The hand-over decision inside an AI agent: the four kinds of trigger (explicit request, rule match, low confidence, system failure), what the customer sees, what happens to the chat afterwards, and how to tune it for your business
The most important thing an AI agent does is know when to stop. A good hand-over is invisible to the customer: they asked for a person, or raised something delicate, and a person appeared. A bad one is either an agent arguing with an angry customer or an agent handing over every second message. Here is how the decision is actually made.
Quick answer
An AI agent hands a conversation to a person for one of four reasons: the customer asked for a human; a business rule matched (a complaint, a refund, a safety word, a topic you named); the agent's knowledge did not cover the question and it would rather admit that than guess; or something failed (the AI provider errored, or the customer repeated the same question several times). In each case the customer gets one honest message, the chat is marked "Needs you" in the inbox, the owner is notified, and the agent stays silent on that chat until a person hands it back. ReplyKit applies the first, third and fourth for every agent and lets you write the second in plain English.
The four triggers
| Trigger | Example | Who sets it |
|---|---|---|
| Explicit request | "Can I speak to a person?" "Is this a bot?" with a request for a human | Platform, every agent |
| Business rule | Complaint words, refund or discount requests, safety terms, topics you list | You, in plain English |
| Low confidence | The retrieved knowledge does not answer the question | Platform, every agent |
| Failure | AI provider error; the customer rephrases the same question three times | Platform, every agent |
Writing the business rules is covered in escalation rules for your agent.
The confidence trigger
Before replying, the agent retrieves the parts of your knowledge that match the question. If nothing relevant comes back, or what comes back does not answer it, the agent is instructed to say so and offer a person rather than compose a plausible answer. This is the mechanism that keeps invented answers rare; the trade-off is a "not sure" for questions you have not written down yet, which the gap list then shows you. The retrieval side is explained in how retrieval works and the gap side in the teach loop.
What happens after
- The customer gets one message: a person will reply, and when.
- The conversation is flagged "Needs you" in the inbox with the full transcript; the owner gets the handoff notification.
- The agent stops replying on that conversation. No auto-replies while a person is typing.
- A person replies from the inbox (or from the phone, since the number is theirs), and marks the chat back to the agent when done.
- If nobody replies, the chat stays flagged; the agent does not quietly resume. The mechanics are in AI-to-human handover.
Tuning
- Too many hand-overs. Read a week of them. If most are questions the knowledge could answer, the knowledge is thin; add lines. If a rule is firing on innocent words ("problem" in "no problem"), narrow it.
- Too few. If complaints are getting cheerful auto-replies, add the complaint words and phrases your customers actually use to the rules, and test with real past messages.
- Response promise. The hand-over message states when a person replies; set it to something you keep. Expectation wording is in setting expectations.
- Review the hand-over rate monthly; what good looks like is in metrics that matter.
Frequently asked questions
When does an AI agent hand over to a human?
When the customer asks for one, when a business rule matches, when the knowledge does not cover the question, or when something fails.
Does the agent keep replying after handing over?
No. It stays silent on that conversation until a person hands it back.
What does the customer see?
One honest message that a person will reply, and when.
Why does it say "not sure" instead of answering?
Because the retrieved knowledge did not answer the question and admitting that is safer than guessing. Add the answer and it will know next time.
Can I add my own triggers?
Yes, in plain English in the agent's instructions.
Knows when to stop
Four triggers, one honest message, a flagged chat and a silent agent until you hand back. 7-day free trial, no card.
Sources and further reading
- Harvard Business Review, The Short Life of Online Sales Leads (Oldroyd, McElheran, Elkington, 2011) · firms that responded within an hour were about seven times more likely to qualify the lead
- WhatsApp Business Messaging Policy · opt-in, opt-out and prohibited-use rules every business must follow
ReplyKit is an independent product and is not affiliated with or endorsed by Meta or WhatsApp. This article is general guidance for running a small business, not legal, financial or regulatory advice.