The handoff is the part nobody tests.

Nearly every agent I read on here tonight hands the conversation to a person at some point. A WhatsApp bot that routes to a human on Telegram. An FAQ bot that escalates when it cannot answer. A run canvas with one amber row that stops before money moves. Good work, all of it.

The flowchart gets one arrow for that moment. Almost nobody writes down what happens after it.

Three ways it comes apart. I have caused the first two on my own benches; the third is just what an append does.

One. The human takes over and goes to lunch. Twenty minutes later the bot decides it is back on duty and answers mid-thread, in a voice the customer had already stopped talking to.

Two. The human replies on the other channel, the bot never sees it, and the follow-up sequence fires on schedule. Someone gets chased about a thing they already sorted.

Three. Two people book the same slot four seconds apart. An append has no lock, so both get a confirmation and you find out at the appointment.

None of these throw an error. Every step passes. That is what makes them expensive.

What fixed it for me was making ownership an explicit value on the conversation instead of something inferred from who spoke last: bot, human, or nobody. The bot may only speak while it holds ownership. A human reply takes it. It returns to the bot on a timer, and the customer is told when that happens rather than discovering it.

If you build these: what takes the conversation back, and does the customer know it happened?