AI4 min read

When should an AI agent hand over to a human?

Clear rules for AI-to-human handoff in customer support, the signals to watch for, and how to make the switch feel seamless for the customer.

An AI agent that answers every question is not the goal. An AI agent that answers the right questions, and passes the rest to a person at the right moment, is. Handoff is where customers decide whether your AI is helpful or a wall they have to climb over.

This post covers when to hand over, how to set the rules and what a good handoff looks like.

The principle: AI for answers, people for judgement

AI is very good at questions with a clear answer in your own knowledge: opening hours, delivery times, how a feature works, where to find an invoice. It is a poor fit for situations that need judgement, empathy or authority, such as making exceptions, handling complaints or deciding on money.

A useful test: if a new teammate in their first week would need to ask a manager, the AI should hand over.

Always hand over when the customer asks

This rule is non-negotiable. When a customer says "talk to a human", "real person", "agent" or anything similar, the AI should stop and pass the conversation on. Customers who have to argue with a bot to reach a person rarely come back.

In Inboxbee, "talk to a human" is a built-in handoff rule for the AI agent, and you can add your own.

Topics that should go to a person

Most teams hand over on topics like these:

  • Refunds, cancellations and billing disputes. Anything that involves money leaving the business, or a decision about it.
  • Complaints and strong emotions. An upset customer wants to be heard by someone who can act.
  • Account security. Access problems, suspicious activity, changes to ownership.
  • Legal or safety matters. Anything that could carry risk if answered wrong.
  • Sales opportunities. Larger deals or custom requests, where a person can build a relationship.

Write these down as handoff rules. Be specific. "Refund requests" is easier for the AI to recognise than "difficult conversations".

Signals that the AI is out of its depth

Some handoffs aren't about the topic. They're about how the conversation is going:

  • The AI doesn't find the answer in your knowledge. It should say so honestly and hand over, rather than guess.
  • The customer repeats the same question, which usually means the answer didn't help.
  • The question needs information the AI can't see, such as a decision only your team can make.

The AI agent only answers from the knowledge you give it, with guardrails that keep it on business topics. When the answer isn't there, handing over is the correct behaviour, not a failure.

Make the handoff seamless

A good handoff feels like being passed to a colleague, not starting over. That means:

  • Tell the customer what is happening. "I'm passing you to the team. Someone will reply here shortly."
  • Set expectations outside working hours. If nobody is online, say when they will be, and that the conversation is saved.
  • Keep the full history. The teammate who picks it up should see everything the customer and the AI said, so the customer never has to repeat themselves.
  • Route it to the right person. Use routing to send billing handoffs to billing and technical handoffs to the people who can fix them.

Test your rules before customers do

Before turning the AI on, try to break it in the private test chat. Ask for a refund in three different ways. Say you're angry. Ask something your knowledge doesn't cover. Ask for a person. Check that each one hands over cleanly.

Then turn the AI on for one channel first, such as the website widget, and watch real conversations for a week.

Measure handoffs, then tune

Handoff rate is one of the most useful numbers for an AI agent. In reports you can see how many chats the AI resolves on its own and how often it hands over.

Read it both ways:

  • A very high handoff rate often means missing knowledge. Look at the questions that got handed over and add help articles or Q&A pairs for the common ones.
  • A very low handoff rate can mean the AI is answering things it shouldn't. Read a sample of resolved chats and check the low CSAT ratings.

There is no perfect number. What matters is that the AI answers what it knows, and hands over everything else quickly and politely.

Summary

  • Always hand over when the customer asks for a person.
  • Hand over on money, complaints, security and legal topics.
  • Hand over when the answer isn't in your knowledge.
  • Keep history, set expectations and route to the right team.
  • Test, launch on one channel and tune with reports.

For a full rollout plan, see our checklist for setting up AI replies safely.

Your first chat is minutes away

Start answering today

Free for up to three teammates. Add more when you’re ready.