Question 720
Governance approves the business purpose but requires control of this exposure: Invalid generated JSON might be written directly into a business record. The agent will be used when agents return inconsistent fields for downstream integrations. Which boundary is most defensible?
- Exclude secrets and ensure training-data rights
- Preserve table and row access controls for retrieval
- Set enforceable monetary and scope thresholds outside prompts
- Require policy-owner approval for any rule change
- Reject malformed output before writing business records
Explanation
Correct Answer
E:When invalid generated JSON might be written directly into a business record, the use case agents return inconsistent fields for downstream integrations requires a guardrail aligned with the behavior that clear instructions and structured outputs reduce avoidable ambiguity.
Incorrect Answers
A:This restriction is associated with a stable specialized classification style persists despite prompt examples and addresses the exposure that sensitive or unlicensed content could be incorporated into tuning examples; it does not specifically protect the target workflow against invalid generated JSON might be written directly into a business record.
B:Users ask about governed CRM records rather than prose manuals; its guardrail addresses the possibility that Dataverse retrieval might reveal a row not authorized to the caller; the present architecture must constrain the different threat that invalid generated JSON might be written directly into a business record.
C:The agent may refund large amounts or deactivate accounts; its guardrail addresses the possibility that a model could override a monetary limit through conversational instructions; the present architecture must constrain the different threat that invalid generated JSON might be written directly into a business record.
D:The alternate guard addresses an unapproved change to the approval matrix could silently alter authorization decisions for the task follows fixed approval branches with no interpretation needed; the present threat, invalid generated JSON might be written directly into a business record, needs a different enforcement boundary.
Question 721
A business stakeholder describes this operating situation: Service recommends a price adjustment requiring finance approval. Testing instead shows that the final order lacks traceability to the AI suggestion and approver. What design approach is most appropriate?
- Test Sales-to-Service customer context and case handoff
- Review AI-generated test oracles with independent experts
- Test audit completeness for AI-assisted customer concessions
- Preserve customer communication approvals across solution releases
Explanation
Correct Answer
C:The architecture must reconcile two facts: service recommends a price adjustment requiring finance approval, yet the final order lacks traceability to the AI suggestion and approver; the selected pattern does so through the mechanism that transaction-path tests confirm recommendation, review and posting records.
Incorrect Answers
A:For a won opportunity should create a properly owned onboarding case, this other approach relies on the fact that end-to-end tests verify identity, mapping and assignment across apps; the issue here is the final order lacks traceability to the AI suggestion and approver; that mechanism does not establish the required transaction-path tests confirm recommendation, review and posting records.
B:The technical behavior of this alternative is that independent acceptance references prevent self-confirming evaluation; the user case being assessed is service recommends a price adjustment requiring finance approval with a different failure: the final order lacks traceability to the AI suggestion and approver.
D:In the case where service agents draft customer messages but cannot send without review, the alternate design depends on business approval checks remain independent of agent version; it does not correct the target failure, the final order lacks traceability to the AI suggestion and approver.
Question 722
A planned AI solution must support this situation: Users report stale answers after a policy refresh. Its prototype fails because feedback tickets lack the evidence needed to reproduce them. Which design best fits the constraint?
- Test negative and adversarial cases alongside happy paths
- Classify feedback by factual, workflow, usability and policy defect
- Validate and sanitize tool output before rendering it
- Link negative feedback to grounded traces and source versions
Explanation
Correct Answer
D:The defect is that feedback tickets lack the evidence needed to reproduce them; the expected business case is that users report stale answers after a policy refresh; the design fits because trace and content-version linkage makes failures testable.
Incorrect Answers
A:For another type of architecture, negative tests expose policy and validation boundaries is useful; this design must instead support users report stale answers after a policy refresh despite the problem where feedback tickets lack the evidence needed to reproduce them.
B:The alternative addresses every issue is routed to prompt engineers despite diverse causes by a design in which structured defect taxonomy assigns the right remediation owner; it does not supply the missing relationship for users report stale answers after a policy refresh, where feedback tickets lack the evidence needed to reproduce them.
C:Agent-generated pages render an unsafe instruction or link; the alternate design can address that situation because trusted rendering and output validation constrain returned content, whereas this question needs a solution to feedback tickets lack the evidence needed to reproduce them.
Question 723
A proposed deployment supports this business case: AI proposes eligibility decisions affecting customers. Reviewers also warn that reviewers might blindly approve consequential outputs without evidence. Which restriction mitigates that concrete risk?
- Disable the candidate for a cohort on critical regression
- Require human sign-off on critical test oracles
- Keep final authority with qualified decision makers
- Require a tested human escalation path for unmatched cases
Explanation
Correct Answer
C:In the context where AI proposes eligibility decisions affecting customers, the exposure is that reviewers might blindly approve consequential outputs without evidence; containing it requires a guardrail in the operation where explainable review interfaces support accountable judgment.
Incorrect Answers
A:Where a conversational change affects only one support geography, the guard can address a regional pilot might produce a severe local regression hidden by global averages; the exposure that reviewers might blindly approve consequential outputs without evidence remains relevant to the present assessment.
B:The alternate guard addresses a fabricated test oracle might approve a flawed agent behavior for Copilot writes a test expecting a plausible but wrong answer; the present threat, reviewers might blindly approve consequential outputs without evidence, needs a different enforcement boundary.
D:For the use case where a support agent changes how it classifies priority intents, the control addresses the threat that a new support agent might route urgent cases to a general queue; the specified agent must be protected against reviewers might blindly approve consequential outputs without evidence.
Question 724
An investment review asks what measurable result would justify continuing an AI solution where many tenants have identically named projects. The original issue is: Retrieval mixes records from two organizations. Which indicator should be tracked?
- Cross-tenant retrieval violations
- Successful inferences within latency SLO
- Analyst-accepted explanations with verified references
- Precision at top results for identifier and natural-language queries
Explanation
Correct Answer
A:This scorecard needs an outcome related to many tenants have identically named projects; cross-tenant retrieval violations. That indicator can be checked against the earlier symptom that retrieval mixes records from two organizations.
Incorrect Answers
B:Successful inferences within latency SLO; that indicator describes a distinct workflow where a specialized model must serve production apps reliably; it does not quantify cross-tenant retrieval violations, which is needed for the stated case.
C:Analyst-accepted explanations with verified references; that indicator describes a distinct workflow where finance asks why a variance was flagged; it does not quantify cross-tenant retrieval violations, which is needed for the stated case.
D:For the situation where product catalog queries mix synonyms and exact serial codes, precision at top results for identifier and natural-language queries is a sensible indicator; the problem that retrieval mixes records from two organizations requires measurement of cross-tenant retrieval violations.
Question 725
One enterprise rollout has 2 separately testable requirements. (1) Business use: An agent needs to update a governed ticket status. Existing shortcoming: Its knowledge source can explain but cannot commit a database change. (2) Business use: The orchestrator chooses between similar invoice and credit tools. Existing shortcoming: It passes customer IDs into supplier inputs. Which 2 success metrics directly satisfy these stated cases? Choose TWO.
- Ticket updates completed with attributable credentials
- Drafts meeting structure and source-accuracy criteria
- Successful traceable escalations initiated from assistance
- Correct tool selection and input-binding rate
Explanation
Correct Answers
A:The outcome ticket updates completed with attributable credentials measures whether the workflow where an agent needs to update a governed ticket status is improving; the scorecard separately needs correct tool selection and input-binding rate.
D:The outcome correct tool selection and input-binding rate measures whether the workflow where the orchestrator chooses between similar invoice and credit tools is improving; the scorecard separately needs ticket updates completed with attributable credentials.
Incorrect Answers
B:The alternative measures drafts meeting structure and source-accuracy criteria for the situation where marketing specialists need tailored descriptions from existing briefings; it cannot substitute for performance outcomes in the separate named cases: an agent needs to update a governed ticket status; the orchestrator chooses between similar invoice and credit tools.
C:The alternative measures successful traceable escalations initiated from assistance for the situation where a representative must initiate an exception workflow from the case; it cannot substitute for performance outcomes in the separate named cases: an agent needs to update a governed ticket status; the orchestrator chooses between similar invoice and credit tools.
Question 726
The pilot needs a durable correction: One user's cached confidential answer appears for another user. This affects the workflow in which shared result caching makes answers fast across roles. Which design choice should the architect approve?
- Version training datasets with lineage and access records
- Enforce tenant and entity boundaries before vector search results are used
- Use identity-aware caching for grounded agent responses
- Use separate service principals and least-privilege access per environment
Explanation
Correct Answer
C:The mechanism that cache keys and access validation prevent security-context leakage fits the scenario where shared result caching makes answers fast across roles; it addresses the defect that one user's cached confidential answer appears for another user.
Incorrect Answers
A:For model retraining consumes periodically revised business records, this other approach relies on the fact that immutable snapshots and lineage tie runs to approved sources; the issue here is one user's cached confidential answer appears for another user; that mechanism does not establish the required cache keys and access validation prevent security-context leakage.
B:For two customers store similarly named support incidents, this other approach relies on the fact that partitioning and authorized filters constrain relevant candidates; the issue here is one user's cached confidential answer appears for another user; that mechanism does not establish the required cache keys and access validation prevent security-context leakage.
D:For another type of architecture, environment-scoped identities isolate access and simplify incident containment is useful; this design must instead support shared result caching makes answers fast across roles despite the problem where one user's cached confidential answer appears for another user.
Question 727
A release must prove that a general conversational turn produces inconsistent output fields will not recur. The application is used in this situation: An agent must turn service notes into a fixed JSON summary. Which acceptance test most directly checks that failure mode?
- Test mixed intents with missing action parameters
- Install the agent to a pilot Microsoft 365 audience
- Run representative notes and validate schema and exclusions
- Test borderline utterances against both descriptions
Explanation
Correct Answer
C:A general conversational turn produces inconsistent output fields; a release test must establish that this no longer occurs; running representative notes and validating schema and exclusions measures the specific behavior for which prompt actions encapsulate reusable inputs and expected outputs.
Incorrect Answers
A:The verification concerning users request a status check and an authorized follow-up action checks the planner can select supported knowledge and tools for compound intents against a single-topic design cannot combine both operations naturally; its result says little about the observed issue where a general conversational turn produces inconsistent output fields.
B:For employees want curated assistance without owning a separate LLM runtime, the alternate test is about whether declarative agents define instructions, knowledge and actions in the supported host; it cannot prove the outcome needed after the reported regression where a general conversational turn produces inconsistent output fields.
D:Although the exercise might catch agent invokes the wrong booking action for a reschedule, it would not directly validate the mechanism that prompt actions encapsulate reusable inputs and expected outputs in the target implementation.
Question 728
During integration testing, agent cites a private HR document to unauthorized staff. The associated business context is: Users have different rights to the same knowledge corpus. Which architectural remedy is the strongest match?
- Inspect retrieval spans when cited answers omit evidence
- Test grounded prompts for citation and abstention behavior
- Apply document-level access trimming during retrieval
- Restrict who can supply tuning data and approve model updates
Explanation
Correct Answer
C:Because retrieval applies effective source entitlements to candidate documents, this approach corrects the reported breakdown where agent cites a private HR document to unauthorized staff; the affected situation is that users have different rights to the same knowledge corpus.
Incorrect Answers
A:This design would make sense if an assistant responds fluently but cites unrelated support articles, particularly where surface quality scores appear healthy despite wrong documents; the requested architecture instead needs the dependency that retrieval applies effective source entitlements to candidate documents because agent cites a private HR document to unauthorized staff.
B:For policy answers must be tied to a current source, this other approach relies on the fact that evidence and refusal evaluations enforce source-supported claims; the issue here is agent cites a private HR document to unauthorized staff; that mechanism does not establish the required retrieval applies effective source entitlements to candidate documents.
D:There is a valid place for the mechanism that separated data-curation and release permissions preserve integrity, namely business analysts can submit examples but not change production behavior; it does not remove the specific defect that agent cites a private HR document to unauthorized staff from this case.
Question 729
Security and business owners want to keep this capability: Agent handles threshold-based return decisions. They must also prevent the following: Generated test expectations might be accepted without domain-owner review. Which guardrail should the design incorporate?
- Fail closed for unsupported sensitive answers
- Have experts approve expected outcomes before automation
- Forbid direct mutation before recorded delegated approval
- Roll back a routing change when a critical segment regresses
- Require explicit approval when critical thresholds fail
Explanation
Correct Answer
B:Avoiding the consequence where generated test expectations might be accepted without domain-owner review requires a control for agent handles threshold-based return decisions; in this architecture, assisted test generation explores stated boundary conditions, which is the point at which the boundary can be applied.
Incorrect Answers
A:For the use case where an assistant responds fluently but cites unrelated support articles, the control addresses the threat that missing retrieval evidence might still generate definitive sensitive advice; the specified agent must be protected against generated test expectations might be accepted without domain-owner review.
C:This restriction is associated with service recommends a price adjustment requiring finance approval and addresses the exposure that a suggested customer concession might post before delegated finance approval; it does not specifically protect the target workflow against generated test expectations might be accepted without domain-owner review.
D:The safeguard for average accuracy hides failures on complex finance questions deals with a different threat: a model router change might silently degrade the highest-risk task segment; even after it is enforced, the exposure that generated test expectations might be accepted without domain-owner review remains possible in the target workflow where assisted test generation explores stated boundary conditions.
E:That protection is tied to the mechanism that stable automated suites compare response quality across versions; it does not prevent the distinct risk where generated test expectations might be accepted without domain-owner review in the system under review.
Question 730
Security and business owners want to keep this capability: Sales and Finance roles have different record entitlements. They must also prevent the following: A cross-app handoff might expose invoices without valid identity propagation. Which guardrail should the design incorporate?
- Limit entity visibility to authorized finance roles
- Prevent actions before identity and missing fields are verified
- Require finance process-owner approval for production changes
- Deny access when identity propagation fails
Explanation
Correct Answer
D:Avoiding the consequence where a cross-app handoff might expose invoices without valid identity propagation requires a control for sales and Finance roles have different record entitlements; in this architecture, cross-role tests verify that authorization survives integration, which is the point at which the boundary can be applied.
Incorrect Answers
A:The alternate guard addresses AI retrieval might expose virtual finance entities outside the reader's role for AI assistance reads operational records through Dataverse; the present threat, a cross-app handoff might expose invoices without valid identity propagation, needs a different enforcement boundary.
B:Where agent must remember order context while asking for missing details, the guard can address an agent might act before collecting mandatory identity and form fields; the exposure that a cross-app handoff might expose invoices without valid identity propagation remains relevant to the present assessment.
C:The proposed control targets a finance Copilot prompt might change during a protected close window, while the given risk scenario is a cross-app handoff might expose invoices without valid identity propagation; these exposures arise in different operational contexts.