RAG & Chatbots
When an AI Chatbot Should Hand Off to a Human (and How to Set It Up)
Muhammad Taha5 min read
An AI chatbot should hand a conversation to a person when the question is outside its approved answers, the customer is upset, the request involves money or a legal or safety risk, or the customer asks for a human. A good handoff passes the full context so the customer never repeats themselves. A bad one leaves people stuck in a loop, which costs more trust than having no chatbot at all.
Key takeaways
- Set clear triggers in advance: low confidence, repeated failure, strong negative sentiment, money or legal topics, and any explicit request for a person.
- Always offer a person when the customer asks, even if the bot thinks it can help. Forcing a loop damages trust.
- Pass the summary, the transcript and the customer's details to the agent, so the conversation continues without a restart.
- Test the handoff at quiet and busy times, and after hours when no person may be available to take it.
The handoff triggers to set
- Explicit request: the customer asks for a person, in any wording.
- Unknown answer: the bot cannot find an approved answer after one clarifying question.
- Repeated failure: the customer rephrases the same question twice.
- Negative sentiment: frustration, threats or abusive language.
- High-risk topics: refunds above a set amount, legal claims, medical or safety concerns, and data requests.
- High-value contacts: VIP customers or leads above a set value, who should get a faster route.
What the agent should receive
A handoff that starts with "how can I help you" wastes the customer's time. Send the agent a short summary at the top of the queue: who the customer is, what they want, what the bot already tried, the order or booking details, and why the handoff happened. Include the full transcript below it, so the agent can see the wording. The customer should be told that a person is joining and roughly how long they will wait.
Handoff options by situation
| Situation | Best route | Wait message |
|---|---|---|
| Team is online | Live handoff to the next free agent | Estimated wait time |
| Outside business hours | Leave a message and a callback time | When the team will reply |
| Urgent safety or medical concern | Show emergency instructions first | Emergency contact |
| Refund or dispute | Ticket with details, assigned to a manager | Response time you can keep |
| Simple question the bot can answer later | Offer a follow-up instead of a transfer | Expected reply |
Avoid the loop
The most common failure is a customer who asks for a person and gets the bot again. Make the request a hard trigger that cannot be overridden by the bot's own confidence. If no person is free, say so honestly, take a message and set a clear reply time. Never let the bot pretend to be a person when a customer asks directly whether they are talking to one.
How to test before launch
- Write twenty test conversations, including angry customers, vague questions and direct requests for a human.
- Check that each trigger fires, and that the agent receives the summary and transcript.
- Test after hours and on a busy day, so you know what the customer sees when no one is free.
- Review the first two weeks of live handoffs, and tune the triggers where the bot handed off too early or too late.
Frequently asked questions
How quickly should a chatbot hand off to a person?
Immediately when the customer asks for one or the topic is high risk. For other cases, after one clarifying question or two failed attempts.
Should the chatbot tell customers it is an AI?
Yes. Customers should know they are talking to an automated assistant, and must not be misled into thinking it is a person.
What if no agent is available?
Take a message with the same context, tell the customer when they will hear back, and make sure the team sees the queue at the start of the next working period.
How do we measure a good handoff?
Track the share of handoffs that get a useful reply quickly, and ask agents to flag transfers that arrived without context. Those flags show where the summary needs work.
For the wider picture on running a chatbot that does not overwhelm your team, read AI chatbot for businesses overwhelmed by daily inquiries and why AI chatbots hallucinate and how to reduce it.