Every chatbot demo goes the same way: the bot answers everything. Nobody demos the moment it steps back.
And yet, after watching a lot of real conversations, we keep arriving at the same conclusion: an AI agent's finest moment is not when it answers well. It is when it recognises that answering is no longer its job.
The three signals to step aside
There are three situations where a bot should raise its hand and ask for the substitution.
The technical question that leaves the script. The customer asks something that is not in the content the AI works from — an odd compatibility, an edge case, an installation detail that depends on seeing the property. The right response is not to improvise. It is to say "someone on the team needs to look at this" and pass it on.
The annoyed customer. When someone writes in frustrated — a job that overran, a quote that does not add up, a promise somebody made on the phone — the last thing they want is to negotiate with a machine. Synthetic empathy irritates more than silence does. Here the bot adds value by stepping aside quickly, not by persisting.
The serious negotiation. A large order, special terms, a volume discount. That is a proper commercial conversation, with margin and judgement involved. A bot should not have the authority to close it, and the customer can tell when it tries.
In all three cases, the failure is not handing over. The failure is handing over late, after two or three replies that have already burned through the customer's patience.
What the human needs on receipt
This is where most implementations fall down, because the typical handover is this: "Let me put you through to an agent." And the agent walks in blind.
The customer, who has already explained their whole situation, has to explain it again. It is the digital version of the switchboard that transfers you from department to department, reciting your postcode each time. Nobody has measured how many sales die right there, but anyone who has been on the receiving end knows the answer: plenty.
A handover done properly feels like passing over a folder, not hanging up and redialling. The person receiving should have in front of them, without asking:
- The full transcript, not a two-line summary.
- The detected intent: what this customer wants, where they are in the decision, what is still missing.
- The details already gathered: name, area, timescales, whatever the conversation has established.
With that, the human's opening line can be "hi Sarah, I can see you're asking about the bathroom quote — here's the picture", rather than "hello, how can I help?". The gap between those two sentences is the gap between a team and a call centre.
The symmetric failure: the human who isn't there
There is a second failure mode that gets far less attention, and it is the exact mirror image.
The bot does its job, spots that a human is needed, hands over... and the human does not respond. They are on site, in a meeting, at lunch — or it is simply Saturday evening. The customer is left hanging at the worst possible moment: right after being told someone is coming to help.
That is worse than having no bot at all. The silence of a business with no chat is neutral; the silence after "a colleague will be with you now" is a broken promise.
Which is why we think handover needs a safety net in both directions: if the human operator does not respond within a few minutes, the conversation should return to the AI — which picks up the thread, keeps helping with what it can, and flags the open item clearly for when the human arrives. At Zatio we call this the automatic rescue, and it came directly from watching conversations die in that limbo.
What a bot should never attempt
We will end with the uncomfortable part, because an article about handover written by a company that builds AI agents has to earn its credibility by saying what its own product must not do.
A bot must not invent a price, a lead time or a product that does not exist — in our case that is prevented by deterministic hard floors, not by the model's good intentions. It must not settle a complaint that requires judgement. It must not pretend to be a person: the law now requires it to disclose that it is an AI, but it also happens to be what works. And it must not be afraid to say "I don't know" — the least impressive line in any demo is the one that builds the most trust in a real conversation.
A good AI agent is not the one that answers everything. It is the one that answers what it knows, steps aside when it should, and makes the human who takes over look better informed, not more lost. If you want to see how we handle that dance, it is laid out on the product page — or get in touch and watch it live.