Five escalation rules to write before your agent goes live

Most people set up an AI agent by writing down what it should answer: opening hours, prices, how to book. Almost nobody writes the other half, which is where the trust actually lives — what it must never answer on its own. An agent can handle a hundred flawless conversations this week and sink you with the hundred and first, the one where it invented a delivery date or brushed off a serious complaint. The fix isn't technical. It's a writing exercise, and it takes about half an hour.

This is for you if you already have an agent running, or you're about to switch one on. By the end you'll know which five situations shouldn't be automated, how to phrase them so they actually trigger, and who gets pinged when they do.

"If it's not sure, escalate" is not a rule

It's the most common instruction and the most useless one. An agent doesn't feel uncertainty the way you do — it processes a message and produces a plausible answer. Asking it to flag things when it "isn't sure" asks it to introspect something it doesn't experience. You get one of two failure modes: it never escalates, or it escalates everything and you're back to answering every message yourself, which is what you were trying to stop doing.

A useful rule hangs on something observable: words that appear in the message, amounts, data types, how many replies have already gone out. Not internal states. "If the customer writes complaint, lawyer, review or refund" is a rule. "If it detects anger" is a wish.

How to make the handover feel seamless to the customer is a separate topic. This piece is about the step before it, the one nearly everyone skips: deciding and writing down when the thing fires.

Four boxes per rule, no more

Every rule fits in four lines. If you need more, you've folded two rules into one.

1. Trigger. What has to show up in the message or the data. Specific and checkable.
2. What the agent does. The exact sentence it says, and what it stops doing (no price, no date, no closing the conversation).
3. Who it tells. A named person. Never "the team".
4. How fast. The window the customer can count on, and that you'll actually hit.

Worked example:

• Trigger: the customer asks for a quote on a job over €3,000.
• What the agent does: "For a job this size I'd rather have Marta look at it and give you a firm number. Can you send me two photos and your postcode? She'll call you today." No figures, no ballparks.
• Who it tells: Marta, on the internal WhatsApp, with the full thread attached.
• How fast: same working day.

Note what's happening there: the agent keeps doing something useful — collecting photos and a postcode — while it waits. Escalating isn't abandoning the customer with "someone will be in touch".

The five that cover most small businesses

Adapt them to your trade, but don't go below this set.

1. Money outside the range

Set a ceiling. Below it the agent quotes from your price list; above it, it escalates. The number depends on your average job: €150 in a salon, €800 in a garage, €3,000 on a renovation. Getting the threshold exactly right on day one doesn't matter — having one does. Same logic for discounts, payment terms and anything that departs from the published prices.

2. Anger, complaints and review threats

An unhappy customer is the highest-value conversation you'll have this week and the worst possible fit for an automated reply. The trigger here is vocabulary (complaint, refund, lawyer, consumer complaint, review, scam) plus a structural one: it's the third message in a row about the same problem. The agent acknowledges, doesn't defend, and pings a human. No "I understand your frustration, however".

3. Health, safety and legal

If you run a clinic, a vet practice, a gym, or anything where a customer might ask a question with physical consequences, this one isn't optional. Symptoms, medication, allergies, pain, a gas appliance that smells wrong: the agent gives logistics (when you're open, how to get an urgent appointment) and never judgement. Same for legal and tax questions in a firm or an accountancy practice.

4. Anything it can't verify

If the answer depends on something the agent can't look up live — exact stock, whether that part lands on Thursday, whether an order shipped — it must not answer from memory. Either wire it to the real source (your calendar, your shop, your system) or escalate. This is the rule that prevents the most silent damage, because a made-up date doesn't look like a mistake until Thursday arrives.

5. The two-attempt rule

If the agent has replied twice and the customer asks the same thing again, the conversation isn't working. Doesn't matter why. Third time, a human takes it. It's the only rule here that ignores content entirely, and that's exactly why it earns its place: it catches all the odd cases you failed to predict.

Who gets told, on which channel, and what happens at 10pm

This is the part everyone skips, and it's what decides whether the rules work at all. A rule that escalates into a mailbox nobody reads is worse than no rule, because now the customer is waiting on something you promised.

Pin down four things for each one:

• A named person and a backup. If Marta's on holiday, then who?
• A channel that person genuinely watches. Usually WhatsApp or Slack. Almost never info@.
• What gets sent. The whole thread, not a bare alert. They shouldn't have to ask what it's about.
• Out of hours. A ping at 10:40pm on a Friday isn't an emergency. Have the agent tell the truth — "Marta will come back to you first thing" — instead of a window you won't meet.

When not to write a rule at all

Three cases where adding rules makes things worse.

When it fires every single day. If your agent escalates twenty times a week for the same reason, that's not an escalation — it's a missing answer. Put it in the agent's FAQ playbook and delete the rule.

When you're insuring against something that's never happened. It's tempting to write a rule for every imaginable disaster. Each extra rule makes the whole set less reliable and slows the agent down. If it hasn't come up in a year, wait until it does.

When the real answer is not to automate that channel. If a whole category of conversation escalates almost every time — complex quotes, contract negotiation — skip the middleman and route it to a person from the start.

---

Write the five this week, in a shared doc, with real names on them. Then set a reminder for thirty days out and count how often each one fired. The ones that never fired: delete. The ones that fired daily: turn into answers. What's left is your business judgement, finally written down — which was the point all along.

If you're setting up the agent that handles first contact — WhatsApp, website, phone — these five rules are genuinely the most important part of it. You can see how it works on the Receptionist page. You approve, it executes.