Handing off to a person, well

A handoff is the moment an agent stops owning a conversation or case and a human starts. Done well, the customer barely notices the seam: they are told who is taking over and when, they never repeat themselves, and the human starts with a summary, the facts, what was already tried and said, and a recommended next step. Done badly, it is where trust in the whole deployment dies: dead ends, loops back to the bot, "I'm just an AI", a coordinator asking for the container number the customer typed four messages ago. Design it as a system: an escalation policy that fires on explicit requests, policy boundaries, high value, repeated failure, low confidence and frustration; a typed packet generated from state, not from the model's memory; skills-based routing with hours and fallbacks; a warm transfer on live channels and a clear expectation on asynchronous ones; a return path that pins the human's decisions; and metrics that reward appropriate escalation, not just containment.

Key takeaways:

• Gate: a human approves an action and the agent continues. Handoff: a human takes over the conversation or case.
• Explicit requests for a human are always honoured. Policy-boundary topics escalate by rule, not by model judgement.
• The packet is built from structured state and evidence. If a fact is not in state, the packet says it is missing.
• Every handoff tells the customer who, when and that they will not need to repeat themselves, in their language and channel.
• Measure escalation precision and recall and the re-ask rate. High containment with low recall is a hidden failure.