When should an AI chatbot hand off to a human?

Short answer. Hand off when the bot has no sourced answer, when the topic is one a person must own (refunds, cancellations, legal and security questions, a customer's own data), when the customer asks for a person, when they are upset, and when its confidence is low. A good handover tells the customer a person will reply right there, never promises a time it can't keep, and tells your team why it handed over.

The most important thing an AI support chatbot does is know when to stop. A wrong answer said with confidence costs more than any number of handovers: the customer acts on it, comes back angrier, and stops trusting the chat. In a Gartner survey of 3,566 customers published in August 2026, 87% said a way to reach a human agent is essential when a company uses generative AI for service.

Here are the five moments to hand over, what the message should say, and what it looked like on a real support chat.

The five moments to hand over

WhenWhy
Nothing it knows covers the questionWithout a source the model fills the gap from general knowledge, which is where made-up answers come from.
The topic belongs to a person: refunds, cancellations, legal and security questions, a customer's own dataThese carry money, risk or trust. Even a correct answer should come from someone who can stand behind it.
The customer asks for a personArguing with that request is the fastest way to lose them.
The customer is upsetAn angry customer reads a cheerful bot as being ignored.
Its confidence is low, or a check on the answer failsAn answer that only half fits the question is worse than a short wait.

The first one does most of the work. If the assistant answers only from a knowledge base you control and shows the source with each answer, "it doesn't know" becomes a clear, checkable state instead of a guess.

What one week looked like

FanWake, a chat tool for creators, runs its support on Afterchat. In its first week with the assistant answering on its own, 8 November to 15 November 2026, 13 customers started a conversation. The assistant handed 8 of them to a person at some point:

Why it handed overChats
Nothing in the knowledge base covered the question5
A topic set to go to a person1
Low confidence in its answer1
The customer was upset1

Counted 16 November 2026 from the chat's handover log.

That is a high share, and it is the right behaviour for a new knowledge base. Each "nothing covered it" is recorded as a gap: the question, in the customer's words. Once someone answers it, the answer is kept, and the next customer who asks gets it straight away instead of a handover.

What the handover message should say

Three things: that a person is coming, that they will reply in the same place, and nothing it can't promise. These are the messages Afterchat sends:

  • When the customer asks for a person: "Of course. I've asked a person from the team to join. They'll reply here."
  • On a topic set to go to a person: "Refunds are handled by a person on the team, so I've passed this on. They'll reply here."
  • When a check holds an answer back: "I want to be sure you get the right answer, so I've passed this to a person on the team. They'll reply here."

What it never says is when. "Someone will be with you shortly" is a promise the bot can't keep at 2am, and a broken promise is worse than none. If nobody is online, the better move is to ask for an email address so the reply reaches the customer after they have left the page.

What your team should get

A handover is only half done if the person picking it up has to read the whole conversation to find out why. The reason should travel with it: no answer in the knowledge base, a refund question, the customer asked, the customer is upset. In Afterchat the inbox shows that reason on every handed-over chat, and the team gets an email or a push when one arrives.

What not to do

  • Don't hide the person. A bot that makes customers try three times before it lets them reach a human turns every handover into a complaint.
  • Don't let it guess on the topics that matter. Refund rules and plan limits are where a plausible wrong answer does the most damage. Klarna, which moved much of its support to AI, said in 2025 it would make sure customers can always reach a person again.
  • Don't loop. If the customer repeats a question after a handover, the bot should not answer it again on its own.
  • Don't count a handover as a failure. Measure it, and work down the "nothing covered it" share by filling the gaps it records.

What to measure

Track three numbers each week: the share of chats handed over, why they were handed over, and how long customers then waited for a person. The last one is easy to forget, because the first reply from the bot already went out in seconds. You can work it out from an export with the first response time calculator.

Questions

What is a chatbot handoff to a human?

It is the moment an AI chatbot stops answering and passes the conversation to a person on the team, with the conversation so far and the reason, so the customer does not have to start again.

Do customers want to be able to reach a human?

Yes. In a Gartner survey of 3,566 customers published in August 2026, 87% said a way to reach a human agent is essential when a company uses generative AI for customer service. Pushback comes when the AI stands between the customer and a person, not from the AI itself.

How do I stop an AI support chatbot from making things up?

Let it answer only from a knowledge base you control, show the source with each answer, and make it hand over when nothing it has covers the question. A handover is a better outcome than a confident wrong answer.

Should a support chatbot say it is an AI?

Yes, in its first message. Customers find out anyway, and saying it up front, with an easy way to reach a person, is what keeps them willing to use it.

Read next

Buyer's guide AI customer support for SaaS: a checklist for small teams What to look for in an AI support chat when your team is small and answers its own customers: sourced answers, honest handovers, learning from every chat, knowing what to build, and working with your coding agent. Knowledge base An AI support chatbot trained on your docs and your code How to give an AI support chatbot your own knowledge: why crawling your docs is only the start, how short checked entries built from your docs, code and chats answer better, and how to keep them true. Free tool First response time calculator Drop in a chat or ticket export and see your median first response time, in calendar and working hours, how many customers waited over an hour, and what happens to chats that come in while you are away. Free, in your browser.