AI-to-human handoff: rules, triggers and message templates

A good chatbot human handoff passes the conversation to a person at the right moment, with everything the bot already learned, to the team that can actually solve it, and tells the customer what happens next. Escalate whenever the customer asks for a person, when the bot cannot answer from approved sources, when the customer is frustrated, and when the case involves money, risk or judgement. Pass a short summary, the customer’s details and what the bot already tried.

Most teams get the first part right and the rest wrong: the bot hands over, but the agent starts with “How can I help you?”, the case lands in a general queue, and the customer waits overnight without knowing. This guide covers the rules, the triggers, the context to pass, routing, and the exact messages to send.

Why the handoff moment matters

Customers judge automation by its worst moment, and the worst moment is usually the handoff. Three things make it fail:

• The bot holds on too long. It loops through “Could you rephrase?” while the customer gets angrier.
• Context is lost. The person who takes over asks for the order number the customer typed five minutes ago.
• Silence after the handoff. “An agent will be with you shortly” at 11 p.m., and nobody replies until morning.

On WhatsApp there is also a policy reason. The Business Messaging Policy allows automated replies inside the 24-hour customer service window, but requires that businesses offer prompt, clear and direct escalation paths, for example an in-chat transfer to a human agent, a phone number, an email address or a support form. A bot without a way out is not only bad service; it is outside the rules.

When should a chatbot hand over to a human? 10 triggers

Split triggers into the ones the customer sets off and the ones your rules detect.

Customer-driven triggers

1. The customer asks for a person. “Agent”, “human”, “real person”, “can I talk to someone”. Hand over immediately, without a “Are you sure?” step.
2. Repeated failure. The bot could not answer two turns in a row, or the customer rephrased the same question.
3. Frustration. Capital letters, insults, “this is useless”, or a negative sentiment score if you use one.
4. Repeat contact. The third message about the same unresolved issue this week.

Rule-driven triggers

5. No grounded answer. The question is not covered by the knowledge base or tools. The bot should hand over or take a message rather than guess.
6. Money above a limit. Refunds, credits or discounts above an amount you set.
7. Sensitive topics. Legal threats, medical questions, safety issues, data deletion requests, complaints about staff.
8. High-value customers. Key accounts, VIP tiers, open deals above a value.
9. Process requires a person. Approvals, exceptions to policy, anything your team must sign off.
10. Technical failure. A tool times out or returns an error twice. Do not leave the customer with “something went wrong”.

Write these triggers into the agent’s instructions and, where possible, into the flow around it. Instructions guide a model; flow rules enforce. A rule like “refund over 100 goes to Billing” is safer as a condition in your flow than as a sentence in a prompt.

What context should the bot pass to the human agent?

The human should be able to reply usefully in the first message. Give them:

Context	Example	Why it matters
Reason for handoff	”Customer asked for a person” / “Refund above limit”	Sets the agent’s first move
One-line summary	”Order 7315 delivered damaged; wants replacement, not refund”	Saves reading 20 messages
Customer identity	Name, phone, account or order number, verified or not	Avoids asking again
What the bot already tried	”Sent returns policy; offered return label; customer declined”	Avoids repeating it
Collected data	Photos, model numbers, dates, address	The agent can act immediately
Sentiment and history	”Second late parcel this month”	Changes the tone and the offer
Priority and SLA	”Urgent, first reply due in 15 min”	Puts it in the right place in the queue

The full transcript should stay visible too. Summaries help, but agents should be able to check what the bot actually said, especially if the bot made a mistake.

A good internal summary looks like this:

Handoff: refund above limit (Billing)
Customer: Omar Haddad, verified by order number
Issue: charged twice for invoice 1042 (two identical card charges on Oct 2)
Bot actions: confirmed both charges in billing system; explained refund policy
Customer wants: one charge refunded today
Sentiment: calm, second contact

How to route the conversation to the right person

A handoff to “the team” is half a handoff. Decide routing rules before launch:

• By topic or intent. Billing, technical support, sales, returns. The bot already knows the topic, so use it.
• By language. If a customer writes in German, route to someone who speaks German.
• By customer value. Key accounts go to their account manager, not the general queue.
• By skill. Product specialists, senior agents for complaints.
• By workload and availability. Load-balance across online agents; skip people who are away.
• By time. Outside working hours, queue for the morning team with a promised reply time.

See routing and handover in the glossary for short definitions.

Pair routing with an SLA: a target for the first human reply after handoff. If nobody picks it up close to the deadline, alert a supervisor or reassign it. A handoff that sits unanswered is worse than no bot at all.

What to tell the customer: handoff message templates

Keep handoff messages short, specific and honest. Name the team, set the expectation, and confirm what you already have.

Customer asks for a person (inside working hours):

Of course. I am passing you to our support team now, together with your order details, so you will not need to repeat anything. A teammate usually replies within a few minutes.

Bot cannot answer:

I do not have a reliable answer to that, and I would rather not guess. I have passed your question to our product team. They will reply here in this chat.

Complaint or frustration:

I am sorry this has happened again. I have flagged your case as a priority and a supervisor is taking over now. They can see your earlier messages about it.

Outside working hours:

Our team is back at 9:00 tomorrow. I have saved your request with the photos you sent, and it is first in the morning queue. You will get a reply here by 10:00.

Money or approval needed:

Refunds of this amount are approved by our billing team. I have sent them your order and the reason. They will confirm here today.

And the first message from the human should prove the handoff worked:

Hi Omar, Selin from billing here. I can see both charges on invoice 1042. I have started the refund for one of them and will confirm here when it is completed.

After 24 hours: the template for late replies

On WhatsApp, once 24 hours have passed since the customer’s last message, you can only write with an approved template message. If your team might reply late, prepare a utility template in advance:

Hi {{1}}, we have an update on your request {{2}}. Reply to this message and we will continue right here.

On Facebook Messenger, Meta’s policy includes a human agent tag that lets a person reply manually within 7 days of the customer’s message. Instagram also lets a human agent’s reply be tagged so it can be sent outside the 24-hour window. Neither is meant for automated messages. Our guide to Messenger’s 24-hour rule and message tags covers the details.

Handing back to the bot

Handoff is not always one-way. After a person resolves the exception, the bot can take back routine steps: send a survey, a payment link, a delivery update. Two rules:

• Tell the customer. Meta’s Messenger policy, for example, expects disclosure that a person is talking to an automated service where the law requires it, including when a chat moves from a human back to automation. It is good practice everywhere.
• Do not let the bot interrupt. While a person owns the conversation, automation should stay quiet unless the agent hands it back.

Common handoff mistakes

• Hiding the human option behind three menus.
• Asking “Are you sure you want to talk to an agent?”
• Handing over without a summary, so the agent starts from zero.
• Routing everything to one general queue.
• Promising “right away” at night.
• Letting the bot keep replying after the person has taken over.
• Never reviewing handed-over conversations to improve the bot.

That last point is the most valuable one. Every handoff tells you something the bot could not do. Tag the reason, review them weekly, and decide whether to add knowledge, add a tool, or keep it with people. Our guide to testing an AI agent before launch includes handoff tests you can run before go-live.

How handoff works in MonoChat

In MonoChat, bots, AI assistants and agents all run inside the same shared team inbox, so a handoff is a change of owner, not a move to another tool.

Handover with the full history. When a flow, an AI assistant or an Agent Harness agent hands over, the conversation moves to your team with its full history, and your agents reply from the same WhatsApp number, Instagram account or other channel. Every bot reply stays in the conversation next to your team’s messages, so you can review and coach. MonoChat keeps a unified customer profile with cross-platform history, so an agent also sees what the customer wrote on other channels.

Triggers in the flow. In workflow automation, a flow can hand over to your team at any step. Combine conditions on customer profile, conversation signals and business rules to decide when, and use timers for follow-ups and SLA escalations.

Routing to the right person. Smart agent routing detects intent, topic, language and priority and assigns the conversation to the right team, queue or agent, with a conversation summary and the customer’s history. Team leads can also assign by hand.

SLA and escalation rules. Escalation and SLA workflows track first response and resolution SLAs with business-hour rules, escalate overdue cases, complaints, VIPs and negative sentiment, and route them to senior agents or managers.

A ready-made handover setup. The MonoChat Start assistant hands over to a person when asked, when the customer is unhappy or when it has no answer. You choose a team or one person, and outside working hours it says when you are back and takes a message.

You can set this up on the free Basic plan (1 agent, 1 channel) at app.monochat.ai; see pricing for larger teams.

Sources

• WhatsApp Business Messaging Policy
• Messenger Platform and IG Messaging API policy (Meta for Developers)
• Instagram API: send messages (Meta for Developers)
• Pricing on the WhatsApp Business Platform (Meta for Developers)