How does AI tell an emergency repair from a routine one?

Not by understanding plumbing. By working through a decision tree you defined, asking specific questions and acting on the answers. That is a less impressive description than people expect, and it is the reason it can be trusted unsupervised.

The categories that matter

Most operators already have an emergency definition, even if it lives in a handbook nobody has opened in a while. The build is largely a matter of writing it down properly.

• Water that is actively running or spreading, as distinct from a drip or a slow drain.
• No heat in cold weather, or no cooling in dangerous heat.
• Anything electrical: sparking, burning smells, exposed wiring, a dead panel.
• Gas, in any quantity, with no judgement call attached.
• Security: a door or window that will not secure, a broken lock, a gate stuck open.
• Anything where somebody is hurt or cannot safely stay in the unit.

Why the order of questions matters

A resident reporting a problem describes the symptom, not the severity. Water in the kitchen could be a burst supply line or a blocked sink, and those are very different calls.

So the questions narrow rather than confirm. Is it still running. Can you shut it off. Is it coming through the ceiling or pooling on the floor. Each answer eliminates a branch, and the sequence reaches a category in three or four exchanges without the resident having to know any of it.

Erring on the side of escalation

The rules are asymmetric on purpose. Escalating a routine issue costs somebody a phone call. Failing to escalate a real emergency costs a flooded unit and a resident who no longer trusts you.

So ambiguity escalates. If the answers do not clearly place something as routine, it goes to a person. That is a deliberate choice to accept some false positives, and it is the right trade in this category even though it would be wrong in most others.

What the resident experiences

From their side it should not feel like triage. It feels like somebody picked up, asked sensible questions and told them what happens next.

That last part is what actually reduces follow-up calls. A resident who has been told a technician will be in touch in the morning, and who has a work order reference, does not call back three times overnight. Being told nothing is what produces repeat calls.

Tuning it after launch

The first version will get some calls wrong in both directions. That is expected, and it is what the weekly review is for: read the week's escalations, find the ones that should not have woken somebody and the ones that should have, and adjust the rules.

Every change is re-tested against the scenario that produced it before it goes back out, so a fix for one case does not quietly break another.

Seasonal rules are worth writing down

One category that catches people out: what counts as an emergency changes with the weather, and a fixed rule set will be wrong for half the year.

No heat is an emergency in January and an inconvenience in May. No cooling is the reverse in a hot climate, and in some places it is a genuine health risk for older residents rather than a comfort issue. A frozen pipe warning matters for about four months and not at all for the rest.

Those thresholds are easy to encode once somebody decides them, and easy to forget entirely. It is worth agreeing the seasonal variations during the build rather than discovering in November that heat complaints are being filed as routine.

What the resident is told

Triage that happens invisibly is unsettling. Somebody who describes a problem and is then told a work order will be raised, with no sense of whether that means tonight or next week, will call back.

So the classification should be communicated in plain terms: either somebody is being contacted now, or this is scheduled and here is when to expect contact. That single sentence prevents most repeat calls, and it costs nothing.