When should an AI agent hand over to a human?

An AI agent should hand over to a human on money, on anger, on anything it has already tried once and failed, and on any request it has no tool to fulfil. Those four are not judgement calls and should not be configurable down to nothing. Everything past them is a decision about your business that you should make in writing before the agent goes live, because a handover rule nobody wrote down is a rule the agent does not have.

The four rules that are not negotiable

Money. Payments, refunds, discounts and price negotiation. Not because the model would get the arithmetic wrong, but because the cost of being wrong is a transaction you have to reverse and a customer who now distrusts the channel. In Apavin these sit behind a per-tool risk level with a human on the other side, so the agent gives the customer your own booking or payment link and hands the conversation over rather than acting.

Anger. A frustrated customer wants to be heard by someone who can decide something. An agent that responds to a complaint with a correct, policy-accurate paragraph has technically answered and practically made it worse. Detect the tone, hand over, and do it on the first message rather than the third.

The second attempt. If the agent has checked the knowledge base once and the answer is not there, checking again is theatre. One attempt, then a person. The failure mode this prevents is the loop every customer has been trapped in, where a system rephrases its inability to help four different ways.

Anything it cannot actually do. If the agent has no tool to cancel an order, a request to cancel an order is a handover, not a conversation. This one is easy to get wrong in a way that looks fine in testing, because a model asked to do something impossible will often produce a helpful-sounding sentence instead of failing. Apavin checks the drafted reply for claims of completed actions before it sends, and hands over when it finds one.

What about the judgement calls?

Write them down as a list of named situations, not as a sensitivity dial.

A threshold slider looks like control and is not one. "Hand over below 70% confidence" means nothing operationally, because the model's confidence is not calibrated against your business risk, and nobody in the room can say what 70 means.

A list, on the other hand, can be argued about and then agreed: enterprise accounts always get a person; anything mentioning a competitor by name goes to sales; anything about a delivery that is already late goes to whoever handles logistics. Those are real rules. New people can learn them, and you can tell whether they were followed.

Should the customer be told?

Yes, in one line, and without ceremony.

There is a version of this that is worse than saying nothing, which is the handover framed as an upgrade: let me connect you with a specialist. Customers read that correctly as the system having given up, and the word specialist sets an expectation the next person has to meet.

A line that says a colleague is picking this up, with a rough idea of when, is enough. Apavin's agent also does not claim to be a person. Asked directly, it says it works for the business, which is true and costs nothing.

What does a good handover carry with it?

Everything the person needs to avoid asking a question the customer already answered.

The full thread, threaded by customer rather than by channel, so someone who arrived by WhatsApp on Tuesday and email on Thursday is one history and not two strangers. The reason for the handover, in a sentence. The draft the agent was going to send, if there was one, because a rejected draft tells the person what the agent thought the answer was. And the knowledge base passages it retrieved, because if the right answer was in there and the agent still handed over, that is a retrieval problem worth knowing about.

A handover that arrives as a bare conversation with no context is a handover that costs the customer a repeat of everything they just typed, which is the exact experience the agent was supposed to prevent.

How do you know the rule is working?

Watch two numbers in opposite directions.

The handover rate on its own tells you nothing. Ninety percent might be correct for a business selling machinery and far too high for one answering delivery questions.

What matters is what happens after. Of the conversations the agent handled alone, how many came back within 48 hours with the same question? That is the false negative rate, and it is the expensive one, because the customer has now had a bad experience and is having a second conversation about it.

Of the conversations it handed over, how many did the person resolve by sending something the agent could have sent? That is the false positive rate, and it is cheap. It costs a few minutes. Pay it.

Given the choice, set the line so that you are paying for false positives. A handover that was not needed costs one person five minutes. A handover that was needed and did not happen costs a customer.

The takeaway

Write the four hard rules into the system, write the judgement calls into a list your team agrees on, and check the return rate every month. The line is not something you set once at launch. It is something you move deliberately, in one direction at a time, with a number in front of you.