When should a chatbot hand off to a human? Six triggers and a clean handover

A chatbot should hand a chat to a person when the customer asks for one, when it cannot answer, when money or an account change is involved, when the customer is upset or complaining, when the topic is sensitive, such as health or law, and when it has failed twice. The person then needs the whole conversation.

The handover itself matters as much as the trigger: what your team sees, what the customer is told while they wait, and what happens when nobody is free. Our guide to AI agents, chatbots and live agents explains how to split the work between them. This post is about the moment a chat moves from the bot to a person.

When should a chatbot hand off to a human?

Microsoft's Bot Framework documentation on handing a conversation to a human names two broad cases: a handoff “might be necessary if the bot doesn't understand the user (because of an AI limitation), or if the request can't be automated and requires a human action.” In practice that splits into six triggers.

1. The customer asks for a person

This one is not up for debate. People ask in many ways: “talk to a human”, “real person”, “customer service”, “agent please”, or a flat “no” after a bad answer. A bot that knows one phrase traps everyone who words it differently, so list every wording you can think of, in every language you serve. Asking one short question about what the chat is about is fair, because the answer can route the chat or settle it. Asking two is an obstacle.

2. The bot does not know

When the answer is not in what the bot was given, it should say so and offer a person rather than guess. In a 2018 study of chatbots, Nielsen Norman Group found that “Recognizing that a question was not understood was disappointing, but better than a blatantly wrong answer.”

3. Money or a change to an account

Refunds, billing errors, double charges, chargebacks, cancellations and anything that changes what a customer has paid for. Even a bot that can look these up should leave the decision to a person, or at least wait for the customer to confirm. Our checklist for AI account actions covers the safeguards.

4. A complaint or an upset customer

A complaint is a request to be heard by someone who can put things right. Insults, shouting in capitals, “this is useless” and “I already told you” are all signs to stop answering and bring someone in. Cheerful bot replies to an angry customer make things worse.

5. Sensitive topics

Symptoms, test results and medication. Legal disputes. A hacked account, fraud or a request to delete personal data. A bot can tell a patient how to book, but it should not judge how serious a symptom is, and anything that sounds like an emergency needs the emergency number in the reply itself, not a place in a queue. Write these topics down as a list the bot never answers on its own.

6. The bot keeps failing

Two poor answers in a row, the same question asked again in other words, or the customer repeating themselves. The same NN/g study found people annoyed when a bot repeated the same answers, and noted that “Owning the failure and offering an escape hatch (phone number or a live agent) were generally perceived favorably.”

Is a business required to offer a person?

Sometimes. Two texts are worth knowing, as we read them on 13 October 2026. This is not legal advice.

On WhatsApp. Meta's WhatsApp Business Messaging Policy, in its terms for the WhatsApp Business Platform, says: “You may use automation when responding during the 24-hour window, but must also have available prompt, clear, and direct escalation paths.” Its list of escalation paths starts with an in-chat transfer to a human agent and also includes a phone number, email and a support form. Our post on WhatsApp's 2026 AI rules covers the rest of the policy.

In the EU, for financial services sold at a distance. Directive (EU) 2023/2673 adds Article 16d to the Consumer Rights Directive. Its paragraph 3 says Member States shall ensure that, where a trader uses online tools, “the consumer shall have a right to request and to obtain human intervention at the pre-contractual stage, and in justified cases after the distance contract has been concluded”. The directive's recitals name chatbots among those tools, and EU countries had to apply their national rules from 19 June 2026. It covers financial services contracts concluded at a distance, not every business.

What the person taking over needs to see

• The whole conversation. Microsoft's handoff design passes “A transcript of the conversation, so an agent can read the conversation that took place between the customer and the bot before the handoff was initiated.”
• Why it was handed over. The customer asked, the bot did not know, or the topic is on your list. That tells the person where to start.
• What the bot already said. Above all any promise, price or link, so the person does not contradict it without explaining why.
• Who the customer is, if they gave a name or an email or are signed in, and which page they were on.

To the customer, it is all one chat. Being asked again for something they already typed, NN/g's 2019 study of customer-service chat says, “shows that the company does not have a well-planned protocol for chat.”

What to tell the customer while they wait

• That a person has been asked. “I've asked our team to join this chat.”
• How long it may take, only if you know. NN/g's chat guidelines prefer an actual time estimate to a place in a queue. If you cannot keep a time, do not give one: “shortly” repeated for twenty minutes is worse than no promise.
• That the bot can still help. An empty wait is worse than being able to ask simple things until the person arrives. When the person writes, the bot should stop, so the two never answer over each other.

Then plan for nobody coming. Set a limit, such as five minutes. When it passes, the bot should say plainly that nobody could join, offer to take an email address, and say the reply will come by email. Promising a person again and again when nobody is coming does more harm than saying so once.

What should happen outside business hours?

Say when the team is back, with a day and a time, and never say that anyone is joining. One of NN/g's chat guidelines reads: “If Live Chat Is Not Always Available, Specify the Hours When It Is.” Then choose one of three: queue the chat for the morning, queue it and take an email address, or let the bot handle it alone if nobody reads the inbox after hours. A line such as “Our team is offline until Monday at 9:00. Leave your email and they will reply then, or ask me anything in the meantime” covers it. Our page on the after hours chatbot goes through the options.

How MagizAI hands a chat to your team

In MagizAI, a visitor can ask for a person in their own words or use the Talk to a human link in the chat window. The AI also hands over when it cannot help, or when a question is on the list of topics you leave to your team. You choose whether it first asks one short question about what the chat is about, and whether it should give one answer before passing the chat on. A request for a person that mentions a refund, a billing error, a legal matter, a hacked account or a complaint skips both. Your team is alerted and the chat lands in the inbox with the whole conversation. While a person is on the way, the AI keeps helping, and it stops as soon as one replies. If nobody picks the chat up within the wait you set, five minutes unless you change it, the AI says so and offers to take an email address. Outside the business hours you set, it says when your team is back. The inbox and handoff guide shows how to take over a chat.