Most AI agent setups focus on capability: what it knows, which systems it's connected to, what actions it can take. In practice, what decides whether customers trust it is the opposite — how quickly it admits it isn't the right one to handle the request.
There are four situations in which an agent should hand over to a person without trying another round: when the customer explicitly asks; when they express frustration or anger; when the action is irreversible, such as a cancellation or refund; and when the agent couldn't find an answer in the organizational knowledge.
The fourth case is the one agents fail on most. An agent instructed to help at any cost will invent a plausible-looking answer, and the customer will act on it. Better an agent that says it has no answer and hands over than one that sounds confident and is wrong.
The handover itself is part of the experience. A representative who picks up a conversation without context will ask the customer to repeat themselves, and that's the moment the whole advantage is lost. The handover must include a summary of what was said, what the agent already checked and what it promised.
In the first month it's worth going through every handed-over conversation once a week. Every recurring question the agent couldn't answer becomes an item on the improvement list, and the handover rate should drop gradually — not because the agent guesses more, but because it knows more.
The bottom line
A good agent isn't measured by the share of conversations it closes but by the share it closes correctly.