Can an AI receptionist handle emergency calls?

Why this is the first rule we write

An AI calmly offering a slot three days out to someone with a flooding bathroom or a dead furnace in January is worse than voicemail. Voicemail at least signals that nobody is handling it; a confident wrong answer does not.

So triage is built and tested before anything else, and it errs toward urgency. A false emergency dispatch costs a call-out. A missed one costs a customer, a reputation and sometimes a property.

How triage actually works

1. Listen for the signal. Specific language, per trade — water flowing now, no heat, burning smell, smoke, no power.
2. Ask the dispatcher's question. One or two questions that separate urgent from routine. "Is water coming out right now?" does more work than any amount of model tuning.
3. Give the safety instruction where there is one — find the shutoff, kill the breaker, leave the property.
4. Capture what matters. Address, callback number, nature of the fault.
5. Reach a human immediately. On-call tech, with context already gathered.
6. Confirm to the caller what happens next and when.

What it must never do

• Diagnose. Not on a phone call, not ever for electrical or gas.
• Talk someone through a repair.
• Book an emergency as a routine appointment.
• Decide an emergency is not urgent. When uncertain, escalate.
• Leave a caller without knowing what happens next.

How to test it before you trust it

Call your own system and describe the worst scenario in your trade. Then describe it badly — the way a panicking customer actually would, vague and out of order. Then describe something that sounds urgent but is not.

If it handles all three correctly it is ready. If it fails any of them, the rules are not finished. This is the test to run on any vendor's demo, not the scripted one they offer you.