A conversational agent cannot safely treat manual takeover as a temporary frontend toggle. The system needs to know who owns the conversation, which actions are still allowed, and how automation may resume.
The state model
A useful handoff model records:
- current ownership;
- who initiated the takeover;
- when automated sending stopped;
- whether background classification may continue;
- pending customer or operator actions;
- the conditions required to return control.
This state belongs on the server. A browser tab can close, an operator can switch devices, and multiple staff members can view the same conversation.
Avoiding split-brain replies
The dangerous failure is not that automation stops. It is that automation and a human both believe they are authorised to reply.
The send path therefore checks ownership at execution time, not only when the interface rendered. Human takeover must remain authoritative across queued work, retries, and delayed provider callbacks.
The interface then exposes the real state: who owns the thread, why automation is paused, and what will happen when control is released.