Escalation rules for an AI agent: the triggers that must hand over, the severity levels, who gets alerted and how fast, the wording the customer sees, and a template to adapt
Every good AI agent has a short list of things it never handles, and the quality of that list decides whether the agent is trusted. Too short and a complaint gets a cheerful auto-reply; too long and everything is "let me get a person". Here is how to write it so it is right for your business.
Quick answer
Write escalation rules in four parts: triggers (the customer asks for a person; a complaint or anger; anything about money beyond the price list; safety, medical or legal matters; sensitive personal topics; and anything the agent is unsure about), severity (urgent gets an immediate alert to a named person; normal waits for the inbox), the customer-facing line the agent says when handing over, and what the agent does afterwards (stays quiet on that chat until a person hands back). Keep the list to ten lines and test each trigger with a real message. ReplyKit applies the universal triggers by default and lets you add your own in plain English.
The triggers
| Category | Example | Severity |
|---|---|---|
| Asks for a person | "Can I talk to someone?" | Normal, immediate hand-over, no argument |
| Complaint or anger | "This is the third time" | Normal, often urgent for a business that cares |
| Money beyond the list | Refunds, discounts, disputes, custom quotes | Normal |
| Safety or emergency | A leak, an injury, a gas smell, a threat | Urgent, alert now |
| Medical, legal, financial advice | "Is this normal?" "Can I sue?" | Normal; the agent declines and offers a person |
| Sensitive personal topics | Distress, abuse, a child's welfare | Urgent, alert now, scripted safe response |
| Uncertainty | The knowledge does not cover it | Normal; the agent says so and offers a person |
| Repeated failure | Customer rephrases the same question three times | Normal |
| VIP or high value | A keyword, a large order | Your choice |
The first, the last two and uncertainty are handled by ReplyKit's platform rules for every agent. The rest you write for your business; a clinic's safety list differs from a plumber's, as the plumbers guide and counselling guide show.
Severity and alerts
- Normal. The chat is marked "Needs you" in the inbox and the owner gets the handoff email, unless notifications are set to in-app only.
- Urgent. Same, plus a line in the customer-facing message that a person will call, and a named person who watches the inbox for that category. ReplyKit does not phone you; the urgency is in who is watching and how fast you promise to respond.
- Response promise. Written into the hand-over line: "within the hour", "today", "a person replies between 9 and 6". Keep it; the expectation wording is in setting expectations.
What the customer sees
Standard hand-over
I'll get a member of the team to help with this. They'll reply here {within the hour / today}. In the meantime, is there anything else I can answer?
Complaint
I'm sorry this happened. I've passed it to {name} with everything you've written, and they'll reply here {today}.
Safety
That sounds urgent. {One line of safety guidance you wrote.} {Name} will call you within {n} minutes; please keep your phone nearby.
Declined advice
That's something {the doctor / a licensed adviser} should answer rather than me. I can book you a consultation; would {slot} work?
After handing over, the agent stays silent on that conversation until a person hands it back, so there is no auto-reply talking over a human; the mechanics are in AI-to-human handover.
The template
ESCALATION RULES
Hand over immediately, and say a person will call within {n} minutes, if a message mentions: {your safety list}.
Hand over, and say a person replies {today}, for: complaints, refunds, discounts, custom quotes, {your list}.
Never give {medical / legal / financial} advice; say {the professional} covers that and offer to book.
If a customer asks for a person, hand over at once without asking why.
If you are not sure, say so and offer a person; do not guess.
After handing over, do not reply again in that conversation until it is handed back.
Test each line with a real message and check the hand-over fires; the routine is in testing before going live.
Frequently asked questions
What should always trigger a hand-over?
A request for a person, complaints, money beyond the price list, safety or emergencies, medical or legal or financial advice, sensitive personal topics, and anything the agent is unsure about.
Does the agent alert me?
The chat is marked "Needs you" and the owner gets a handoff email. Urgent categories need a named person watching the inbox and a response promise in the wording.
What does the agent do after escalating?
Stays silent in that conversation until a person hands it back.
Can I add my own triggers?
Yes, in plain English in the instructions. Test each with a real message.
How many rules is too many?
More than about ten usually means the agent hands over things it could answer. Keep the list tight and specific.
Safe by rules, not by luck
Universal triggers built in, your own added in plain English, every hand-over logged. 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.