Question 70
The architect wants evidence before altering the deployment. In a workflow where answers depend on policies revised every week, a trained model remembers obsolete guidance after source changes. What should the team examine first?
- Compare gap size to supported extension interfaces
- Inspect connector crawl scope, schema and ACL mappings
- Analyze route decisions by task class and observed error
- Compare errors for historical and newly effective rules
Explanation
Correct Answer
D:The reported fault is that a trained model remembers obsolete guidance after source changes in a system where answers depend on policies revised every week; evidence from comparing errors for historical and newly effective rules can show whether retrieval separates volatile knowledge from model parameters.
Incorrect Answers
A:Comparing gap size to supported extension interfaces explores a separate connection, where extension retains familiar user experience and adds bounded custom behavior; the diagnosed requirement is answers depend on policies revised every week, and the actual defect is a trained model remembers obsolete guidance after source changes.
B:In the separate scenario where teams want approved external records searchable in Copilot, the same investigation could explain the agent lacks knowledge of non-Microsoft content; it does not isolate the reported cause in a system where answers depend on policies revised every week.
C:In the process where many calls require only a small fixed-label output, the diagnostic procedure could explain universal use of a premium model inflates cost; this particular incident instead turns on whether retrieval separates volatile knowledge from model parameters and why a trained model remembers obsolete guidance after source changes.
Question 71
A quality engineer is writing a regression case after the incident where its knowledge source can explain but cannot commit a database change. The underlying business scenario is: An agent needs to update a governed ticket status. What should the test exercise?
- Ask for a variance explanation with incomplete data
- Install the agent to a pilot Microsoft 365 audience
- Change a test case status and inspect audit events
- Test borderline utterances against both descriptions
Explanation
Correct Answer
C:Changing a test case status and inspecting audit events challenges the path where an authorized connector action invokes an operational service for the business setting an agent needs to update a governed ticket status; the specific defect to exclude is that its knowledge source can explain but cannot commit a database change.
Incorrect Answers
A:For finance asks why a variance was flagged, the alternate test is about whether grounded explanatory responses distinguish evidence from hypothesis; it cannot prove the outcome needed after the reported regression where its knowledge source can explain but cannot commit a database change.
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 its knowledge source can explain but cannot commit a database change.
D:A pass on that verification would speak to two topics claim nearly identical scheduling intents; it would not show that an authorized connector action invokes an operational service, which is needed to avoid the observed defect its knowledge source can explain but cannot commit a database change.
Question 72
A planned AI solution must support this situation: Executives request AI before defining an operational improvement target. Its prototype fails because prototype activity increases but no business value is measured. Which design best fits the constraint?
- Provide a unified approval handoff for AI-identified account risks
- Start with documented business outcomes and adoption readiness
- Establish a product-oriented operating model for AI services
- Build a bespoke agent when differentiated orchestration is essential
- Use a custom model only when general models fail a defined task
Explanation
Correct Answer
B:Without the intended behavior, prototype activity increases but no business value is measured; the proposed architecture establishes the specific design principle that cloud adoption planning begins from motivations, outcomes and readiness constraints for the process where executives request AI before defining an operational improvement target.
Incorrect Answers
A:A suggested discount bypasses delegated approval authority; the alternate design can address that situation because cross-app workflow routes the recommendation to the accountable approver, whereas this question needs a solution to prototype activity increases but no business value is measured.
C:Long-lived product ownership funds ongoing evaluation and operations; its main purpose is an agent will evolve continuously after its initial project delivery, not correction of the reported defect where prototype activity increases but no business value is measured in the target process.
D:The alternative addresses a boxed product cannot express the execution and ownership rules by a design in which custom engineering trades more control for more operating responsibility; it does not supply the missing relationship for executives request AI before defining an operational improvement target, where prototype activity increases but no business value is measured.
E:The technical behavior of this alternative is that a measured capability gap, appropriate dataset and MLOps plan justify customization; the user case being assessed is executives request AI before defining an operational improvement target with a different failure: prototype activity increases but no business value is measured.
Question 73
Implementation alternatives must be compared for a workflow where standard capabilities draft customer communications but policy requires approval. The pilot shows this gap: Automatically generated notices are sent without required review. Which alternative should lead the solution?
- Retain process-specific human review alongside prebuilt automation
- Use Dataverse as the governed handoff for customer cases
- Adopt a supported prebuilt agent for standard productivity workflows
- Require approval before irreversible agent actions
Explanation
Correct Answer
A:The architecture must reconcile two facts: standard capabilities draft customer communications but policy requires approval, yet automatically generated notices are sent without required review; the selected pattern does so through the mechanism that a review step preserves accountability without discarding assisted drafting.
Incorrect Answers
B:The technical behavior of this alternative is that shared authorized records support coordinated business context; the user case being assessed is standard capabilities draft customer communications but policy requires approval with a different failure: automatically generated notices are sent without required review.
C:For users need ordinary meeting summaries and follow-up drafts, this other approach relies on the fact that a prebuilt Microsoft 365 capability minimizes custom maintenance for commodity tasks; the issue here is automatically generated notices are sent without required review; that mechanism does not establish the required a review step preserves accountability without discarding assisted drafting.
D:The technical behavior of this alternative is that confirmation and human approval separate advice from execution authority; the user case being assessed is standard capabilities draft customer communications but policy requires approval with a different failure: automatically generated notices are sent without required review.
Question 74
The architect wants evidence before altering the deployment. In a workflow where the orchestrator chooses between similar invoice and credit tools, it passes customer IDs into supplier inputs. What should the team examine first?
- Compare targeted label accuracy and processing cost
- Inspect outbound payloads and role-based access decisions
- Inspect transcript alignment and transcription errors
- Review tool selection traces and parameter-binding errors
Explanation
Correct Answer
D:It passes customer IDs into supplier inputs; to attribute the defect rather than guess at a fix, reviewing tool selection traces and parameter-binding errors examines the dependency where specific tool metadata helps the planner choose and parameterize actions.
Incorrect Answers
A:A full agent workflow adds cost to simple text tagging; examining that other path does not determine whether specific tool metadata helps the planner choose and parameterize actions, which is what distinguishes the issue in the stated use case.
B:Inspecting outbound payloads and role-based access decisions explores a separate connection, where server-side authorization and minimized inputs protect data flow; the diagnosed requirement is the orchestrator chooses between similar invoice and credit tools, and the actual defect is it passes customer IDs into supplier inputs.
C:Quality teams need searchable insights from recorded calls; its matching diagnostic check is reasonable for manual notes omit important customer statements; the present investigation needs evidence for the different failure where it passes customer IDs into supplier inputs.
Question 75
The architecture board has evidence that the agent follows lower-trust text as an order. The affected situation is: A retrieved email contains instructions to bypass approval. What change most directly establishes the missing behavior?
- Test prompt-injection resistance across trust boundaries
- Test negative and adversarial cases alongside happy paths
- Require independent backend authorization for generated tool calls
- Test structured outputs against explicit format contracts
Explanation
Correct Answer
A:Because adversarial fixtures verify separation of sources and instructions, this approach corrects the reported breakdown where the agent follows lower-trust text as an order; the affected situation is that a retrieved email contains instructions to bypass approval.
Incorrect Answers
B:In the case where benchmarks cover only cooperative user wording, the alternate design depends on negative tests expose policy and validation boundaries; it does not correct the target failure, the agent follows lower-trust text as an order.
C:For prompt wording persuades the agent to submit a privileged operation, this other approach relies on the fact that server-side access checks enforce real actor entitlements; the issue here is the agent follows lower-trust text as an order; that mechanism does not establish the required adversarial fixtures verify separation of sources and instructions.
D:This design would make sense if a Copilot prompt feeds records into downstream automation, particularly where generated output occasionally omits required keys; the requested architecture instead needs the dependency that adversarial fixtures verify separation of sources and instructions because the agent follows lower-trust text as an order.
Question 76
A release must prove that agent invokes the wrong booking action for a reschedule will not recur. The application is used in this situation: Two topics claim nearly identical scheduling intents. Which acceptance test most directly checks that failure mode?
- Submit missing and malformed values and check validation behavior
- Request drafts using multiple brief formats
- Test borderline utterances against both descriptions
- Create an escalation from a test case and inspect audit records
Explanation
Correct Answer
C:Agent invokes the wrong booking action for a reschedule; a release test must establish that this no longer occurs; testing borderline utterances against both descriptions measures the specific behavior for which distinct descriptions and inputs improve topic selection.
Incorrect Answers
A:In a solution where inspectors submit a short description and need suggested categorization, submitting missing and malformed values and checking validation behavior could catch the defect that free-form AI output is written to the wrong Dataverse field; the current release gate must challenge the failure where agent invokes the wrong booking action for a reschedule.
B:The suggested test serves marketing specialists need tailored descriptions from existing briefings, in which a broad action agent can modify records unnecessarily; the present scenario instead concerns the reported problem where agent invokes the wrong booking action for a reschedule.
D:The verification concerning a representative must initiate an exception workflow from the case checks an authorized action creates a governed workflow item from the current case against AI suggestions are never converted into traceable escalation records; its result says little about the observed issue where agent invokes the wrong booking action for a reschedule.
Question 77
Before introducing a fix, the team must explain why unvetted labeled data poisons a retrained classifier. The affected setting is: Business analysts can submit examples but not change production behavior. What should be inspected?
- Reconcile active model versions with promotion events
- Review endpoint access logs and allowed network paths
- Audit training-data writers and model approvers
- Match release manifest to transformation and model checksums
Explanation
Correct Answer
C:In the setting where business analysts can submit examples but not change production behavior, auditing training-data writers and model approvers can isolate the failure in which unvetted labeled data poisons a retrained classifier, because separated data-curation and release permissions preserve integrity.
Incorrect Answers
A:That probe evaluates deployment audit links actor, approved artifact and release time in a workflow where a model is replaced without a traceable change ticket; the reported failure concerns the path for which separated data-curation and release permissions preserve integrity instead.
B:That probe evaluates identity-based access and supported network isolation reduce exposure in a workflow where a sensitive model endpoint is reachable using a broadly shared key; the reported failure concerns the path for which separated data-curation and release permissions preserve integrity instead.
D:In the process where a preprocessing release changes category labels incorrectly, the diagnostic procedure could explain restoring code alone does not restore previous predictions; this particular incident instead turns on whether separated data-curation and release permissions preserve integrity and why unvetted labeled data poisons a retrained classifier.
Question 78
Before introducing a fix, the team must explain why a failing agent shows only a generic HTTP error. The affected setting is: Operations need to correlate prompts, model calls and tools in a Foundry workflow. What should be inspected?
- Compare intended task with executed tool names and inputs
- Inspect model and tool spans under one operation identifier
- Compare query, filters and top retrieved source identifiers
- Review critical-path durations for slow requests
Explanation
Correct Answer
B:Where a failing agent shows only a generic HTTP error, the useful evidence comes from inspecting model and tool spans under one operation identifier; its focus matches the architectural mechanism that agent trace spans provide component-level execution evidence through connected telemetry in the process where operations need to correlate prompts, model calls and tools in a Foundry workflow.
Incorrect Answers
A:For the failure that final answer text conceals the erroneous tool choice, comparing intended task with executed tool names and inputs is relevant; this question instead reports that a failing agent shows only a generic HTTP error, so different evidence is needed.
C:The proposed inspection is associated with an assistant responds fluently but cites unrelated support articles, where retrieval telemetry distinguishes evidence search from answer generation; it would not expose the failure mode that a failing agent shows only a generic HTTP error with the required specificity.
D:For the failure that engineers blame generation although the backend is slow, reviewing critical-path durations for slow requests is relevant; this question instead reports that a failing agent shows only a generic HTTP error, so different evidence is needed.
Question 79
In a business situation where a third-party AI tool changes where submitted documents are processed, a previous build showed this defect: Approved data crosses an unreviewed processor. Which reproducible test belongs in the release gate?
- Recreate an evaluation from the recorded dataset snapshot
- Execute an authorized and a denied deletion in staging
- Test synthetic flows and compare actual destinations
- Roll back a staged faulty transformation and compare output
Explanation
Correct Answer
C:The historical failure is that approved data crosses an unreviewed processor; for the setting where a third-party AI tool changes where submitted documents are processed, testing synthetic flows and comparing actual destinations verifies the dependency where contract and privacy assessment cover service location and subprocessor use.
Incorrect Answers
A:This check exercises a different system behavior: immutable snapshots and lineage tie runs to approved sources; a pass there cannot establish that approved data crosses an unreviewed processor no longer occurs in the required workflow.
B:The suggested test serves a tool deletes a record following an ambiguous conversation, in which logs show deletion but not the requesting principal or approval; the present scenario instead concerns the reported problem where approved data crosses an unreviewed processor.
D:Although the exercise might catch restoring code alone does not restore previous predictions, it would not directly validate the mechanism that contract and privacy assessment cover service location and subprocessor use in the target implementation.
Question 80
A business stakeholder describes this operating situation: An internal helper produces attractive but unsupported policy answers. Testing instead shows that satisfaction scores hide factual mistakes. What design approach is most appropriate?
- Integrate runtime tracing and alerting into release readiness
- Inspect retrieval spans when cited answers omit evidence
- Measure tool success, retries and side-effect duplication
- Measure grounded factual accuracy on a curated corpus
Explanation
Correct Answer
D:The defect is that satisfaction scores hide factual mistakes; the expected business case is that an internal helper produces attractive but unsupported policy answers; the design fits because expert-checked evaluation examples quantify evidence fidelity.
Incorrect Answers
A:The alternative addresses the on-call team cannot explain failed tool calls by a design in which connected observability supplies traces and operational metrics; it does not supply the missing relationship for an internal helper produces attractive but unsupported policy answers, where satisfaction scores hide factual mistakes.
B:Surface quality scores appear healthy despite wrong documents; the alternate design can address that situation because retrieval telemetry distinguishes evidence search from answer generation, whereas this question needs a solution to satisfaction scores hide factual mistakes.
C:The technical behavior of this alternative is that tool metrics reveal mutation failures and retry consequences; the user case being assessed is an internal helper produces attractive but unsupported policy answers with a different failure: satisfaction scores hide factual mistakes.