A good chatbot-to-human handoff happens when the customer asks for a person, when the bot isn’t confident in its answer, when the topic is sensitive, or when the conversation is going round in circles. The agent should receive the whole transcript, a short summary and the customer’s contact details. The customer should be told honestly who will reply and roughly when. Small teams usually get handoff wrong by making the human hard to reach, not by having too few people.
Customers are clear about this. In a Gartner survey of 3,566 customers run in February and March 2026, 87% said companies using generative AI for service must offer a way to reach a human agent (Gartner press release). An earlier Gartner survey found that 64% of customers would prefer companies didn’t use AI for service at all, and their main worry was that it would be harder to reach a person. The nine rules below are designed to address that worry.
Rule 1: Show the human option from the first message
Tell customers at the start that they can reach a person: “Type ‘talk to someone’ at any time.” A hidden handoff doesn’t reduce work. It makes customers more frustrated by the time they get through, and some leave before they do. A visible option also builds trust in the bot’s answers, because people know they aren’t trapped.
Rule 2: When someone asks for a person, hand over
An explicit request should trigger a handoff immediately, with no “Before I transfer you, can I try one more thing?”. Watch for the obvious phrases and your own customers’ wording:
- “human”, “real person”, “agent”, “someone from your team”
- “speak to”, “call me”, “manager”, “complaint”
- industry terms your customers use, such as “reception” for a hotel or “the vet” for a clinic
One exception: “are you a real person?” is a question, not a request. The bot should answer it truthfully (“No, I’m an AI assistant”) and offer a handover, rather than silently escalating every curious visitor.
Rule 3: Hand over when the bot doesn’t know, instead of guessing
The most harmful chatbot failure isn’t “I don’t know”. It’s a confident answer that is wrong. Set up your assistant to answer only from your own content. When nothing in that content covers the question, or the model’s confidence is low, it should say so and offer a person. “I don’t have that information, but our team does” is a good answer. A made-up returns policy is not.
Rule 4: Send sensitive topics straight to a person
Some conversations shouldn’t be automated even when the bot could produce an answer:
- complaints and anything involving a refund or compensation decision;
- legal threats, chargebacks and data-protection requests;
- health, safety or medical questions, especially in clinics and pharmacies;
- account security: suspected fraud, locked accounts, changes of email or address;
- bereavement, hardship or a customer who is clearly upset.
Add these topics as handoff keywords and write them into the bot’s instructions, so it offers a person rather than trying to handle them.
Rule 5: Use a two-strike rule for failed answers
If the customer rephrases the same question twice, says “that’s not what I asked”, or starts writing in capitals, the bot has failed. Stop and hand over. The US Consumer Financial Protection Bureau described the alternative in its 2023 report on chatbots: “doom loops” of repetitive, unhelpful answers with no way through to a human. Two strikes is a simple, testable rule that prevents them.
Rule 6: Pass on the full context
Making customers repeat themselves makes a handoff feel like starting again. The agent should see all of this before typing a word:
| Pass along | Why it matters | Example |
|---|---|---|
| Full transcript | The agent sees exactly what was asked and answered | All messages, including anything the bot got wrong |
| A two-line summary | Faster than reading 20 messages | “Wants to change delivery address on order #4102, already shipped. Frustrated.” |
| Why it was handed over | Tells the agent how urgent it is | Explicit request, keyword, low confidence, sensitive topic |
| Contact details | Lets you reply if the customer leaves the page | Email address, captured before or during the chat |
| Verified order or booking context | Saves looking it up again | Order number and status, only if the customer has been verified |
| What the bot already suggested | Avoids giving the same answer twice | “Sent the returns policy link; customer says item is damaged” |
Rule 7: Tell the customer what happens next, honestly
“Connecting you to an agent” followed by silence is worse than no chatbot. Match the message to reality:
- Someone is online: “I’ve passed this to the team. Someone usually replies within a few minutes.”
- Nobody is online: “Our team is offline until 9am [time zone]. Leave your email and a person will reply then. I’ve saved our conversation, so you won’t need to repeat anything.”
- Busy periods: give a realistic range rather than a hopeful one. A promised five minutes that becomes an hour costs more goodwill than an honest “within two hours”.
Set office hours in your chat tool so these messages switch automatically. Ask for an email address when a reply may come later, but don’t make the customer fill in a form before they can ask for a person.
Rule 8: Be honest about who is talking
The bot should say it is an AI, and the person who takes over should say they are a person, preferably by first name. Don’t give the bot a human name and photo to make the handover feel smoother. Customers notice when “Emma” suddenly writes differently, and since 2 August 2026 the EU AI Act requires AI chat systems to disclose that they are AI. Our AI transparency page explains how we approach this.
Rule 9: Measure handoffs and fix the causes
Track a few numbers each month:
- Handoff rate = conversations handed to a person ÷ conversations the AI started. For example, 88 handoffs from 400 AI conversations is 22%.
- Handoff reasons: explicit request, keyword, low confidence, sensitive topic, repeated failure.
- Time to first human reply after the handoff, split into office hours and out of hours.
- Satisfaction (CSAT) for handed-off conversations compared with AI-only ones.
- Repeat contacts: customers who come back about the same issue within a few days.
There is no universal “good” handoff rate. It depends on how complex your products are and how complete your content is. What matters is the trend and the reasons. If most handoffs come from low confidence on the same three topics, write those answers into your knowledge base. A very low rate isn’t automatically good either: it can mean the human option is hard to find.
How to word the handover
A handover message should cover three things in a sentence or two: that a person is taking over, that they can see the conversation, and when they will reply. Adapt these to your tone:
When the customer asks for a person
Of course. I’ve passed this to our team, and they can see everything we’ve discussed. Someone usually replies within [time].
When the bot doesn’t know the answer
I don’t have reliable information on that, and I’d rather not guess. I’ve sent your question to our team, and a person will reply by [time].
When a person joins the chat
Hi, I’m [first name] from the [Business] team. I’ve read your conversation with our AI assistant, so there’s no need to repeat anything. Let me look into this now.
Avoid “Please hold while I transfer you”, which suggests a phone queue, and avoid apologising for the bot at length. Short and specific works best.
Anti-patterns to remove this week
- Loops: the bot answering the same thing after the customer has said it didn’t help.
- A hidden human option: a buried link, or a handoff that only works if the customer types an exact phrase.
- Handoffs that go nowhere: escalating into an inbox nobody checks, with no reply time given.
- Making customers repeat themselves: the agent opens with “How can I help?” when the transcript is right there.
- Bots pretending to be people: human names, stock photos and scripted denials.
- Forms before help: asking for name, email and phone before answering a simple opening-hours question.
- Promising what you can’t deliver: “An agent will be with you shortly” at 11pm on a Saturday.
A setup checklist for small teams
- The first message says how to reach a person.
- Handoff keywords include your industry’s terms and your sensitive topics.
- The bot answers only from your content and offers a person when it can’t.
- Two failed answers in a row trigger a handoff.
- Agents see the transcript, a summary and the reason for the handoff.
- Office hours and realistic reply times are set, with email capture when offline.
- Someone checks the handoff reasons every month and fixes the top gap.
KuraChat was built around these rules. It hands over on request, on your keywords or when confidence is low, gives the agent the transcript and an AI summary, and respects your office hours. It’s worth a look if your current tool makes reaching a person harder than it should be. See the full feature list or compare plans and pricing.
Frequently asked questions
What is a good chatbot handoff rate?
There is no universal benchmark, because it depends on how complex your questions are and how complete your content is. Track the rate and the reasons each month, and treat a falling rate from better content as good news and a very low rate as a sign to check the human option is easy to find.
How should a chatbot hand off when nobody is online?
Say that the team is offline and when they will be back, ask for an email address, and confirm that the conversation is saved. Don't promise an immediate reply that won't happen.
Should the chatbot ask for contact details before handing off?
Only when you need them to reply, for example outside office hours or if the customer might leave the page. Don't make a form a condition of reaching a person.
Should the bot use a human name to make the handoff feel seamless?
No. The bot can have a name, but it should say it is an AI, and the human who takes over should introduce themselves. Posing as a person damages trust and, for EU customers, conflicts with the AI Act's disclosure rules.