The handoff: teach your agent when to get a human
Most AI support failures aren't wrong answers. They're an agent bluffing its way through a question it should have handed over. Whichever assistant you use, the handoff is the cheapest reliability feature you can add, and it takes an afternoon.
- Write the stop list. Refunds, complaints, legal or medical questions, pricing exceptions, and anyone who simply asks for a human. Anything on that list ends the agent's turn — no exceptions, no judgement calls.
- Add the honesty rule. If the answer isn't in the material you gave it, it says so and hands over. "I don't know — let me get Anna" beats a confident invention every single time.
- Route it somewhere a person actually watches. A shared inbox, Slack, WhatsApp, Telegram. On my own intake bot every client message mirrors to my phone; I reply, the agent pauses itself, and I have the thread.
- Tell the customer what happens next, with a time. "Nikolai will reply here within a few hours." Silence after a handoff does more damage than a bad answer.
- Log the reason every time. Each escalation names a gap: a missing price list, a policy nobody ever wrote down. Fix one a week and the handoff rate falls on its own.
An agent that hands over well earns trust faster than one that tries to answer everything. Aim for a low handoff rate — never zero.