Chatbot human handoff: 9 rules for when an AI support bot should step aside
By The Relay team · Published · 5 min read
The short answer
An AI support bot should hand off to a person when it can't answer from content you trust, when the customer asks for a human, when money or account changes are involved, and when it fails. The handoff should keep the conversation, ask for contact details once, and never leave the customer at a dead end. This post lists nine rules and shows how Relay applies each one.
A support bot is judged on the conversations it shouldn't have finished. Answering the easy questions is the job; the harm comes from the ones it kept going on. These rules describe what to expect from any bot. Where it helps, we say how Relay does it, because Relay's rules are written down and you can check them against the live demo.
The nine rules
1. Answer only from content you control, and hand off when it doesn't cover the question
A bot that fills gaps from general knowledge will eventually invent a policy. Relay's AI is told to answer only from your published help center articles, to keep answers brief, and to reply with a handoff marker when the articles don't cover the question. When that happens the customer is never given a loosely related article excerpt in place of an answer.
2. Treat a request for a person as an instruction
If someone writes "human", "agent", "real person", "someone" or "talk to a person", the conversation moves to your inbox. Relay does this without asking the model, so it can't be talked out of it.
3. Send money and account changes to people
Messages with "refund", "charged twice", "double charged" or "cancel my account" go to a human. The bot also never promises refunds, discounts or timelines. Those are decisions only your team can make.
4. Send breakage to people
Words such as "bug", "not working", "broken", "error" and "crash" hand off too. This is blunt on purpose: a customer asking how to fix an error message may be handed to you when an article would have worked. A false handoff costs you one reply. A bot that argues with someone whose product is broken costs you a customer.
5. Fail toward a human, not toward silence
If the AI refuses to answer or the call to the model fails, Relay hands off instead of showing an error or a vague apology. If AI answers are off, or no article matches, the same rule applies.
6. Keep "talk to a human" in view
The latest bot message in Relay always shows two buttons: "That helped", which resolves the conversation, and "Talk to a human". Customers shouldn't have to find the right phrase to escape the bot.
7. Hand over the conversation, and ask for contact details once
Relay doesn't start a new ticket on handoff; the same conversation moves to your inbox as open, with the whole thread. If the customer hasn't given an email and your app didn't pass one, the widget asks for it once, right after the handoff ("Get notified when we reply"), so they aren't blocked from asking.
8. Give files to people
A bot can't read a screenshot of an error, so a message containing only attachments goes straight to your team.
9. Label the bot and keep private data out of it
Relay tells customers they are chatting with an AI assistant, with a notice under the header and a label on every bot message. If answers come from keyword search only, with no AI, it says "automated assistant" instead so the label matches what is happening. Both disappear once a person takes over. Article text is passed to the model as data, not instructions, and internal notes and other customers' data are never sent to it.
What to test before you launch any bot
- Type "I want to talk to a person". It should hand off at once.
- Ask something your help center doesn't cover. It should say so and hand off, not improvise.
- Ask for a refund or say you were charged twice. It should hand off and promise nothing.
- Send a screenshot with no text. It should reach a person.
- Ask for a discount or a deadline ("by Friday?"). It should not commit.
- Break it: turn the AI off, or remove the API key. Customers should still be able to ask and be handed to a human.
- Read the conversation as the customer. Is it obvious that it's a bot, and obvious how to reach a person?
What Relay doesn't do
Relay's bot answers questions; it doesn't take actions such as issuing refunds or changing plans, so those always go to a person. Every handoff lands in one shared inbox, and Relay has no routing rules or round-robin assignment yet. If you need those, see how Relay compares with Intercom and others.
Questions
Common questions
- When should an AI chatbot hand off to a human?
- When it can't answer from trusted content, when the customer asks for a person, when the topic involves money or account changes, when something is broken, when the customer sends only a file, and whenever the AI fails or refuses.
- Should a support bot ask for the customer's email before handing off?
- No. Hand off first, then ask once for contact details if you don't have them, so the customer can reach a person straight away and you can follow up if they leave. That is how Relay's widget works.
- Does Relay's bot say that it is an AI?
- Yes. The widget shows a notice that the customer is chatting with an AI assistant, and an "AI assistant" label on every bot message. If answers come from keyword search only, it says "automated assistant". The labels go away when a human takes over.
- What happens if the AI is switched off?
- The keyword bot still works: it replies with the best matching help center article's first paragraph and a link, and hands off when nothing matches. You can turn AI answers on and off in Inbox settings.
Sources and how we wrote this
Relay is made by the team that wrote this post. Figures for other tools come from their own public pages, listed under Sources, and are list prices in US dollars that can change. If something here is out of date, tell us on the feedback board and we'll correct it.