approval, transaction-path tests confirm recommendation, review and posting records; in the stated workflow, however, happy-path fixtures omit unauthorized and invalid inputs; the relevant architectural behavior is that specification-guided generation covers missing fields and denied roles.
B:In the case where agent handles threshold-based return decisions, the alternate design depends on assisted test generation explores stated boundary conditions; it does not correct the target failure, happy-path fixtures omit unauthorized and invalid inputs.
D:In the case where agent supports refunds, order status and generic FAQs, the alternate design depends on risk-weighting prioritizes high-impact actions and failure paths; it does not correct the target failure, happy-path fixtures omit unauthorized and invalid inputs.
Question 509
A release must prove that agent actions fail only after the connector reaches production will not recur. The application is used in this situation: An external API changes a required field used by an agent. Which acceptance test most directly checks that failure mode?
- Replay representative requests against the new connector version
- Try to pass sensitive records to a prohibited external connector
- Rebuild a test index and switch traffic under observation
- Try a forbidden refund despite a permissive model instruction
Explanation
Correct Answer
A:For the required workflow where an external API changes a required field used by an agent, the test needs to reproduce or prevent the failure where agent actions fail only after the connector reaches production; replaying representative requests against the new connector version exercises the path where versioned API contracts and consumer checks detect breaking changes.
Incorrect Answers
B:The verification concerning a Copilot Studio agent can read corporate data and use public messaging checks Power Platform data policies constrain connector use and movement pathways against a permitted connector combination enables uncontrolled disclosure; its result says little about the observed issue where agent actions fail only after the connector reaches production.
C:This check exercises a different system behavior: explicit index versions support compatible retrieval and rollback; a pass there cannot establish that agent actions fail only after the connector reaches production no longer occurs in the required workflow.
D:The verification concerning prompt wording persuades the agent to submit a privileged operation checks server-side access checks enforce real actor entitlements against API trusts the model's text assertion of permission; its result says little about the observed issue where agent actions fail only after the connector reaches production.
Question 510
The operating team must monitor whether an architectural correction works. Business context: Customer terminology changes over time. Historical symptom: Team retrains monthly despite unchanged model behavior. Which outcome measure is the best match?
- Deployed models with complete registry evidence
- Model changes with complete actor and approval records
- True performance drift detected before critical SLO breach
- Time needed to recover a known-good model/data pair
Explanation
Correct Answer
C:For the business use where customer terminology changes over time, the result to quantify is true performance drift detected before critical SLO breach; it connects measurement to the mechanism where drift monitoring and outcomes justify retraining when performance shifts, not merely to agent traffic.
Incorrect Answers
A:For the workflow where data scientists cannot explain why a newer classifier behaves differently, deployed models with complete registry evidence is an actionable measure; it does not establish whether team retrains monthly despite unchanged model behavior in the situation under review.
B:The premise of this indicator is deployment audit links actor, approved artifact and release time; the architecture being evaluated instead serves customer terminology changes over time, and its success is measured by true performance drift detected before critical SLO breach.
D:For the workflow where a preprocessing release changes category labels incorrectly, time needed to recover a known-good model/data pair is an actionable measure; it does not establish whether team retrains monthly despite unchanged model behavior in the situation under review.