Question 320
A business agent creates a refund whenever a user writes Please refund without checking entitlement and approval. Which Power Platform extension or architecture change BEST addresses the core requirement?
Explanation
Correct Answer
Evaluate agent tool-choice quality with representative test cases and require human approval for high-impact data writes. This choice implements the controlling platform requirement rather than masking the symptom. LLM reasoning is probabilistic; prompt-only prohibitions are not a substitute for enforced tool permissions. Verify the change using: Inspect evaluator cases, tool traces, policy refusal tests, human approval status, error rates and rollback evidence.
Incorrect Answers
Use the Dataverse Web API with logical names, $select, encoded $filter and @odata.nextLink pagination instead of retrieving unrestricted rows. That action addresses Dataverse OData Web API select filter and pagination instead. A page-size hint or first result set does not guarantee the entire data collection has been returned. For this case, the targeted action is: Evaluate agent tool-choice quality with representative test cases and require human approval for high-impact data writes.
Use expressions such as coalesce, empty, formatDateTime and guarded JSON parsing with explicit type assumptions. That action addresses Power Automate expressions null coalesce and safe parsing instead. A missing property is not always equivalent to an empty string; expression behavior depends on data type. For this case, the targeted action is: Evaluate agent tool-choice quality with representative test cases and require human approval for high-impact data writes.
Register a signed plug-in assembly, target message, step mode/stage and image configuration with supported tooling and deploy it in a solution. That action addresses Plug-in Registration Tool and signed assembly deployment instead. Publishing a DLL without a registered step does not automatically attach code to Dataverse messages. For this case, the targeted action is: Evaluate agent tool-choice quality with representative test cases and require human approval for high-impact data writes.

