A digital assistant can answer frequently asked questions and orient a user. The path becomes fragile when the system continues to propose generic responses to a problem it cannot understand. The possibility of involving a person must be designed from the beginning.
I propose to define scope of service and situations that require a passage. A complex request, complaint, or missing information may necessitate another channel. The system should explain the available path without inventing times or commitments that the organization does not support.
A hypothetical example is a customer who reports an order other than the one received. Repeating general delivery instructions can increase frustration. The service collects the necessary information and makes it available to the authorized contact person, under the applicable conditions.
The assessment should include unresolved cases and repeated requests, in addition to the number of conversations. A closed dialogue does not prove that the need has been met. I propose tests using realistic questions and content updates. The NIST AI RMF offers a general reference on risks; the path described here remains an organizational choice to be adapted to the actual capabilities of the system.