When should an AI agent hand a conversation to a person?

An AI agent that answers well is useful. An AI agent that knows when to stop is trustworthy. Most frustration with chatbots doesn’t come from wrong answers — it comes from being unable to get past them.

This article is about the second skill: deciding when a conversation belongs to a person, and making the switch so smooth the customer barely notices.

Why hand-off is the real product

Think about the last time you were stuck in a bot loop. You rephrased your question three times, typed “human” twice, and finally closed the tab. The bot may have answered dozens of other people perfectly that day. You won’t remember that. You’ll remember the loop.

For a business, every loop is a lost order or a one-star review. So the most important design question for AI support isn’t “how much can the AI answer?” It’s “how quickly and gracefully does it step aside?”

Five triggers that should always hand over

1. There’s no grounded answer

If the AI can’t find the answer in your own material — help pages, policies, Q&A pairs — it should say so and pass the conversation on. Not improvise. Not answer something adjacent. A grounded “I don’t know, let me get a colleague” beats a confident guess every time.

In AskCub, Cubby answers only from your knowledge base, shows the source with each answer, and hands over when nothing relevant is found.

2. The material isn’t enough

This one is subtler. Your returns policy covers returns — but not returns of sale items bought with a gift card. The AI found something, but not the answer. A good agent treats “partially relevant” as “not enough” and hands over rather than stretching the document to fit.

3. The customer asks for a person

No negotiation, no “are you sure?”, no third attempt. If someone types “can I talk to a human”, they should be talking to one — or leaving a message for one — in the next reply.

4. The conversation goes in circles

Set a turn limit. After a handful of AI answers in one conversation, the odds that the next one will solve it drop sharply; the odds the customer is getting annoyed rise. AskCub uses a maximum number of AI turns per conversation (six by default, adjustable from one to twenty) and hands over when it’s reached.

5. A person is already involved

Once an agent takes the conversation or writes a reply, the AI should go quiet. Two voices answering at once is confusing, and the agent should be in control.

Make the switch invisible to the customer

A hand-off fails when the customer has to start again. The person picking up should see:

• the whole conversation so far, including what the AI already answered and which sources it used;
• the visitor’s context — the page they’re on, where they came from, whether they’ve chatted before;
• why the AI handed over — no match, the customer asked, the turn limit.

With that, the agent can open with “I see you’re asking about returning a sale item — let me check that for you” instead of “How can I help?”

What if nobody is online?

Hand-off at 2 a.m. is still hand-off. The honest version is: “I can’t answer that, and my colleagues are offline. Leave your email and they’ll reply first thing tomorrow.” Then make sure someone actually does. A message form that nobody reads is worse than no form.

Three mistakes that make AI support worse

1. Hiding the AI. Pretending a bot is a person feels clever until a customer notices. Label AI messages clearly and tell people up front. (In many places it’s also becoming a legal requirement.)
2. Making “human” hard to reach. If your hand-off button is three menus deep, you’ve built a wall, not a helper.
3. Never reviewing hand-offs. Every hand-off is a free lesson. If the same question hands over five times a week, write the answer down.

A weekly hand-off review

Spend fifteen minutes a week on this:

1. List the hand-offs with the reason “no match”.
2. Group them by topic.
3. For the top three topics, add a Q&A pair or update a document.
4. Ask the same questions in a private test conversation to confirm the AI now answers.

Do this for a month and you’ll see the share of conversations the AI resolves alone go up — not because the AI got smarter, but because your knowledge base did.

The takeaway

The goal isn’t an AI that never hands over. It’s an AI that hands over at the right moment, with everything the next person needs. Customers forgive a bot that says “let me get someone”. They don’t forgive one that won’t let them leave.

Related articles

AI customer service·Oct 9, 2026·4 min read

How to write a knowledge base an AI agent can actually use

AI agents are only as good as the material they answer from. Seven writing habits that turn a messy help center into reliable answers.

Read article

Cross-border & multilingual·Sep 29, 2026·3 min read

Seven questions overseas shoppers ask before they buy

Shipping, customs, payment, returns, sizing, trust and timing — what cross-border buyers need to hear, and how to answer each one in chat.

Read article

AI customer service·Sep 25, 2026·3 min read

Measuring an AI agent without fooling yourself

“The AI resolved 70%!” sounds great until you ask how it was counted. Four numbers that tell you whether AI support is actually helping.

Read article