Question 1170
The pilot needs a durable correction: New planning behavior occasionally leaks internal notes. This affects the workflow in which a hosted agent gains an additional research tool. Which design choice should the architect approve?
- Preserve customer communication approvals across solution releases
- Monitor newly released service agent behavior by customer cohort
- Integrate runtime tracing and alerting into release readiness
- Use preproduction evaluation gates for agent workflow changes
- Test posting permissions and approvals after agent updates
Explanation
Correct Answer
D:For the need that a hosted agent gains an additional research tool, the relevant mechanism is that evaluation suites validate quality and safety before publication; that architectural dependency addresses the observed failure where new planning behavior occasionally leaks internal notes.
Incorrect Answers
A: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, new planning behavior occasionally leaks internal notes.
B:Cohort telemetry supports safe staged validation; its main purpose is a conversational change affects only one support geography, not correction of the reported defect where new planning behavior occasionally leaks internal notes in the target process.
C:In the case where a newly deployed Foundry agent has no detailed diagnostics, the alternate design depends on connected observability supplies traces and operational metrics; it does not correct the target failure, new planning behavior occasionally leaks internal notes.
E:For agents suggest payment adjustments and inventory holds, this other approach relies on the fact that authorization regression gates verify posting authority remains constrained; the issue here is new planning behavior occasionally leaks internal notes; that mechanism does not establish the required evaluation suites validate quality and safety before publication.
Question 1171
A product owner must measure whether an AI-enabled change actually improves this process: A classifier underperforms on one language group. The earlier problem was: Overall accuracy conceals disproportionate error. Which outcome indicator is most relevant?
- Worst-cohort error rate and unjustified outcome gap
- Worst-performing segment error rate
- Candidate releases satisfying every critical acceptance metric
- Ingested documents passing quality and access gates
Explanation
Correct Answer
B:The observed shortcoming is that overall accuracy conceals disproportionate error; measuring worst-performing segment error rate gives business owners a directly relevant indicator for the process where a classifier underperforms on one language group.
Incorrect Answers
A:The alternative tracks worst-cohort error rate and unjustified outcome gap, arising from the mechanism where segment-aware evaluations surface fairness risks; its movement alone says little about the failure that overall accuracy conceals disproportionate error.
C:The proposed metric concerns teams promote a custom model because it feels better, whereas the failure in this case is that overall accuracy conceals disproportionate error; the indicator needed here is worst-performing segment error rate.
D:That scorecard concerns the activity in which new documents enter a production source automatically and its result of ingested documents passing quality and access gates; the reported shortcoming here is overall accuracy conceals disproportionate error.
Question 1172
Implementation alternatives must be compared for a workflow where an internal model sometimes returns confidential text verbatim. The pilot shows this gap: A client can request arbitrary portions of training examples. Which alternative should lead the solution?
- Apply output filtering and usage restrictions to model interfaces
- Require independent backend authorization for generated tool calls
- Isolate model evaluation and fine-tuning data by classification
- Restrict who can supply tuning data and approve model updates
Explanation
Correct Answer
A:For the need that an internal model sometimes returns confidential text verbatim, the relevant mechanism is that safety controls and authorization checks limit misuse and disclosure; that architectural dependency addresses the observed failure where a client can request arbitrary portions of training examples.
Incorrect Answers
B:When prompt wording persuades the agent to submit a privileged operation, server-side access checks enforce real actor entitlements; in the stated workflow, however, a client can request arbitrary portions of training examples; the relevant architectural behavior is that safety controls and authorization checks limit misuse and disclosure.
C:For another type of architecture, segregated datasets and access controls protect sensitive data is useful; this design must instead support an internal model sometimes returns confidential text verbatim despite the problem where a client can request arbitrary portions of training examples.
D:Separated data-curation and release permissions preserve integrity; its main purpose is business analysts can submit examples but not change production behavior, not correction of the reported defect where a client can request arbitrary portions of training examples in the target process.
Question 1173
An enterprise team is making a platform design decision. Business context: Multiple systems use inconsistent customer identifiers and duplicate cases. Present failure: Responses merge two unrelated customers with similar names. Which approach resolves the mismatch?
- Expose governed operational entities through Dataverse
- Normalize identifiers and remove duplicated grounding records
- Use quality-and-cost telemetry to tune routing thresholds
- Filter retrieval by effective dates and authoritative version
Explanation
Correct Answer
B:The defect is that responses merge two unrelated customers with similar names; the expected business case is that multiple systems use inconsistent customer identifiers and duplicate cases; the design fits because canonical keys and deduplication prevent incorrect cross-record joins.
Incorrect Answers
A:There is a valid place for the mechanism that Dataverse tables and governed relationships provide a reusable business-data surface, namely agents and apps need shared definitions of customers and work orders; it does not remove the specific defect that responses merge two unrelated customers with similar names from this case.
C:For another type of architecture, feedback-driven thresholds rebalance accuracy, latency and spend is useful; this design must instead support multiple systems use inconsistent customer identifiers and duplicate cases despite the problem where responses merge two unrelated customers with similar names.
D:This design would make sense if policy documents contain draft, superseded, and currently effective revisions, particularly where the agent cites a superseded rule despite a newer policy existing; the requested architecture instead needs the dependency that canonical keys and deduplication prevent incorrect cross-record joins because responses merge two unrelated customers with similar names.
Question 1174
A proposed deployment supports this business case: Debugging logs include prompts containing customer data. Reviewers also warn that agent transcripts might expose direct identifiers to roles without a legitimate need. Which restriction mitigates that concrete risk?
- Redact identifiers and restrict transcript access
- Block release when production variables or connections are unresolved
- Impose bounded timeouts and safe partial response rules
- Deny unapproved destinations and excessive scopes
Explanation
Correct Answer
A:The business setting is debugging logs include prompts containing customer data, with the exposure that agent transcripts might expose direct identifiers to roles without a legitimate need; the enforceable protection must correspond to the path where retention schedules and minimization control telemetry exposure.
Incorrect Answers
B:The alternate boundary mitigates imported actions might retain staging endpoints after production deployment, which is relevant to a flow moved to production still points to a test endpoint; it leaves the current workflow vulnerable to agent transcripts might expose direct identifiers to roles without a legitimate need.
C: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 agent transcripts might expose direct identifiers to roles without a legitimate need.
D:Another process could need this guardrail because a third-party agent might send business data to an unapproved destination; the design being assessed supports debugging logs include prompts containing customer data and has the explicit exposure that agent transcripts might expose direct identifiers to roles without a legitimate need.
Question 1175
Business leaders reject an activity-only metric for this process: A legacy portal has no integration API for a clerical workflow. They need to detect the issue where teams propose browser clicks despite an available stable API. Which performance indicator fits?
- Holdout error rate by business-critical segment
- Correctly completed legacy UI tasks without unsafe side effects
- User understanding of verified versus suggested content
- Critical voice actions with verified entity confirmation
Explanation
Correct Answer
B:The use case involves a legacy portal has no integration API for a clerical workflow and the baseline failure that teams propose browser clicks despite an available stable API; the KPI tracks correctly completed legacy UI tasks without unsafe side effects, so improvement has an interpretable operational meaning.
Incorrect Answers
A:The premise of this indicator is separation of training and evaluation yields realistic generalization checks; the architecture being evaluated instead serves a legacy portal has no integration API for a clerical workflow, and its success is measured by correctly completed legacy UI tasks without unsafe side effects.
C:The premise of this indicator is visual provenance separates authoritative records from generated advice; the architecture being evaluated instead serves a legacy portal has no integration API for a clerical workflow, and its success is measured by correctly completed legacy UI tasks without unsafe side effects.
D:That scorecard concerns the activity in which callers provide account numbers and amounts by speech and its result of critical voice actions with verified entity confirmation; the reported shortcoming here is teams propose browser clicks despite an available stable API.
Question 1176
Release reviewers will not accept a generic success message. The business scenario is that an internal helper produces attractive but unsupported policy answers, and the historical failure is: Satisfaction scores hide factual mistakes. Which test is appropriately targeted?
- Ask for a privileged record using an unprivileged test identity
- Run a frozen policy question set after every model change
- Invoke an agent after deployment and locate its trace
- Use an annotated query with a known authoritative paragraph
Explanation
Correct Answer
B:Running a frozen policy question set after every model change challenges the path where expert-checked evaluation examples quantify evidence fidelity for the business setting an internal helper produces attractive but unsupported policy answers; the specific defect to exclude is that satisfaction scores hide factual mistakes.
Incorrect Answers
A:A pass on that verification would speak to a user asks the agent to read an unrelated executive record; it would not show that expert-checked evaluation examples quantify evidence fidelity, which is needed to avoid the observed defect satisfaction scores hide factual mistakes.
C:A pass on that verification would speak to a newly deployed Foundry agent has no detailed diagnostics; it would not show that expert-checked evaluation examples quantify evidence fidelity, which is needed to avoid the observed defect satisfaction scores hide factual mistakes.
D:Although the exercise might catch surface quality scores appear healthy despite wrong documents, it would not directly validate the mechanism that expert-checked evaluation examples quantify evidence fidelity in the target implementation.
Question 1177
Security and business owners want to keep this capability: A request spans procurement, finance and policy expertise. They must also prevent the following: A specialized agent could receive broader tools than its assigned business role. Which guardrail should the design incorporate?
- Limit each specialist to its own permissions and tools
- Separate recommendation privileges from posting permissions
- Restrict access through source permissions, not prompt wording
- Enforce backend permissions separately from frontend display
Explanation
Correct Answer
A:The business setting is a request spans procurement, finance and policy expertise, with the exposure that a specialized agent could receive broader tools than its assigned business role; the enforceable protection must correspond to the path where a coordinating agent selects bounded specialists and merges their outputs.
Incorrect Answers
B:The alternate boundary mitigates an AI recommendation could become an actual posting without an authorized reviewer, which is relevant to service, sales and finance jointly decide high-value concessions; it leaves the current workflow vulnerable to a specialized agent could receive broader tools than its assigned business role.
C:The alternate boundary mitigates an agent prompt could claim permission to read a private SharePoint document, which is relevant to employee policies reside in governed Microsoft 365 documents; it leaves the current workflow vulnerable to a specialized agent could receive broader tools than its assigned business role.
D:That protection is tied to the mechanism that app-hosted integration brings relevant AI actions into an owned user interface; it does not prevent the distinct risk where a specialized agent could receive broader tools than its assigned business role in the system under review.
Question 1178
An agent-assisted service does not behave as specified: A background agent has triggered irreversible duplicate actions. It is used in this situation: Work can proceed from events without human initiation but needs bounded autonomy. Which observable diagnostic procedure is best?
- Trace triggering events and idempotency keys for repeated actions
- Inspect pending tasks and final-result callbacks by correlation ID
- Capture advertised capabilities and actual authorized tool calls
- Examine tool calls by transaction impact and approval requirement
Explanation
Correct Answer
A:In the setting where work can proceed from events without human initiation but needs bounded autonomy, tracing triggering events and idempotency keys for repeated actions can isolate the failure in which a background agent has triggered irreversible duplicate actions, because event triggers and tool permissions can automate predefined safe steps within explicit limits.
Incorrect Answers
B:In the separate scenario where downstream research agents run longer than the interactive request, the same investigation could explain orchestrator returns a partial result as though the workflow finished; it does not isolate the reported cause in a system where work can proceed from events without human initiation but needs bounded autonomy.
C:For the failure that every agent reimplements the same proprietary tool signatures, capturing advertised capabilities and actual authorized tool calls is relevant; this question instead reports that a background agent has triggered irreversible duplicate actions, so different evidence is needed.
D:In the separate scenario where teams propose that an agent perform both low-risk lookups and financial postings, the same investigation could explain the same agent can read an invoice and commit a payment unchecked; it does not isolate the reported cause in a system where work can proceed from events without human initiation but needs bounded autonomy.
Question 1179
A risk analysis highlights a threat in this workflow: An imported agent can authenticate in dev but not production. Exposure: Production consent might grant a wider tenant scope than necessary. Which enforceable safeguard addresses it at the relevant boundary?
- Require approval and a separate action identity for mutations
- Reject broad admin consent not required by the workload
- Block publication without accountable owner and classification
- Scope the identity to only required resource permissions
- Avoid privileged shared accounts for user-specific retrieval
Explanation
Correct Answer
B:Because production consent might grant a wider tenant scope than necessary, a guardrail must constrain the actual system behavior: production consent, identity and connection setup are distinct release checks; protecting a response alone cannot enforce that boundary.
Incorrect Answers
A:The proposed control targets a read-only assistant might acquire write operations by sharing a tool identity, while the given risk scenario is production consent might grant a wider tenant scope than necessary; these exposures arise in different operational contexts.
C:For the use case where business units create agents with no central visibility, the control addresses the threat that production agents might have no recorded accountable owner or risk classification; the specified agent must be protected against production consent might grant a wider tenant scope than necessary.
D:The proposed control targets long-lived credentials might grant an agent unnecessary Azure resource privileges, while the given risk scenario is production consent might grant a wider tenant scope than necessary; these exposures arise in different operational contexts.
E:The alternate boundary mitigates a shared admin token might let ordinary users retrieve executive records, which is relevant to agent looks up records on behalf of employees with different rights; it leaves the current workflow vulnerable to production consent might grant a wider tenant scope than necessary.
Question 1180
A risk analysis highlights a threat in this workflow: An employee requests a standard equipment return initiation. Exposure: A single-purpose return agent might invoke unrelated inventory or purchasing tools. Which enforceable safeguard addresses it at the relevant boundary?
- Limit tool availability and confirmations for risky steps
- Require least-privilege service permissions and action confirmation
- Give the agent only the required return operation
- Keep business-critical formulas out of free-form generation
Explanation
Correct Answer
C:In the context where an employee requests a standard equipment return initiation, the exposure is that a single-purpose return agent might invoke unrelated inventory or purchasing tools; containing it requires a guardrail in the operation where single-purpose tools and typed parameters bound the task.
Incorrect Answers
A:The alternate guard addresses a generative planner might chain tools that exceed the approved workflow scope for users request a status check and an authorized follow-up action; the present threat, a single-purpose return agent might invoke unrelated inventory or purchasing tools, needs a different enforcement boundary.
B:The alternate guard addresses a tool might update tickets using a shared identity with too many permissions for an agent needs to update a governed ticket status; the present threat, a single-purpose return agent might invoke unrelated inventory or purchasing tools, needs a different enforcement boundary.
D:The alternate guard addresses free-form model arithmetic might change a critical service-level calculation for a page must show a reproducible service-level metric; the present threat, a single-purpose return agent might invoke unrelated inventory or purchasing tools, needs a different enforcement boundary.