The short answer
A chatbot should hand the chat to a person in five moments: when the customer asks for one, when nothing in your own knowledge answers the question, when two of your sources disagree, when the AI errors or times out, and when the bot itself judges the case needs a person. Every handover should carry a short note on what the customer needs, plus their name and email if you have them, so your teammate never has to ask the customer to repeat themselves. Once a person replies, the bot has to stop completely, including any reply it was halfway through writing, until your team hands the chat back.
Most guides on chatbot human handoff stop at “pass the transcript and keep wait times short”. That’s right, but it misses the parts that break in real use: the bot answering over your teammate, the bot guessing instead of handing over, and the customer left waiting at 11 pm with no idea whether anyone saw their message. This guide covers those too, and ends with a ten-point checklist you can run on any tool in about fifteen minutes.
If you want to see this working on your own site, SupportVia is an AI chatbot for your website that answers from your own pages and hands over to a shared team inbox. See plans, with a free trial of 7 days or 50 AI conversations and no card.
Why the handover matters more than the bot
A wrong answer from a chatbot costs more than no answer. In February 2024 a Canadian tribunal ordered Air Canada to pay a customer after its website chatbot told him he could claim a bereavement fare after traveling, which the airline’s actual policy didn’t allow. The airline argued the chatbot was responsible for its own words. The tribunal disagreed (Moffatt v. Air Canada, 2024 BCCRT 149).
The lesson for a small business isn’t “don’t use AI”. It’s that the moment a bot doesn’t know, it should say so and pass the chat to someone who does. A good chatbot to live agent handoff is the safety net that makes the AI worth turning on.
The five moments to hand over
Most tools only handle the first of these. Check that yours handles all five.
1. The customer asks for a person
This is the simplest trigger and the one most often made difficult. Some bots hide the option, or only offer it after the customer has rephrased a question three times. Don’t. If someone types “can I talk to a human” or presses a button for it, hand over straight away and say so.
A common worry is that customers will skip the bot entirely. In practice most people ask the bot first because it’s faster. The ones who ask for a person straight away usually have a reason: a complaint, a complicated order, or a question they already know isn’t on your website.
2. Nothing in your knowledge answers the question
This is the most important trigger, and the one that separates a safe bot from a risky one. When a customer asks “do you deliver to Brooklyn on Sundays?” and your website says nothing about Sunday delivery, the bot has two options. It can guess from what AI models generally know about delivery companies, or it can say “I’m not sure about that. Want me to get someone from the team?”
Only the second is acceptable for prices, policies, dates, fees and contact details. A bot that answers only from your own pages and notes needs to know when it found nothing, and treat that as a reason to hand over rather than a gap to fill. We cover the rules a bot should follow in more detail in why chatbots give wrong answers, and how to stop it.
3. Two of your sources disagree
This happens more than people expect. Your returns page says 14 days, an old FAQ says 30, and a blog post from 2022 says 7. A bot that picks one at random will be wrong a third of the time and won’t know it.
The right behavior is to say the information differs and offer a person. A teammate can then give the right answer, and you’ve found a page to fix.
4. The AI errors or times out
AI providers have outages. Requests time out. A model sometimes refuses a harmless question. If your bot simply goes quiet, the customer sits looking at a typing indicator that never ends.
Every failure should turn into a handover, with a short apology to the customer (“Sorry, I’m having trouble answering right now. I’ve asked a teammate to pick this up.”) and a note for your team saying what went wrong. Set a time limit, such as 60 seconds, after which the chat goes to people regardless.
5. The bot decides on its own
Some cases are answerable but shouldn’t be answered by a bot: an angry customer, a complaint about a staff member, a question about a medical symptom on a clinic site, a request for a refund exception. A good AI chatbot human handoff lets the model call for a person when it judges the conversation needs one, and write down why.
| Moment | What the customer should see | What your team should see |
|---|---|---|
| Customer asks | “I’m bringing in a teammate. They’ll reply right here.” | Customer asked for a person, plus what they were asking about |
| No source | “I’m not sure about that. Want me to get someone from the team?” | The question that had no answer |
| Sources disagree | “I’m seeing different information on this. Want me to get someone from the team?” | Which question, and that the pages conflict |
| Error or timeout | A short apology and the promise of a person | What failed |
| Bot decides | A plain line that a person will reply | The bot’s reason |
What the person needs to see
The point of a handover is that the customer never has to start again. Your teammate needs three things the moment they open the chat.
The whole conversation. Every message from the customer and every answer from the bot, in order, including any links the bot showed. If the bot gave a wrong answer, your teammate needs to see it to correct it.
A note on what the customer needs. Reading a long transcript takes time. A one or two line note at the top saves it: “Wants to change the delivery address on an order placed yesterday. Asked for a person after the bot couldn’t find an order policy.” The note should be visible to your team only. A customer should never see an internal note like “customer seems annoyed”.
Who the customer is. At least a name and an email, so you can follow up if the customer closes the tab. If your website already knows who is signed in, pass that through to the chat.
What your teammate doesn’t need is a form to fill in, a ticket number to create or a queue to claim from before they can type. Every step between “a customer is waiting” and “I’m replying” adds to the wait.
One state at a time: why the bot must stop once a person replies
This is the part most comparisons skip, and it causes the most embarrassing failures.
Picture it. A customer asks a question. The bot starts writing a reply, which takes the AI a few seconds. In those seconds a teammate sees the chat, takes over and types “Hi, I’m Sam, I can help with that.” Then the bot’s reply lands underneath, answering a question Sam is already handling, possibly with a different answer.
To the customer this looks broken. Two voices are talking at once and they don’t agree.
The fix is a simple rule: a conversation has one owner at a time, and people win. In practice that means:
- Each conversation is in exactly one state: the bot is answering, it’s waiting for a person, a person has it, or it’s closed.
- When a teammate takes over, or simply replies, the conversation moves to “a person has it”.
- From that moment nothing automatic is sent. Any bot reply still being written is thrown away, not posted late.
- The bot only answers again when someone hands the chat back.
When you test a tool, try exactly the race described above. Ask a question, then take over in the inbox within a second or two. If a bot reply appears after your own message, the tool doesn’t enforce one owner.
When nobody is online
Small teams can’t watch a chat window all day. That’s fine, as long as the chat is honest about it.
When a handover happens and nobody picks it up, the chat should:
- Ask for a name and email, so the conversation becomes a lead and not a dead end. Ask one thing at a time, inside the chat, not with a form that hides the conversation.
- Tell the customer how they’ll hear back. “Thanks, we’ll contact you by email” is the minimum. If you can promise a time, such as “by 11 am tomorrow”, say it, but only if your team will keep it.
- Keep the conversation waiting in your inbox, with the handover note, so whoever opens it first can reply in one go.
- Not keep pretending. A status that says “Online” when nobody is there sets up a disappointment. The customer should see that they’re waiting for a teammate.
Alerts matter here too. Your team needs to notice a waiting customer while they’re doing other work. Desktop notifications and a count of waiting chats in the browser tab are enough for most small teams, as long as somebody has the inbox open.
Handing back to the bot
Handovers shouldn’t be one-way. A customer asks about a refund exception, a teammate sorts it out, and then the customer asks “what are your opening hours on Saturday?” There’s no reason a person needs to answer that.
A good tool lets your teammate hand the chat back with one click. From then on, the next question gets an answer from the bot again, and if it hits one of the five moments it comes straight back to the team.
Two rules make hand-back safe:
- The bot sees the whole conversation, including what your teammate said, so it doesn’t contradict them.
- Your current pages beat earlier replies. If a teammate quoted an old price and you’ve since fixed the page, the bot should go by the page, not by what was said earlier in the chat.
Closing works the same way. A closed conversation should reopen if the customer writes again, rather than starting a fresh chat with no history.
Checklist: ten tests for any tool’s handoff
Run these on a trial account before you put any chatbot live. Each takes a minute or two.
- Ask for a person in your first message. Do you get one straight away, without being made to rephrase?
- Ask something that isn’t on your website. Does the bot say it isn’t sure and offer a person, or does it make up a plausible answer?
- Put two different numbers on two pages (a delivery time, a fee). Ask about it. Does the bot pick one, or say the information differs?
- Break the AI on purpose by using a wrong API key or an unreachable model, if the tool lets you. Does the chat go to people with an apology, or hang?
- Open the handed-over chat as a teammate. Is there a note on what the customer needs, visible only to your team?
- Run the race. Ask a question and take over within two seconds. Does a bot reply ever land after yours?
- Hand over with nobody online. Does the chat ask for a name and email, and does the conversation show up in your inbox as a lead?
- Check the customer’s side. Does the chat show they’re waiting for a person, and then who joined?
- Hand the chat back and ask a simple question. Does the bot answer again, and does it respect what your teammate said?
- Count what you’re charged for. Are chats your team answers billed, or only the ones the AI answers? Some tools charge per AI outcome or resolution on top of seats, so check how a handed-over chat is counted.
If a tool fails tests 2, 3 or 6, the handover isn’t a safety net. It’s decoration.
For a shortlist of tools tested on these points, see the best AI chatbots with human handoff. If you’re comparing a bot-only tool with one that includes a team inbox, our Chatbase alternative page walks through the difference.
How SupportVia does it
SupportVia was built around the handover rather than adding it later, so each of the points above maps to something in the product.
The five triggers. The customer can ask for a person from the chat at any time. The AI is told to answer only from your website’s pages, your Q&As and your notes, and when nothing matches, it says it isn’t sure and offers your team. When two sources disagree, it says the information differs and offers a teammate. If the AI errors, refuses or doesn’t answer within the time limit (60 seconds by default), the chat goes to your team with an apology to the customer. The AI can also hand over on its own judgment. The full rules are in how it answers.
The note. Every handover comes with a note for your team: that the customer asked for a person, what went wrong if the AI failed, or, when the AI hands over by itself, its note on what the customer needs. Your team sees it in the inbox. The customer never does.
The inbox folders. Conversations sit in a shared live inbox with these folders:
| Folder | What’s in it |
|---|---|
| All | Every conversation |
| Needs a human | Handed over and waiting for a teammate |
| My open | Conversations you’ve taken over |
| With the team | Conversations any teammate has taken over |
| Assistant | Conversations the AI is answering |
| Leads | Customers who left their name and email while waiting |
| Closed | Finished conversations |
You can filter the inbox by website if you run more than one.
People win. Take over in one click, or just reply, which takes over too. Once a teammate has the chat, nothing automatic is sent, and if the AI was still writing, its reply is dropped. Replies are signed with the teammate’s name, and the customer sees that they’ve joined.
Nobody online. While a customer waits for a person, the chat asks for their name and email, tells them you’ll contact them by email, and the conversation appears under Leads.
Alerts. Desktop notifications when a customer is waiting, and the waiting count in your browser tab. There are no Slack, email or mobile app alerts, so someone needs the inbox open in a browser.
Hand it back. One click returns the chat to the AI, and the next question gets an answer again.
What’s counted. Plans count AI conversations only. Chats your team answers are unlimited on every plan. Plans start at $29 a month; the details are on the pricing page.
What it doesn’t do: there’s no ticketing, no assignment rules, no canned replies and no notes between teammates beyond the AI’s handover note. It’s a chat on your website, not a helpdesk. If you need email, phone and social channels in one queue, a helpdesk suite will suit you better.
Getting started
Write down the five moments, run the ten tests, and pick the tool that passes them on your own website, not on a demo. If you’d like to try it with SupportVia, add your website and it reads your pages, then try its answers in the private playground before you paste one line onto your site. Start with the AI chatbot for your website, or compare plans.