When intelligent customer service has no basis, it is necessary to let humans continue to deal with the same service problem. Users should not re-describe the entire context, and manual handovers should distinguish between facts provided by users and suggestions generated by the model. The handover content should be clear and traceable.

Transfer conditions

Set manual transfer portals based on problem types, missing data and user requests to allow manual takeover. In order to reduce the number of transfers, the model cannot continue to answer questions beyond the scope of knowledge.

Context handover

Provide necessary questions, displayed answers and cited sources to humans to reduce repeated inquiries. Handover data is still subject to access restrictions, and irrelevant historical records cannot be attached to the current service.

Continuous improvement

The results of manual processing can be used as clues for knowledge updating, but the content and scope of authorization must be verified before use. Customer feedback is used to identify issues and should not go directly into the formal knowledge base without inspection.

Scenario verification

The handover test can select problems for which the model cannot find a basis, check how the user transfers to manual, and whether the manual can see the necessary context. After the transfer is completed, subsequent replies and task records should still be attributed to the same service process.

Specific verification method

Raise issues that are not covered by the knowledge base and check that the manual entrance is available; the person taking over reads the necessary conversations and unresolved issues; after the transfer, manual replies continue to be associated with the original conversation; repeated clicks do not generate multiple independent work orders, and users can also know the current waiting or connected status.

Transfer to manual connection for the complete service process, highlighting problems, existing information and content to be confirmed. The source of the model and manual responses are retained so that subsequent posts can know which information has been verified and which still needs to be processed.