It is twenty past nine on a Saturday evening. Someone messages a bathroom fitter's WhatsApp asking about a refit. The agent does well: asks the postcode, the size of the room, whether the tiles are coming off. On the fourth reply comes the line that turns up in almost every long conversation.
"Can I speak to someone?"
What happens in the next thirty seconds decides the job. And it is, by a distance, the worst-built part of almost every chatbot we have looked inside, including several expensive ones.
You and your customer are not searching for the same thing
It is worth comparing the vocabulary on each side of the counter.
The person buying a chatbot searches for "human handoff", "escalation to agent", "live agent transfer". The person on the other end uses none of those words. They type "speak to a human", "how to talk to a real person", "get past the chatbot". That last one is the tell: a good chunk of the search demand around customer service bots is people looking for the fire exit, not the feature.
That is the whole article in two lines. Whoever pays for the system is shopping for a capability. Whoever is stuck in it is looking for a way out.
The number everyone repeats and nobody stands behind
Search "chatbot human handoff" and you will find ten near-identical 2026 guides. Nearly all of them quote the same figure: a healthy handover rate is 15 to 25 per cent of conversations, under 10 per cent means customers cannot find the exit, over 30 per cent means the bot is badly trained.
We went looking for where that range comes from. It comes from nowhere. Not one of those pages cites a study, a sample, a sector or a year. They are copying each other.
We are not going to add another invented figure. Your healthy rate depends on what you sell, what it costs, whether the decision needs a site visit and what time of night people message. There is no reliable public benchmark for it. The only useful move is to measure your own, and there are five tests at the end of this piece to do that today.
In the UK there is no right to reach a human. Elsewhere there now is.
Worth being precise, because this is where most articles wave their hands.
There is no general legal right in the UK to be put through to a person instead of a bot. Two places nearby have gone further. Spain's Ley 10/2025, in force since 28 December 2025, bans automated systems as the sole means of customer service, requires human attention to be an option on the main menu, and requires 95 per cent of those requests to be answered in under three minutes on average — though only for basic-service sectors and firms over roughly 250 staff. And on 3 July 2026 Ireland signed into law the rules implementing EU Directive 2023/2673, giving consumers buying financial products at a distance the right to request human intervention rather than being left with an automated system.
The nearest UK equivalent is narrower and softer, and applies only to FCA-regulated firms. The Consumer Duty's consumer support outcome, PRIN 2A.6, requires firms not to put unreasonable barriers in the way of customers getting support, and the FCA's guidance names unreasonable delays in engaging with the firm as an example of falling short. The regulator has also been explicit about stripping out "sludge" — friction that makes it harder for people to act in their own interest.
So if you fit bathrooms in Kent, none of this binds you. What does reach you is the second-order effect: over the next year your customers get used to asking for a person and getting one, at the bank and at the insurer. By the time they reach your chat window, that is simply the expectation.
The four places handover breaks
1. There is no exit. The commonest failure and the daftest. The agent is trained to resolve, not to leave, so "can I speak to someone" gets answered with another helpful question. Twice in a row and the person closes the window. That is not a model failure. Nobody wrote the route.
2. The alert reaches nobody. The handover "works": the system emails info@, or pings an app only the person who installed it has. On a Saturday night that is identical to doing nothing. The conversation gets marked as escalated, which on a dashboard looks a lot like resolved.
3. It is picked up outside the 24-hour window. This part is technical, specific to WhatsApp, and almost nobody explains it. Meta's own documentation is blunt: when a user messages you, a 24-hour timer called a customer service window starts; it resets each time they message again; and when the window closes you can only send pre-approved template messages.
Translated into your Monday morning: you can no longer reply to Saturday's question. You can send a Meta-approved template, which reads like a bank notification rather than an answer. There is a billing difference too, because Meta's pricing page states that non-template messages are free inside an open window, while outside it you pay per template delivered, under the per-message billing that started on 1 July 2025. Replying on Saturday is free and sounds human. Replying on Monday costs money and sounds like a machine. Rates move and changes have been announced for October 2026, so check the source on the day you need it; the mechanism itself has been stable for years. The full picture is in our API guide.
4. The bot does not shut up. The human steps in and the agent keeps replying over the top — or the reverse: someone answers from their phone and the system never notices. The customer gets two versions of the same answer in two different voices. This is the one that stings, because it exposes that there is no team behind the window, just two things badly glued together.
There is a fifth: the human arriving blind, so the customer has to start again. We covered that at length in When the bot should shut up and will not repeat it here.
Five tests you can run this afternoon
Pick up your phone, message your own WhatsApp or your own website chat as if you were a customer, and write down what happens. Use a stopwatch.
- The plain exit. Type "I want to speak to a human", nothing else. Pass if the next message offers a real route, not another question.
- The angry exit. Type a complaint with some heat in it. Pass if it steps aside quickly, without trying to soothe you.
- Saturday at nine. Repeat test one out of hours. Pass if something honest happens: a stated time you will hear back, not silence after "a colleague will be with you shortly".
- The memory. Explain your situation across three messages, then ask for a person. Pass if the human's first line proves they read all three.
- The clock. Time it from your request to the first sentence written by a human. There is no correct number. There is your number, and until today you did not know it.
Fail three of five and your problem is not the AI. It is the alert and whose shift it lands on.
What none of this fixes
It does not conjure people. If nobody is watching the phone on Saturday night, no system hands over to anyone — at best it creates an expectation, which is worse than having no chat at all.
It does not fix a shared inbox with no owner. If three people can see escalated conversations and none of them is accountable for them, the best detection in the world lands in limbo.
It does not turn a handover into a sale. It only guarantees the conversation arrives alive at someone who can close it.
And it does not get you off the phone. If you are an FCA-regulated firm, an immaculate chat window does nothing for the queue on your 0800 number.
We built Zatio around this moment more than the one before it: the agent says it is an AI, steps aside when it should, and passes the whole conversation to whoever picks it up. It does not answer voice calls and it cannot invent an on-call rota you do not have. If you want to see how it behaves on your own conversations, get in touch.