Knowing when to stop and hand over. The trigger should be defined in advance rather than left to the model, and the handover needs to carry the context, because nothing annoys a person more than being asked to repeat everything they just told the machine.
Technical view
Triggers include low confidence, detected sensitive categories such as distress or complaint, repeated failure, explicit user request, and high-stakes action classes. The handover should transfer conversation history, a summary of what was attempted, and any partial work. Measure escalation rate and post-escalation resolution, since a system optimised only for deflection will suppress escalations that should have happened, which is the most common failure in customer service deployments.
Related
Part of AI Terminology Index, section: Product and Adoption.