Human handoff
Chatbot human handoff, without the customer starting over
The agent answers first. When it cannot, or when the customer asks, a person takes over the same thread with everything already said in front of them. On Optlo that conversation costs the same as one the agent finished.
The short answer
A chatbot human handoff is passing a conversation from an AI agent to a person, with everything already said going along with it. The test is simple: the customer never has to repeat themselves. Some people write handover. It is the same thing.
A handoff has three jobs
- The customer can ask for a person at any point and get one, without arguing with a machine about whether their problem counts.
- The agent stops on its own when it is out of its depth, instead of producing a confident guess.
- The whole conversation travels, so whoever picks it up already knows what has been tried.
Most tools claim all three. They differ in what your team receives, what fires the handoff, and what the customer sees while they wait.
What arrives
Your team gets the transcript, and a brief with it
Your team should not have to read the whole conversation before they can reply. They need three things: why the agent stopped, who the customer is, and the thread. The note answers the first and the details panel the second, before anyone opens the third.
The note is the part most setups leave blank. On Optlo the workflow asks the model to write one as it hands off, covering what the customer wants, what the agent tried, and why it stopped. Your team can act on that in ten seconds. A transcript takes a minute to skim and still leaves them guessing.
The identity comes from your site, not the chat. A signed-in customer arrives verified through a signed token, with whatever attributes you chose to send. A refund request from a verified customer on an annual plan is not the same conversation as the same words from a stranger, and whoever replies should know which one they have.
When it fires
Some triggers are fixed rules. Some are judgment calls.
Every guide lists the moments a chatbot should hand off. Few say how firm each one is, and a language model sounds just as sure when it is wrong. The model usually notices the moment, so the question is what it is allowed to do next.
Fixed rules
Each of these has a clear yes or no. The model notices the moment, and the rule decides the rest: hand off, and do not try again.
The customer asks for a person
The clearest trigger there is, and the one most bots still argue with.
Hand off the moment they ask. The agent replies once, escalates, and the waiting workflow takes over. No second attempt.
The request needs their account
Plan changes, password resets, anything where the answer depends on who is asking.
If the customer is not verified, the agent asks them to sign in or leave an email. If they are, it can look up their billing in Stripe, or hand off.
Money is involved
Refunds, payments, credits. A wrong answer here is a wrong transaction.
The agent can look up billing but not change it. Anything that moves money goes to a person, unless you decide otherwise.
A complaint, a safety issue or a legal question
Topics where a wrong answer costs more than a slow one.
Straight to a person. The agent does not attempt an answer first.
Judgment calls
No clean rule exists. The model weighs the conversation and chooses between three outcomes, and these are the cases where it should choose the third.
- Answer
- Search again
- Hand off
The search comes back empty
Nothing in your sources answers the question.
The agent searches, then guesses at an answer and searches again with it. If that also finds nothing, it hands off rather than invent one.
The customer is getting frustrated
Short replies, capitals, a second "as I already said".
A listed reason to hand off. The model reads tone better than a keyword list, so this is where its judgment earns its place.
The same question, asked a third time
Rephrasing is a customer who cares, not a customer being difficult.
The agent sees the last twenty messages. On the third try it hands over rather than answering a fourth time.
The question could mean two things
"Cancel" might be the subscription or the order.
Ask one clarifying question. If that does not settle it, escalate and put both readings in the note.
The minutes after
What the agent says while nobody has arrived
The handoff is not an instant. On most teams nobody is watching a queue, and the customer finds that out by waiting. What the agent says at that moment decides whether they stay in the thread or leave to look for a phone number.
After an escalation, Optlo moves the conversation into a waiting workflow. It answers what it can still find in your sources and otherwise says the team has been notified, and it stops the moment a person joins.
Who pays
On per-resolution pricing, the handoff is the outcome nobody bills for
Zendesk and Intercom charge each time their AI resolves a conversation. Their own help centers explain what happens to that charge when a person takes over, and the two answers are different.
Intercom
The AI resolves it
$0.99
per Fin resolution
A person takes over
$0.99
Still billed if a teammate joins after Fin has answered and the customer did not ask for a person. Free only when Fin escalates on its own.
"A resolution is counted when, following Fin's last answer in a conversation, the customer either confirms the answer was satisfactory, or exits the conversation without requesting further assistance."
Intercom Help, Fin AI Agent outcomesZendesk
The AI resolves it
$1.50
per automated resolution
A person takes over
$0
Not counted. An escalation a person finishes does not come off your resolution allowance.
"You pay only for customer requests that were successfully resolved by an AI agent, without any escalation to a human agent."
Zendesk help, About automated resolution tiersOptlo
The AI resolves it
$0.03
per conversation past your plan, plus about 2¢ of model tokens at your provider's list rate
A person takes over
$0.03
The same. A person joining adds nothing to the bill.
Per-resolution pricing pays the AI to finish, not to hand off.
Neither rule is unfair. But a vendor paid per resolution has a reason to tune its agent to keep going, so check how quickly it hands off before you trust it. Optlo has no such meter. You pay per conversation, and the model tokens go to your own provider account, so a handoff costs the agent nothing. The ticket deflection page makes the same point about the deflection number.
Measuring it
Escalation rate is a map, not a score
Escalation rate is the share of conversations the agent hands to a person, the mirror of the deflection rate. A low rate is not a target. An agent that refuses to escalate posts a lovely deflection figure and a falling satisfaction score, because the customers it deflected gave up. Read the two together, as the metrics guide lays out.
Read it by topic. A rate that climbs on one subject is pointing at a hole in your help content or your workflow, and filling the hole moves the number on its own. Then measure what no dashboard shows. Once a week, read ten handed-off conversations and count how often your team asked something the customer had already said. That count is the quality of your handoff, and it should be zero.
Side by side
Where Optlo differs, and where it does not
Every tool here hands the thread to a person, shows them the conversation, and lets the customer ask. The difference is what a handoff does to the bill, and whether you can see the rule that fired it.
| What you are paying for | Zendesk | Intercom | Optlo |
|---|---|---|---|
| A person can take over the thread mid-conversation | Yes | Yes | Yes |
| Your team sees the whole conversation | Yes | Yes | Yes |
| The customer can ask for a person at any point | Yes | Yes | Yes |
| A handoff is never billed as a success | YesAssisted escalations are not counted | NoBilled if a teammate joins after Fin has answered | YesConversations are a capacity unit, not a success fee |
| AI charged at your provider's list rate | No | No | YesYou hold the key and the account |
| Every escalation rule is visible and editable | Partial | Partial | YesEvery rule is visible in the workflow editor |
Capabilities and prices checked September 2026. Vendors change plans often; the sources below are where each was read from.
Want one of these in depth?Optlo vs ZendeskOptlo vs Intercom
Getting started
A working handoff this afternoon
Three steps. The escalation logic comes wired into the default workflow, so the third one is mostly reading.
Connect your AI provider
Create a key with OpenAI, Anthropic or Google and paste it in. The 7-day trial includes AI usage, so this can wait until the trial ends.
Start from the default workflow
It comes with the handoff already wired in: it asks for an email first when there is none, and keeps the customer company while they wait. Open the handoff rules and write them in your own words: when to hand off, and what the note should say.
Take one over from the inbox
Ask your own agent for a person, then open the conversation in the shared inbox and reply. Check that the note is at the top, the thread is intact, and the agent has gone quiet. That is the whole test, and it is the one to run on any tool you are comparing.
Questions about handoff, answered plainly
What is a chatbot human handoff?
A chatbot human handoff is the moment a conversation passes from an AI agent to a person, with everything already said going along with it. It happens when the customer asks for someone, when the agent runs out of answers, or when a rule says a person has to make the call. Some people say handover instead. It is the same thing.
When should a chatbot hand off to a human?
The moment the customer asks, every time. Beyond that, hand off whenever a request needs authority the agent does not have, such as a refund, a plan change or anything to do with an account it cannot verify, and whenever the search behind the answer comes back empty. Write those as rules. Leave the model to judge the softer cases: a customer getting short with it, the same question asked three ways, a question that could mean two things.
What happens if nobody is online when the chatbot hands off?
On Optlo the conversation moves into a waiting workflow. It tells the customer the team has been notified, collects an email if there is none on file, and keeps answering the simple questions it can find in your sources. It stops doing that the moment a person joins the thread. The customer is not left with a spinner, and they are not told someone is coming when nobody is.
Can a person take over a conversation the AI has already started?
Yes. A team member opens the conversation in the shared inbox and replies. The workflow checks whether an operator is active on the thread and steps aside while one is. There is no separate ticket and no new window for the customer; the typing indicator simply belongs to someone else.
Do customers have to repeat themselves after a handoff?
They should not, and that is the test to run on any tool. Whoever picks it up should see every message, the reason the agent stopped, and who the customer is. Read ten handed-off conversations and count how often the person asks something the customer had already answered. That number is the quality of your handoff.
Is a chatbot handoff a warm transfer or a cold transfer?
A warm transfer is one where the person receiving the call already knows what it is about. A handoff that carries the transcript and an escalation note is a warm transfer. A handoff that replaces the chat with a ticket number and an email address is a cold one, and it is the version customers remember.
Does a handoff count as a resolution on per-resolution pricing?
It depends on the vendor, and the answer is in each one's help center. Zendesk does not count an assisted escalation. Intercom bills a conversation as resolved if a teammate joins after Fin has answered and the customer did not explicitly ask for a person, and bills a handoff made under your instructions as a separate outcome. Optlo bills per active conversation, so a conversation costs the same whether the agent finishes it or a person does.
Build AI customer support on your terms
Bring your own AI, stay in control, and optimize as you scale.
* No credit card required