Every chatbot eventually meets a question it should not answer. A customer wants to know why a parcel marked "delivered" is not on the doorstep. Someone asks for a refund on an order that shipped three weeks ago. A wholesale buyer wants a price for 400 units. None of those are knowledge base questions, and an AI that bluffs its way through them costs you more than one that simply says: let me get a colleague.
That moment — the transition from automated reply to live human agent — is the chatbot human handover. It is the least glamorous part of any support chat and the part that decides whether customers ever trust the bot again. This guide covers what triggers a handover, what the visitor sees, what context the agent receives, and the specific ways a badly built handoff flow loses the conversation.
What is chatbot human handover?
Chatbot human handover is the moment an AI chatbot stops answering and passes the conversation to a human agent, ideally with the full conversation history attached and without asking the customer to start again. The bot handles routine queries; the human team takes the rest.
You will see three words used for the same event. Handover and human handoff describe it from the customer's side: the chat window changes hands. Escalation describes it from the support side: a conversation moves up from automation to a person. All three mean the same thing, and vendors mix them freely, so do not read too much into the label.
The thing worth measuring is not whether a tool advertises the feature. Nearly all chatbots do. It is whether the handoff happens at the right moment, and whether the agent who picks it up can see what already went wrong.
What triggers a handover?
A handover can be triggered by the visitor, by the AI, or by a rule you set. A support workflow that only supports one of those three will leak conversations. Here are the five triggers that matter in practice.
1. The visitor asks for a human
The simplest trigger and the one most often broken. If someone types "talk to a human", "speak with a human", "agent", or just "this is useless", the bot should stop selling its own competence and escalate. No confirmation step, no "are you sure I can't help?", no three-option menu.
Every phrasing has to work. "I'd like to speak to a person", "get me support", "customer service please" and a plain "human" are all the same user request. If your bot only recognises one exact phrase, most people will never reach a human.
2. The AI is not confident in its own answer
This is the trigger that separates a genuinely useful assistant from a liability. If the answer is not in the shop's knowledge base, the AI has two options: guess, or admit it does not know and hand over. Guessing is how you end up promising a next-day delivery you do not offer.
Detecting low confidence is not magic. When an assistant answers from your own documents, it can check whether the retrieved material actually covers the query. If nothing relevant comes back, that is a handover, not an invitation to improvise. This is one of the strongest arguments for an assistant that is trained on your own data rather than one running on general world knowledge — a bot with no source to point at cannot tell the difference between knowing and inventing.
3. The query touches money or something irreversible
Some topics should escalate on sight, regardless of how confident the AI is. Refunds, chargebacks, cancelling an order that has already shipped, changing a delivery address after dispatch, warranty claims, anything involving a dispute. The bot can gather details — order number, postcode, what went wrong — and then pass the conversation to a human agent with all of it collected.
This is not a limitation of AI capabilities. It is a business decision. You do not want an automated system approving refunds at 2am, and customers do not want to negotiate with one.
4. The visitor is going in circles
If the same customer asks a near-identical question three times, the AI has not understood it and repeating a fourth variation will not help. Repetition, rising frustration, or a message consisting mostly of capital letters are all signals to escalate. A good handoff flow treats being asked the same thing twice as a failure state, not as a normal conversation.
5. Topics you have deliberately fenced off
Every shop has subjects it does not want automated. Medical advice on supplements. Sizing guarantees for expensive items. B2B pricing. Anything legal. These are rules, not judgement calls: if the message matches, the conversation escalates to a human before the bot has said anything it should not have.
What a seamless handover looks like to the visitor
From the customer's side, a seamless chatbot to human handoff has four properties.
- It is announced. One short line — "I'm not sure about this one, I'm passing you to a colleague" — so the change of hands is not a surprise.
- It happens in the same chat window. No new tab, no "please email us instead", no ticket form. The conversation continues where it started.
- Nothing has to be repeated. The agent opens with something that shows they already read it, not with "hi, how can I help?"
- The wait is honest. If nobody is free for twenty minutes, say twenty minutes. A spinner that promises nothing is worse than a number.
Note what is not on that list: the customer does not need to be told which parts were AI and which were human. Some shops keep the bot and the human under one display name; others label them separately. Both work. What breaks trust is a bot that pretends to be a person right up until it fails.
What the human agent receives
The handover is only as good as the context that travels with it. When an agent opens an escalated conversation, the full conversation history should already be visible to the agent, along with everything the widget knows about the visitor. At minimum:
- The complete transcript, including what the AI answered and where it got stuck.
- Page URL and referrer — which product the customer was looking at when they asked.
- Order details, if the customer already supplied and verified them.
- Why the conversation escalated: user request, low confidence, or a rule that fired.
- Any earlier chats from the same visitor.
That last item on the list — the reason for the handoff — is the one most dashboards omit, and the one that makes the agent fastest. "Customer asked three times about a delivery date the knowledge base does not cover" tells a colleague exactly where to start.
It is also the data that tells you what to fix. A pile of handovers all triggered by the same missing answer is not a staffing problem; it is a gap in your knowledge base that you can close in ten minutes. Reviewing escalation reasons weekly is the cheapest form of customer service automation there is.
Where chatbot human handover goes wrong
Most handoff processes fail in one of five ways. Each is easy to spot once you know the shape.
The visitor has to repeat themselves
The single most common failure. The bot collects an order number, a postcode and a description of the problem, then hands over to an agent who opens with "Hello, what can I do for you?" The customer has now explained the same thing twice, and the automation has cost them time instead of saving it. If your chat platform passes only the last message rather than the full conversation context, this will happen on every handover.
The bot guesses instead of escalating
AI chatbots that answer everything look impressive in a demo and create work in production. Confident wrong answers about returns windows, stock, or delivery dates generate a second contact — usually an angrier one, often by email, where you cannot see it alongside the chat. A bot that escalates one conversation in five is doing its job better than one that escalates none.
The conversation dies in a queue
The handover fires, the customer is told a colleague is coming, and then nothing happens because the escalation went to a channel nobody watches. This is a routing problem, not an AI problem, and it is entirely preventable: escalations need a notification somewhere a person will actually see it, and an unanswered escalation needs to be visibly overdue in the dashboard after a set number of minutes.
There is no route back
Some handoff flows are one-way. Once a conversation escalates, it stays with the human forever, even after the actual problem is solved and the customer asks a simple follow-up about postage. Handing the conversation back to a bot when the agent closes their part keeps your team out of routine queries they already automated.
The escape hatch is hidden
Burying the option to reach a human behind a menu, or leaving it out of the chat entirely to protect an AI deflection metric, does not reduce the need for human contact. It moves it to your inbox, your phone line, or a review. Deflection you achieve by making people give up is not deflection.
What happens outside business hours?
Outside business hours, an honest escalation collects rather than promises. The AI should still answer what it can from the knowledge base, and when a conversation needs a person it should say so plainly: nobody is available until 9am, leave an email address and you will get a reply then.
What it must not do is offer a live handover that cannot happen. A widget that says "connecting you to an agent" at midnight and then goes quiet is worse than one that admits the office is closed. Set your opening hours in the tool and let the handover message change with them.
Can the conversation go back to a bot after a handover?
Yes, and it usually should. Once the human agent has resolved the part that needed a person, returning the chat to the AI means the next routine query — order status, opening times, where the size chart is — gets an instant answer instead of waiting for a colleague to be free.
The rule that keeps this from annoying people: a customer who has already been escalated once should reach a human faster the second time. If they asked for a person an hour ago, they have earned a shorter path back.
Frequently asked questions
How does an AI chatbot detect that it should escalate?
Through three mechanisms working together: keyword and intent matching on explicit requests such as "speak to a human", a confidence check on whether the knowledge base actually contains an answer to the query, and hard rules on topics you have marked as human-only. Conversational AI is good at the first, decent at the second, and should not be trusted alone on the third — that is what rules are for.
What share of conversations should end in a handover?
There is no correct number, and anyone quoting one has not seen your catalogue. What matters is the direction over time. If handovers fall because you keep answering the same gaps in your knowledge base, that is real progress. If they fall because the bot got more willing to guess, your email volume will tell you within a fortnight.
Do I need a helpdesk like Zendesk or Freshdesk for handover?
No. A helpdesk turns the escalated chat into a support ticket, which is useful if you already run one and want everything in a single queue. For a webshop with a small human team, a live chat widget with a shared inbox and a notification to Slack or Teams covers the same ground with less setup. Add the ticketing layer when the volume justifies it, not before.
Does chatbot to human handoff work on WhatsApp and other channels?
It works anywhere the conversation is persistent, which includes WhatsApp, website chat and email threads. Multi-channel handover has one extra requirement: the agent needs the history from that channel, not a generic transcript. Where it gets harder is voice, because there is no chat window to hand over inside.
Is human handover only for problems?
No. Some of the most valuable handoffs are commercial. A visitor comparing two products, asking about bulk pricing, or hesitating on a high-value basket is a case where a person closes what an AI agent cannot. Handover for sales questions tends to pay for itself faster than handover for complaints.
Building a handover you can trust
The pattern behind all of this is simple: let AI and human support each do the part they are actually good at. AI handles volume, availability and recall — it answers the same delivery question for the four-hundredth time at three in the morning without getting bored. Human agents bring judgement, empathy and the authority to make exceptions. The handoff is the seam between them, and it is where most implementations are weakest.
If you are choosing a tool, the questions worth asking are narrow. Does the AI escalate when it does not know, or does it always produce an answer? Does the agent see the full conversation history and the reason for the escalation? Can the customer reach a human in one message? Does an escalation that nobody picks up become visible to you? A vendor who cannot answer those four is selling you a demo, not a handover.
Veyra is built around this seam. The assistant answers from your own knowledge base, escalates to your team when it is not confident enough to answer, and hands over the full conversation context so nobody has to repeat themselves. If you want the wider picture first, our guide to the customer service chatbot covers how the pieces fit together, and the practical setup for adding an AI chatbot to your website takes about as long as reading this article. The why Veyra page explains what we deliberately do not automate.