When Should an AI Chatbot Hand Off to a Human?

A chatbot should hand off the moment it hits a question outside the short list of things it is actually built to answer. Same thing if someone asks for a person directly, or if it gives the same wrong answer twice. Most builds get this backward. They try to make the bot handle more over time instead of drawing a tight line up front and keeping it there.

This shows up constantly in text and DM bots built to reach leads back and forth. Builders spend weeks chasing bugs, wrong timezones, a bot that forgets where the conversation stood, when the real problem is scope. The fix is not a smarter bot. It is a short allowed list: yes, a specific time, reschedule, stop. Anything outside that list goes straight to a person. The bot's replies got more reliable once the list got shorter. It never got smarter.

What happens when a bot doesn't know when to stop?

It keeps trying. An unsupervised bot does not know it is out of its depth, so it loops, repeats itself, or answers something close enough to the question that it feels wrong to the person on the other end.

A Shopify community member named dstuart put it plainly in a support thread from October 2024: "I just need to speak to a customer service representative but the chat bot will not allow me to do so." The bot was not rude. It just never had a rule that said stop trying and get a person.

That is worse for the business than the bot answering nothing at all. Now the person is stuck and annoyed at the same time, and the first human they eventually reach is dealing with someone who already had a bad ten minutes.

That is a different problem from a bot giving conflicting answers because its own knowledge base contradicts itself (we wrote about that here), but the fix rhymes. Know the edges of what the bot actually knows, on purpose, before it has to find out live.

How do you draw the line before you build?

Write down the actual list of things the bot is allowed to answer, in plain language, before you write a single line about its personality. Pull the list from what you already have: your FAQs, your last few hundred support messages, the questions people ask in their first month.

• On the list: order status, pricing, hours, how to access something, the basic troubleshooting steps you would type back yourself anyway

• Off the list: refunds above a set amount, complaints, anything where your honest answer starts with "it depends," any question you have never gotten before

If a question does not fit cleanly on the allowed list, it does not get answered. It gets handed off. The list is not fixed forever either. When the same new question shows up three or four times, that is a signal to add it on purpose, not a sign the bot should have guessed it the first time.

Once you have the list, how you write the actual instructions matters just as much as what is on it.

What should automatically trigger a handoff?

• Someone asks for a human, in any wording

• The bot gives the same non-answer twice in a row

• Money moves outside the normal range: a refund, a discount, an exception

• The tone turns frustrated or upset

• The question is real but sits outside what the bot actually knows

Does a narrow bot feel like a downgrade to customers?

No. People do not mind hearing "let me get someone for that." What actually annoys them is a bot pretending it can help and looping instead.

Escalation guides written for large support teams keep landing on the same short list of triggers: an explicit ask for a human, repeated failed answers, and a conversation that "involves complex decision-making" or "requires empathy and nuanced understanding."

That line is from a chatbot escalation guide at Cobbai (cobbai.com/blog/chatbot-escalation-best-practices), written for teams running big support desks. It is the same short list a one-person business needs. It is just smaller.

---

Common questions

Should a chatbot ever guess instead of handing off?

No. A bot that guesses gives a wrong answer that sounds sure of itself, and the person acts on it before anyone catches the mistake. Saying "I don't have that answer, let me get someone" costs you nothing. A confident wrong answer costs you the account.

How many failed tries should a bot get before it hands off?

One repeat is enough. If it gives the same answer twice and the person is still asking, hand it off on the second try, not the fifth.

Does a handoff lose the conversation history?

It should not. Whoever picks up the thread, you or a teammate, needs to see the whole conversation, not start the person over at "hi, how can I help." Making someone repeat themselves after a bot already failed them is the fastest way to lose them.

Is a bot with a short allowed list worse than one built to handle everything?

It is the opposite. A bot with a short list gets every one of those answers right, every time. A bot built to handle everything gets most things almost right, which is worse, because you cannot tell which of its answers to trust.

Does this change if the bot is trained on your own content instead of a generic script?

No. Training it on your FAQs and your own past messages makes the answers sound like you. It does not move the edge of its job. You still have to draw that line yourself, on purpose, before the bot goes live.

---

When we build an agent for a client, the allowed list is one of the first things we write down, before the bot's personality or its tone. It is the difference between a bot that quietly earns trust and one that quietly loses it. If you want to see what a trained agent looks like on your own content, our demo builds one from a YouTube link in about a minute.