Define a bounded task
Specify approved knowledge sources, supported topics and actions the assistant may take. Keep sensitive decisions and unsupported requests on a clear human path.
Build a useful evaluation set
Include ordinary questions, incomplete context, conflicting information and requests outside scope. Record the expected behavior, not only a preferred sentence.
Check permissions and actions
Test whether the assistant can reveal another customer's data or perform an action without appropriate authority. Tool access must enforce the same boundaries as the rest of the application.
Repeat after change
Models, prompts and knowledge can change behavior. Re-run the evaluation and monitor real failures, with a way to restrict or disable the assistant if needed.