When AI should stop talking and hand over to a person

The measure of a customer-facing AI is not how long it talks but how well it stops.

The market keeps competing over which assistant sounds more human, and that competition aims at the wrong target. In a customer-facing AI system, the design decision that separates a useful product from a liability is not the opening line; it is knowing when the system must stop talking and hand the matter to a person.

The honest categories are few and learnable: the customer asks outright for a human; the matter carries a legal or financial consequence the system should not decide; the system catches itself guessing; the conversation turns to complaint or distress; or the same question comes back a second time, which usually means the first answer missed. A system that treats these as signals rather than failures was built by people who respected its limits.

Pretending to know is the expensive failure. A confident wrong answer on a consequential question costs more than silence ever did, because the customer rarely announces the damage: they quietly stop trusting the channel, then the company behind it, and neither fact reaches your dashboard.

A handover only works when it carries the conversation over: the full context moves with the customer, nobody repeats their story from zero, one prepared message summarizes the situation, and a named person actually answers. A transfer that forgets the context is not a handover; it is a restart dressed up as service.

A quieter point sits underneath: the handover must be rare by design and honest in form. A system that passes everything to humans is a contact form with better manners; a system that never passes anything is a wall. Restricting it to at most once per conversation keeps the human route meaningful instead of routine.

Our own practice follows the same law: our assistant is transparent that it is AI from the first exchange, it routes matters it should not decide to human colleagues at most once per conversation, and it does so without template filler. We arrived at this after watching what pretense costs when the truth surfaces mid-conversation.

So when we design systems that face customers, ours or our clients', the handover is designed first and never bolted on after the first failure. Guidance, not guarantees: the machine handles what it can verify, and everything carrying a consequence meets a person who carries a name.