When should a chatbot hand over to a human? Eight situations and response templates
A practical handover policy for small support teams: eight escalation triggers, honest response templates, and a workflow for customers who arrive when nobody is available.
A chatbot should request a human when the visitor asks for one, repeated attempts have not helped, the available information cannot support an answer, or the next step requires authority the assistant does not have. A handover succeeds when someone accepts responsibility for the next step. Displaying “connecting you” is only a message.
For a small team, a useful policy fits on one page: when to escalate, who checks the inbox, what to tell a waiting visitor, and what happens outside staffed hours. The following examples are templates to adapt and test, not promises of automatic behaviour in every conversation.
Eight situations worth defining
| Situation | What the assistant should do |
|---|---|
| An explicit request for a person | Start the available handover without another troubleshooting loop |
| Repeated unsuccessful help | Stop repeating advice and preserve the context |
| Missing or conflicting information | State what cannot be confirmed and request a check |
| An action outside its authority | Explain the boundary and involve an authorized person |
| An unresolved identity check | Keep access controls in place and offer a supported recovery route |
| A complaint requiring discretion | Let a person assess the case before promising a remedy |
| The offered interaction does not work for the visitor | Offer a usable alternative and human help |
| Several teams must coordinate | Give the case a clear owner and next step |
1. “I would like to speak to someone”
An explicit request is enough. The visitor may already know that a person must decide the issue. Do not require a reason, email address, or a second attempt at the FAQ before requesting help.
After the system confirms that the request was created, suitable wording is:
I have requested help from our team. You can add any further details here while you wait.
If the request fails, say so and offer your published contact route. “An agent has joined” belongs only after an agent actually joins.
2. The suggested solution has failed repeatedly
A customer says they have followed the instructions and the error remains. Repeating the same steps in different words does not move the case forward. Define a practical rule for your team, such as escalating after a second unsuccessful attempt at the same issue.
The steps we tried have not resolved this. This needs a closer look from our team, so you do not have to repeat the same troubleshooting again.
Test that your assistant follows this rule; the instruction alone does not guarantee it will.
3. The sources cannot establish the answer
Suppose one page says that Saturday delivery is available and another excludes it. Or a product description omits the measurement the visitor needs. The assistant should identify the uncertainty instead of guessing from a similar product.
I cannot confirm Saturday delivery for this order from the information available. Our team needs to check it before you rely on a delivery date.
Record the source gap afterwards. Our guide to fixing an AI agent's answer explains how to turn a repeated escalation into better supporting information.
4. The next step exceeds the assistant's authority
Explaining a cancellation policy is different from cancelling an order. The assistant may perform an action only when the necessary integration, permissions, and confirmation steps exist.
I can explain the cancellation process, but I cannot cancel this order here. A member of our team needs to handle that request.
An attempted action is not a completed action. If an integration fails, do not tell the visitor that the refund, address change, or cancellation succeeded.
5. The visitor cannot complete verification
A mistyped address or an inaccessible mailbox can interrupt a legitimate request. Human help is appropriate, but it must not become a shortcut around ownership checks.
I cannot show the order details without completing verification. Our team can explain the alternative verification steps available for your account.
Your staff need their own approved recovery procedure. A name or order number typed into chat is not, by itself, proof of ownership.
6. A complaint needs an individual decision
A published policy may answer where to send a return. It cannot decide every request for an exception, compensation, or a negotiated solution. Let a person assess the circumstances.
I understand that the item arrived damaged. Our team needs to review what happened and confirm the next step. I cannot promise a particular remedy before that review.
Ask only for information needed at this stage. If evidence must be sent through a separate channel, name a channel your business actually supports.
7. The current interaction is not usable
A visitor may be unable to complete a form, follow a long procedure, or use the offered contact method. Offer a simpler explanation if it helps, and respect a request for personal assistance without asking the visitor to disclose why they need it.
We can continue with our team instead. You do not need to complete that form to ask for help.
Keep an accessible contact alternative visible on your website as well as in chat.
8. More than one team must act
A business order might combine a delivery constraint, an invoice correction, and a custom quotation. An assistant can clarify the request, but somebody must coordinate the dependencies.
This request involves delivery and billing. Our team needs to review both before confirming a solution.
Choose an owner internally. Do not promise that a case has been assigned to a named department unless that assignment has happened.
Make the handover work in Breezaro
An operator can open a conversation, read its history, and use Take Control to reply while the assistant pauses. AI-initiated requests require the Call an operator action to be enabled. A successful request pauses AI replies and creates an inbox notification; visitors can continue writing. Optional push and email alerts depend on your settings and delivery conditions. See the Take Control guide and operator action setup.
Take Control is available on Pro and Business and during an active trial. Check the current plan comparison before choosing the workflow. Website chat and connected messaging channels have different delivery rules; review the relevant channel guide rather than assuming a reply can always be sent later.
When nobody is available
Publish the hours you can actually staff. A request does not prove that someone is online. Use a message such as “Our team handles enquiries during [published hours]. For follow-up, please use [supported contact route].” Replace the brackets with verified information and promise a response deadline only if your team can meet it.
When no operator is available, offer an existing contact page or another established contact route.
Before launch, test one successful handover, one failed notification, and one out-of-hours conversation. Confirm that an operator can find the request, understand what was already tried, reply, and return control to the assistant. Our chatbot testing checklist gives you more scenarios to add to that review.