Question 920
A pilot scorecard needs evidence of business improvement rather than only invocation counts. The use case is: Current users need one governed custom action inside their workflow. Known defect: Full replacement increases retraining and migration risk. Which KPI should be selected?
- End-to-end tasks completed when dependencies are healthy
- New-policy answer accuracy after ingestion
- Incremental maintenance effort of the extension
- Relevant external content found without access violations
Explanation
Correct Answer
C:For the use case where current users need one governed custom action inside their workflow, the indicator incremental maintenance effort of the extension is worth baselining against the earlier issue where full replacement increases retraining and migration risk.
Incorrect Answers
A:For the workflow where standard conversation needs a proprietary scoring service, end-to-end tasks completed when dependencies are healthy is an actionable measure; it does not establish whether full replacement increases retraining and migration risk in the situation under review.
B:Although new-policy answer accuracy after ingestion is observable, the affected process involves current users need one governed custom action inside their workflow; its acceptance outcome is incremental maintenance effort of the extension, not the alternative measurement.
D:The proposed metric concerns teams want approved external records searchable in Copilot, whereas the failure in this case is that full replacement increases retraining and migration risk; the indicator needed here is incremental maintenance effort of the extension.
Question 921
A policy review covers an agent for which downstream research agents run longer than the interactive request. Its most relevant exposure is that retried delegated work could charge a customer or create a record more than once. What safeguard should be required before release?
- Prevent duplicate side effects during task retries
- Operate independent identity, safety and observability controls
- Enforce backend permissions separately from frontend display
- Separate recommendation privileges from posting permissions
Explanation
Correct Answer
A:Because retried delegated work could charge a customer or create a record more than once, a guardrail must constrain the actual system behavior: task IDs, status transitions, deadlines and durable handoff state support coordination; protecting a response alone cannot enforce that boundary.
Incorrect Answers
B:For the use case where the solution demands its own model selection and hosting policies, the control addresses the threat that an owned custom runtime might lack the host's identity and safety governance; the specified agent must be protected against retried delegated work could charge a customer or create a record more than once.
C:Where users must remain in a proprietary work-order application, the guard can address a hidden app field could still be retrieved through an over-privileged backend call; the exposure that retried delegated work could charge a customer or create a record more than once remains relevant to the present assessment.
D: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 retried delegated work could charge a customer or create a record more than once.
Question 922
A product owner must measure whether an AI-enabled change actually improves this process: Employees want curated assistance without owning a separate LLM runtime. The earlier problem was: A custom service would duplicate built-in host orchestration. Which outcome indicator is most relevant?
- Correct tool selection and input-binding rate
- Prompt action responses conforming to output contract
- Successful authorized external service actions
- Role-specific tasks completed in supported Copilot surfaces
Explanation
Correct Answer
D:Because declarative agents define instructions, knowledge and actions in the supported host, the outcome worth following is role-specific tasks completed in supported Copilot surfaces; the metric is tied to the reported issue—a custom service would duplicate built-in host orchestration—for the scenario where employees want curated assistance without owning a separate LLM runtime.
Incorrect Answers
A:The alternative tracks correct tool selection and input-binding rate, arising from the mechanism where specific tool metadata helps the planner choose and parameterize actions; its movement alone says little about the failure that a custom service would duplicate built-in host orchestration.
B:For the situation where an agent must turn service notes into a fixed JSON summary, prompt action responses conforming to output contract is a sensible indicator; the problem that a custom service would duplicate built-in host orchestration requires measurement of role-specific tasks completed in supported Copilot surfaces.
C:Successful authorized external service actions; that indicator describes a distinct workflow where a Teams user should query and update a supported external service; it does not quantify role-specific tasks completed in supported Copilot surfaces, which is needed for the stated case.
Question 923
A risk analysis highlights a threat in this workflow: Model retraining consumes periodically revised business records. Exposure: Training could ingest data whose use rights expired or were never approved. Which enforceable safeguard addresses it at the relevant boundary?
- Require reviewed datasets and independent promotion sign-off
- Keep rollback available until all critical criteria pass
- Retain audit records of source replacements and withdrawals
- Prevent training with unapproved or expired sensitive data
Explanation
Correct Answer
D:In the scenario where model retraining consumes periodically revised business records, the explicit risk is that training could ingest data whose use rights expired or were never approved; enforcement must follow the mechanism by which immutable snapshots and lineage tie runs to approved sources.
Incorrect Answers
A:Business analysts can submit examples but not change production behavior; its guardrail addresses the possibility that unreviewed training examples might poison a promoted model; the present architecture must constrain the different threat that training could ingest data whose use rights expired or were never approved.
B:This restriction is associated with a tuned model claims higher accuracy but adds latency and addresses the exposure that a new classifier might replace its safe baseline before comparisons finish; it does not specifically protect the target workflow against training could ingest data whose use rights expired or were never approved.
C:The proposed control targets a disputed answer might be untraceable to its source and index revision, while the given risk scenario is training could ingest data whose use rights expired or were never approved; these exposures arise in different operational contexts.
Question 924
An incident review has one confirmed symptom: A general conversational turn produces inconsistent output fields. For the use case where an agent must turn service notes into a fixed JSON summary, which targeted diagnostic check would reveal the most pertinent evidence?
- Review agent capability manifest against requested functions
- Review plans for omitted or duplicate tool calls
- Review action inputs, examples and schema-parse logs
- Inspect missing-field collection and final action payload
Explanation
Correct Answer
C:The investigation directly tests the causal connection: prompt actions encapsulate reusable inputs and expected outputs; it collects evidence for a workflow where an agent must turn service notes into a fixed JSON summary but the outcome is instead that a general conversational turn produces inconsistent output fields.
Incorrect Answers
A:In the process where employees want curated assistance without owning a separate LLM runtime, the diagnostic procedure could explain a custom service would duplicate built-in host orchestration; this particular incident instead turns on whether prompt actions encapsulate reusable inputs and expected outputs and why a general conversational turn produces inconsistent output fields.
B:Reviewing plans for omitted or duplicate tool calls explores a separate connection, where the planner can select supported knowledge and tools for compound intents; the diagnosed requirement is an agent must turn service notes into a fixed JSON summary, and the actual defect is a general conversational turn produces inconsistent output fields.
D:Inspecting missing-field collection and final action payload explores a separate connection, where state tracking captures required inputs before any action; the diagnosed requirement is an agent must turn service notes into a fixed JSON summary, and the actual defect is a general conversational turn produces inconsistent output fields.
Question 925
The pilot needs a durable correction: A prompt update changes decisions for all users at once. This affects the workflow in which new behavior must be evaluated without exposing the full workforce. Which design choice should the architect approve?
- Decompose end-to-end workflows into agent-appropriate tasks
- Use staged rollout with feature flags and pilot cohorts
- Implement an abstain-or-escalate response policy for uncertain outputs
- Extend a prebuilt agent with an approved business action
Explanation
Correct Answer
B:Without the intended behavior, a prompt update changes decisions for all users at once; the proposed architecture establishes the specific design principle that controlled release gates restrict blast radius and support rollback for the process where new behavior must be evaluated without exposing the full workforce.
Incorrect Answers
A:There is a valid place for the mechanism that task decomposition separates deterministic services from bounded agent reasoning, namely a business process mixes deterministic calculations and open-ended judgments; it does not remove the specific defect that a prompt update changes decisions for all users at once from this case.
C:Confidence and evidence checks route insufficiently supported cases to humans; its main purpose is the system cannot always find evidence for regulated guidance, not correction of the reported defect where a prompt update changes decisions for all users at once in the target process.
D: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 new behavior must be evaluated without exposing the full workforce with a different failure: a prompt update changes decisions for all users at once.
Question 926
Governance approves the business purpose but requires control of this exposure: A canvas app might send unnecessary protected fields into an inference prompt. The agent will be used when a canvas app passes customer records to a model action. Which boundary is most defensible?
- Remove unneeded attributes before model invocation
- Limit creation privileges and prevent duplicate escalations
- Reject generated fields outside the approved schema
- Stop and compensate when a required approval fails
Explanation
Correct Answer
A:The business setting is a canvas app passes customer records to a model action, with the exposure that a canvas app might send unnecessary protected fields into an inference prompt; the enforceable protection must correspond to the path where server-side authorization and minimized inputs protect data flow.
Incorrect Answers
B:Another process could need this guardrail because repeated agent calls might create duplicate escalation records or excess privileges; the design being assessed supports a canvas app passes customer records to a model action and has the explicit exposure that a canvas app might send unnecessary protected fields into an inference prompt.
C:An agent must turn service notes into a fixed JSON summary; its guardrail addresses the possibility that generated response keys might be trusted even when they violate the schema; the present architecture must constrain the different threat that a canvas app might send unnecessary protected fields into an inference prompt.
D:The proposed control targets a flow might notify the customer even though a required approval failed, while the given risk scenario is a canvas app might send unnecessary protected fields into an inference prompt; these exposures arise in different operational contexts.
Question 927
The acceptance board asks how to verify a corrected design. The setting is: Multiple systems use inconsistent customer identifiers and duplicate cases. Last time, responses merge two unrelated customers with similar names. Which scenario should the team execute?
- Query the same entity from two authorized clients and reconcile attributes
- Re-evaluate the router on production-like stratified examples
- Search the catalogue by task and verify current ownership
- Seed lookalike records and check correct entity separation
- Run multi-hop evaluation cases for the strong-model route
Explanation
Correct Answer
D:For the required workflow where multiple systems use inconsistent customer identifiers and duplicate cases, the test needs to reproduce or prevent the failure where responses merge two unrelated customers with similar names; seeding lookalike records and checking correct entity separation exercises the path where canonical keys and deduplication prevent incorrect cross-record joins.
Incorrect Answers
A:The tested behavior is that Dataverse tables and governed relationships provide a reusable business-data surface; the necessary evidence here concerns canonical keys and deduplication prevent incorrect cross-record joins and the previously observed responses merge two unrelated customers with similar names.
B:A pass on that verification would speak to real task mix differs from prelaunch performance tests; it would not show that canonical keys and deduplication prevent incorrect cross-record joins, which is needed to avoid the observed defect responses merge two unrelated customers with similar names.
C:In a solution where teams rebuild similar assistants without seeing existing assets, searching the catalogue by task and verifying current ownership could catch the defect that multiple agents offer overlapping tools with inconsistent permissions; the current release gate must challenge the failure where responses merge two unrelated customers with similar names.
E:A pass on that verification would speak to some cases need cross-document reasoning and detailed justification; it would not show that canonical keys and deduplication prevent incorrect cross-record joins, which is needed to avoid the observed defect responses merge two unrelated customers with similar names.
Question 928
One enterprise rollout has 2 separately testable requirements. (1) Business use: Personal read-only assistants and payment agents have different risk. Existing shortcoming: Every agent receives the same minimal review. (2) Business use: Employees install external agents with unknown data pathways. Existing shortcoming: An unreviewed plugin exports sensitive tenant information. Which 2 success metrics directly satisfy these stated cases? Choose TWO.
- User-impacting failures during agent release changes
- High-severity agent defects remediated within policy target
- Releases conforming to risk-tier controls
- Third-party agents completing security review before activation
Explanation
Correct Answers
C:The outcome releases conforming to risk-tier controls measures whether the workflow where personal read-only assistants and payment agents have different risk is improving; the scorecard separately needs third-party agents completing security review before activation.
D:The outcome third-party agents completing security review before activation measures whether the workflow where employees install external agents with unknown data pathways is improving; the scorecard separately needs releases conforming to risk-tier controls.
Incorrect Answers
A:The alternative measures user-impacting failures during agent release changes for the situation where business applications depend on the agent endpoint continuously; it cannot substitute for performance outcomes in the separate named cases: personal read-only assistants and payment agents have different risk; employees install external agents with unknown data pathways.
B:The alternative measures high-severity agent defects remediated within policy target for the situation where product owners favor easy fixes over high-consequence failures; it cannot substitute for performance outcomes in the separate named cases: personal read-only assistants and payment agents have different risk; employees install external agents with unknown data pathways.
Question 929
The acceptance board asks how to verify a corrected design. The setting is: An AI-driven replenishment proposal influences procurement. Last time, recommendation changes spread to all planners without evaluation. Which scenario should the team execute?
- Try to send high-impact notices without approval in staging
- Upgrade one app while retaining the prior dependent version
- Introduce a bad demand forecast and verify rollback
- Test extraction and secret-disclosure prompts with synthetic fixtures
Explanation
Correct Answer
C:Introducing a bad demand forecast and verifying rollback challenges the path where a pilot cohort and rollback path isolate operational impact for the business setting an AI-driven replenishment proposal influences procurement; the specific defect to exclude is that recommendation changes spread to all planners without evaluation.
Incorrect Answers
A:The tested behavior is that business approval checks remain independent of agent version; the necessary evidence here concerns a pilot cohort and rollback path isolate operational impact and the previously observed recommendation changes spread to all planners without evaluation.
B:For a customer handoff depends on shared record and term mappings, the alternate test is about whether cross-app release plans validate shared contracts and timing; it cannot prove the outcome needed after the reported regression where recommendation changes spread to all planners without evaluation.
D:The suggested test serves an internal model sometimes returns confidential text verbatim, in which a client can request arbitrary portions of training examples; the present scenario instead concerns the reported problem where recommendation changes spread to all planners without evaluation.
Question 930
An incident review has one confirmed symptom: Copilot attaches one customer's risk note to another account. For the use case where two systems share company names but not primary keys, which targeted diagnostic check would reveal the most pertinent evidence?
- Review confusion pairs against the customer-service glossary
- Compare predicted severity against signed service-impact rules
- Trace each suggested promise to the customer's entitlement and policy
- Inspect match decisions and false-positive entity joins
Explanation
Correct Answer
D:For the case where two systems share company names but not primary keys, the diagnostic path must explain the failure that Copilot attaches one customer's risk note to another account; inspecting match decisions and false-positive entity joins targets the mechanism by which a mapping service resolves stable identities before enrichment.
Incorrect Answers
A:The proposed inspection is associated with agents must distinguish an incident, complaint, inquiry and entitlement, where business-term definitions align user vocabulary with governed service entities; it would not expose the failure mode that Copilot attaches one customer's risk note to another account with the required specificity.
B:The proposed probe concerns a specialized support team uses a nonstandard severity matrix, where custom business rules align AI-assisted triage with actual impact definitions; this incident instead requires evidence about a mapping service resolves stable identities before enrichment and the failure that Copilot attaches one customer's risk note to another account.
C:Replies cite generic guidance instead of contract-specific policy; examining that other path does not determine whether a mapping service resolves stable identities before enrichment, which is what distinguishes the issue in the stated use case.