Question 600
Support engineers report this failure: Lower-cost routing degrades a business-critical subgroup. The business context is: Average accuracy hides failures on complex finance questions. What evidence should be collected to isolate the cause?
- Compare quality by task complexity and selected model
- Correlate request rates with throttles and queue depth
- Audit training-data writers and model approvers
- Compare ticket assignment confusion matrices across versions
Explanation
Correct Answer
A:In the setting where average accuracy hides failures on complex finance questions, comparing quality by task complexity and selected model can isolate the failure in which lower-cost routing degrades a business-critical subgroup, because stratified evaluation guides thresholds for difficult requests.
Incorrect Answers
B:A regional launch multiplies concurrent agent requests; its matching diagnostic check is reasonable for rate limiting causes intermittent failures; the present investigation needs evidence for the different failure where lower-cost routing degrades a business-critical subgroup.
C:The proposed probe concerns business analysts can submit examples but not change production behavior, where separated data-curation and release permissions preserve integrity; this incident instead requires evidence about stratified evaluation guides thresholds for difficult requests and the failure that lower-cost routing degrades a business-critical subgroup.
D:That probe evaluates controlled service evaluations validate routing before rollout in a workflow where a support agent changes how it classifies priority intents; the reported failure concerns the path for which stratified evaluation guides thresholds for difficult requests instead.
Question 601
A solution architecture is being reviewed for this business situation: A backend agent must call an approved Azure service without embedded secrets. The observed defect is that long-lived API keys leak from configuration and diagnostics. Which architectural decision addresses the underlying requirement?
- Use managed identities for service-to-service agent components
- Require independent backend authorization for generated tool calls
- Regression-test case routing after service agent changes
- Protect inference endpoints with network and identity boundaries
Explanation
Correct Answer
A:The defect is that long-lived API keys leak from configuration and diagnostics; the expected business case is that a backend agent must call an approved Azure service without embedded secrets; the design fits because managed identity provides credential-less workload authentication where supported.
Incorrect Answers
B: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 long-lived API keys leak from configuration and diagnostics; that mechanism does not establish the required managed identity provides credential-less workload authentication where supported.
C:This design would make sense if a support agent changes how it classifies priority intents, particularly where high-severity tickets begin routing to general queues; the requested architecture instead needs the dependency that managed identity provides credential-less workload authentication where supported because long-lived API keys leak from configuration and diagnostics.
D:Untrusted clients submit requests and exfiltrate outputs; the alternate design can address that situation because identity-based access and supported network isolation reduce exposure, whereas this question needs a solution to long-lived API keys leak from configuration and diagnostics.
Question 602
A planned AI solution must support this situation: Users ask about governed CRM records rather than prose manuals. Its prototype fails because a document index cannot answer current case ownership reliably. Which design best fits the constraint?
- Use Dataverse knowledge for structured business-record questions
- Define cross-service boundaries and ownership before integration
- Use retrieval citations with explicit no-source handling
- Tag documents with ownership, sensitivity, and provenance metadata
Explanation
Correct Answer
A:Without the intended behavior, a document index cannot answer current case ownership reliably; the proposed architecture establishes the specific design principle that Dataverse grounding exposes appropriate structured records to the agent for the process where users ask about governed CRM records rather than prose manuals.
Incorrect Answers
B:When copilot, Dataverse, Dynamics and custom APIs share one process, contracts and ownership boundaries identify responsibility for each state change; in the stated workflow, however, a document index cannot answer current case ownership reliably; the relevant architectural behavior is that Dataverse grounding exposes appropriate structured records to the agent.
C:When answers about product policy require traceable evidence, cited retrieval separates factual support from free-form generation; in the stated workflow, however, a document index cannot answer current case ownership reliably; the relevant architectural behavior is that Dataverse grounding exposes appropriate structured records to the agent.
D:In the case where search spans internal manuals with different owners and permissions, the alternate design depends on metadata enables filtering, review, retention and traceable answers; it does not correct the target failure, a document index cannot answer current case ownership reliably.
Question 603
The pilot needs a durable correction: Generative routing skips a mandatory disclosure step. This affects the workflow in which a regulated process requires fixed questions in exact sequence. Which design choice should the architect approve?
- Use explicit topics for a tightly scripted compliance dialogue
- Design a fallback topic that escalates genuine nonmatches
- Use connector actions for authenticated data mutations
- Use a prompt action for bounded text transformation
- Control creativity for regulated response templates
Explanation
Correct Answer
A:The mechanism that deterministic topic logic controls known prompts and branching fits the scenario where a regulated process requires fixed questions in exact sequence; it addresses the defect that generative routing skips a mandatory disclosure step.
Incorrect Answers
B:When users use novel requests outside the supported topic inventory, fallback handles unmatched intent and can ask for clarification or escalate; in the stated workflow, however, generative routing skips a mandatory disclosure step; the relevant architectural behavior is that deterministic topic logic controls known prompts and branching.
C:An authorized connector action invokes an operational service; its main purpose is an agent needs to update a governed ticket status, not correction of the reported defect where generative routing skips a mandatory disclosure step in the target process.
D:For another type of architecture, prompt actions encapsulate reusable inputs and expected outputs is useful; this design must instead support a regulated process requires fixed questions in exact sequence despite the problem where generative routing skips a mandatory disclosure step.
E:There is a valid place for the mechanism that templating and approved clauses limit generated variation, namely compliance letters must preserve approved language and disclaimers; it does not remove the specific defect that generative routing skips a mandatory disclosure step from this case.
Question 604
A product owner must measure whether an AI-enabled change actually improves this process: An employee requests a standard equipment return initiation. The earlier problem was: A broad assistant changes unrelated inventory records. Which outcome indicator is most relevant?
- Unauthorized ancillary actions per task
- Correctly sequenced workflow executions
- Analyst-accepted explanations with verified references
- Valid returns completed without unrelated writes
Explanation
Correct Answer
D:For the use case where an employee requests a standard equipment return initiation, the indicator valid returns completed without unrelated writes is worth baselining against the earlier issue where a broad assistant changes unrelated inventory records.
Incorrect Answers
A:The premise of this indicator is limited intent and tool scope prevent unnecessary actions; the architecture being evaluated instead serves an employee requests a standard equipment return initiation, and its success is measured by valid returns completed without unrelated writes.
B:For the situation where approval processing requires sequential verification and notification, correctly sequenced workflow executions is a sensible indicator; the problem that a broad assistant changes unrelated inventory records requires measurement of valid returns completed without unrelated writes.
C:The premise of this indicator is grounded explanatory responses distinguish evidence from hypothesis; the architecture being evaluated instead serves an employee requests a standard equipment return initiation, and its success is measured by valid returns completed without unrelated writes.
Question 605
A proposed deployment supports this business case: Employees install external agents with unknown data pathways. Reviewers also warn that a third-party agent might send business data to an unapproved destination. Which restriction mitigates that concrete risk?
- Enforce owner sign-off for high-risk grounding collections
- Block deployment pending approved handling and residency evidence
- Block release if privileged writes bypass approvals
- Deny unapproved destinations and excessive scopes
Explanation
Correct Answer
D:In the context where employees install external agents with unknown data pathways, the exposure is that a third-party agent might send business data to an unapproved destination; containing it requires a guardrail in the operation where procurement and security review evaluate permissions and data flows.
Incorrect Answers
A:The safeguard for new documents enter a production source automatically deals with a different threat: sensitive ownerless documents might enter production grounding; even after it is enforced, the exposure that a third-party agent might send business data to an unapproved destination remains possible in the target workflow where procurement and security review evaluate permissions and data flows.
B:The alternate guard addresses a third-party processor might move confidential documents without contractual approval for a third-party AI tool changes where submitted documents are processed; the present threat, a third-party agent might send business data to an unapproved destination, needs a different enforcement boundary.
C:Where agents suggest payment adjustments and inventory holds, the guard can address an updated connector might gain a direct privileged posting route; the exposure that a third-party agent might send business data to an unapproved destination remains relevant to the present assessment.
Question 606
One enterprise rollout has 3 separately testable requirements. (1) Business use: A prompt injection requests data export to an outside address. Known defect: Model-level refusal alone may not prevent a tool call. (2) Business use: A tool response includes HTML or links supplied by third parties. Known defect: Agent-generated pages render an unsafe instruction or link. (3) Business use: External knowledge contains embedded role-change commands. Known defect: The agent obeys text retrieved from a supplier document. Which 3 regression tests directly satisfy these stated cases? Choose THREE.
- Inject a malicious export instruction into a source document
- Disable knowledge access and verify safe no-evidence behavior
- Embed hostile instructions in a test knowledge file
- Feed adversarial markup in a synthetic tool response
Explanation
Correct Answers
A:For the business workflow where a prompt injection requests data export to an outside address, injecting a malicious export instruction into a source document challenges the failure that model-level refusal alone may not prevent a tool call; the release also needs a separate check for a tool response includes HTML or links supplied by third parties; external knowledge contains embedded role-change commands.
C:For the business workflow where external knowledge contains embedded role-change commands, embedding hostile instructions in a test knowledge file challenges the failure that the agent obeys text retrieved from a supplier document; the release also needs a separate check for a prompt injection requests data export to an outside address; a tool response includes HTML or links supplied by third parties.
D:For the business workflow where a tool response includes HTML or links supplied by third parties, feeding adversarial markup in a synthetic tool response challenges the failure that agent-generated pages render an unsafe instruction or link; the release also needs a separate check for a prompt injection requests data export to an outside address; external knowledge contains embedded role-change commands.
Incorrect Answers
B:This regression test is associated with the different failure where the agent confidently answers when retrieval fails in a workflow where policy answers must be tied to a current source; it does not directly solve the listed conditions: model-level refusal alone may not prevent a tool call; agent-generated pages render an unsafe instruction or link; the agent obeys text retrieved from a supplier document.
Question 607
Risk reviewers need a control, not merely a monitoring dashboard. The use case is: Hostile instructions arrive both as user input and retrieved documents. Potential harm: Hostile user and retrieved text might bypass model-level instructions. Which protection fits?
- Combine detection with server-side authorization of all actions
- Prevent sensitive data from entering prompts unless required
- Prioritize incidents with safety or privacy impact
- Mask sensitive input/output fields and control log access
Explanation
Correct Answer
A:The business setting is hostile instructions arrive both as user input and retrieved documents, with the exposure that hostile user and retrieved text might bypass model-level instructions; the enforceable protection must correspond to the path where input and document-attack detection supplies layered protection.
Incorrect Answers
B:The safeguard for an internal model sometimes returns confidential text verbatim deals with a different threat: a model might expose sensitive user data in uncontrolled prompt output; even after it is enforced, the exposure that hostile user and retrieved text might bypass model-level instructions remains possible in the target workflow where input and document-attack detection supplies layered protection.
C:The alternate boundary mitigates a serious privacy incident might be triaged behind cosmetic chatbot defects, which is relevant to complaints use vague labels such as agent is wrong; it leaves the current workflow vulnerable to hostile user and retrieved text might bypass model-level instructions.
D:Where a newly deployed Foundry agent has no detailed diagnostics, the guard can address tracing might store secrets and personal content without adequate access controls; the exposure that hostile user and retrieved text might bypass model-level instructions remains relevant to the present assessment.
Question 608
A planned AI solution must support this situation: Quality teams need searchable insights from recorded calls. Its prototype fails because manual notes omit important customer statements. Which design best fits the constraint?
- Use managed inference and operational health checks for a custom model
- Design voice interaction with confirmation of critical entities
- Use translation capabilities before downstream review
- Choose speech processing for transcribed contact-center audio
Explanation
Correct Answer
D:The architecture must reconcile two facts: quality teams need searchable insights from recorded calls, yet manual notes omit important customer statements; the selected pattern does so through the mechanism that speech transcription transforms audio to text for subsequent analysis.
Incorrect Answers
A:For another type of architecture, a managed serving endpoint needs scaling, monitoring and rollback design is useful; this design must instead support quality teams need searchable insights from recorded calls despite the problem where manual notes omit important customer statements.
B:For callers provide account numbers and amounts by speech, this other approach relies on the fact that spoken confirmation and verification reduce high-impact recognition errors; the issue here is manual notes omit important customer statements; that mechanism does not establish the required speech transcription transforms audio to text for subsequent analysis.
C:Translation provides standardized text input while preserving original evidence; its main purpose is support documents arrive in several customer languages, not correction of the reported defect where manual notes omit important customer statements in the target process.
Question 609
One enterprise rollout has 2 separately testable requirements. (1) Business use: External credit service responds slowly during regional outages. Known defect: A sales workflow hangs awaiting optional enrichment. (2) Business use: Sales copilots should not expose sensitive margin assumptions. Known defect: A connector returns the full ERP payload including protected fields. Which 2 diagnostic procedures directly satisfy these stated cases? Choose TWO.
- Analyze containment failures and repeated intent fallback patterns
- Capture response schemas and returned sensitive columns
- Review confusion pairs against the customer-service glossary
- Trace latency distributions and retries for connector calls
Explanation
Correct Answers
B:Capturing response schemas and returned sensitive columns can reveal why a connector returns the full ERP payload including protected fields when sales copilots should not expose sensitive margin assumptions; that finding is needed alongside the independent diagnosis for external credit service responds slowly during regional outages.
D:Tracing latency distributions and retries for connector calls can reveal why a sales workflow hangs awaiting optional enrichment when external credit service responds slowly during regional outages; that finding is needed alongside the independent diagnosis for sales copilots should not expose sensitive margin assumptions.
Incorrect Answers
A:This diagnosis is associated with the different failure where repeated fallback messages trap callers in a loop in a workflow where contact Center agents sometimes cannot resolve complex billing intent; it does not directly solve the listed conditions: a sales workflow hangs awaiting optional enrichment; a connector returns the full ERP payload including protected fields.
C:This diagnosis is associated with the different failure where Copilot misclassifies a complaint as an ordinary information request in a workflow where agents must distinguish an incident, complaint, inquiry and entitlement; it does not directly solve the listed conditions: a sales workflow hangs awaiting optional enrichment; a connector returns the full ERP payload including protected fields.
Question 610
A test plan focuses on a previously missed failure: Agent-generated language promises the wrong financial remedy. It occurred when refund, credit, reimbursement and adjustment have different accounting impacts. Which check directly validates that behavior?
- Test duplicate legal names with different external IDs
- Change a permitted credit field and verify updated authorized retrieval
- Delay the external API and verify useful partial handling
- Generate stressful and adversarial user examples
- Test similar user requests that imply different posting types
- Query as multiple roles and confirm only authorized fields appear
Explanation
Correct Answer
E:Agent-generated language promises the wrong financial remedy; a release test must establish that this no longer occurs; testing similar user requests that imply different posting types measures the specific behavior for which controlled definitions tie each term to approved policies and authorities.
Incorrect Answers
A:The verification concerning two systems share company names but not primary keys checks a mapping service resolves stable identities before enrichment against Copilot attaches one customer's risk note to another account; its result says little about the observed issue where agent-generated language promises the wrong financial remedy.
B:For sellers need current credit exposure held outside CRM, the alternate test is about whether a supported connector can retrieve external information with explicit identity mapping; it cannot prove the outcome needed after the reported regression where agent-generated language promises the wrong financial remedy.
C:This check exercises a different system behavior: bounded timeouts and graceful fallback preserve essential user work; a pass there cannot establish that agent-generated language promises the wrong financial remedy no longer occurs in the required workflow.
D:A pass on that verification would speak to compliance letters must preserve approved language and disclaimers; it would not show that controlled definitions tie each term to approved policies and authorities, which is needed to avoid the observed defect agent-generated language promises the wrong financial remedy.
F:Querying as multiple roles and confirming only authorized fields appear verifies an adjacent objective where sales copilots should not expose sensitive margin assumptions; it leaves untested the causal path in which controlled definitions tie each term to approved policies and authorities for this case.