Tier zero: the things that must never be triaged
Gas. Carbon monoxide. An electrical or burning smell. Anything on fire. These are not urgency categories to be weighed against schedule availability; they are safety events, and the correct behaviour is fixed.
The agent gives a scripted safety instruction, which for a suspected gas or carbon monoxide event means leave the building and call the emergency number or the utility, and it escalates to a human immediately. It does not attempt a diagnosis, it does not ask clarifying questions to narrow the fault, and it does not offer an appointment slot.
This is the part that has to be built first and tested hardest, and it is the reason "it uses AI" is not a sufficient answer to how a system handles emergencies. The right design here is deliberately not intelligent. It is a hard rule that fires on a keyword and gets out of the way.
The short versionSafety events are not triage. Scripted instruction, immediate human escalation, no diagnosis, no booking.
Tier one: genuine emergencies, dispatch tonight
No heat in freezing conditions, or no cooling in dangerous heat, becomes an emergency based on who is in the building. A household with an infant, an elderly resident, or someone with a medical condition is a different call from a healthy adult in a mild snap, and the agent has to ask.
Also here: water actively escaping where it should not be, because the damage compounds by the hour, and a complete failure at a commercial site where the business cannot trade or stock is at risk.
The question that separates this tier from the next is not "how bad is the fault" but "what happens between now and morning". A technician knows this instinctively. A system has to be told.
The short versionAsk who is in the building and what happens by morning. Those two answers decide the tier, not the fault code.
Tier two: urgent, first slot tomorrow
Partial failures, intermittent faults, and a system that is running but clearly on its way out. Uncomfortable rather than unsafe, and mild weather rather than extreme.
The agent books the earliest genuine slot, captures the diagnostic detail, and says plainly when someone will arrive. The specific thing to avoid is a promised callback, which is where most goodwill is lost: the caller has been handled but not resolved, and they will keep ringing other numbers while they wait.
A booked time, even a less convenient one, ends the caller's search. That is the whole job.
The short versionA real slot beats a promised callback, even when the slot is later than the caller wanted.
Tier three: routine, book it properly
Maintenance, quotes, seasonal service, filter changes, a question about a bill. These are the calls that should never have woken anybody, and they are the bulk of the volume.
The value here is not urgency handling, it is that the call gets captured at all. These are the ones that arrive at nine in the evening, go to voicemail, and quietly do not become customers.
The short versionMost calls are routine. Capturing them at all is the return, not clever triage.
What the agent has to capture, whatever the tier
System type, make and model where the caller can find it, any error code on the display, what the symptom actually is in the caller's own words, when it started, and what changed just before it started.
That last one is the question amateurs leave out and technicians always ask. "It started after the power cut" or "it began when we turned the heating on for the first time this autumn" changes what goes on the van.
Capturing this properly is most of the value of the whole system. A booked job with no diagnostic detail is a technician arriving twice.
The short versionMake, model, error code, symptom in their words, when it started, and what changed. Especially the last one.
How to test any vendor on this, in one phone call
Ring the demo line and describe a smell of gas. The correct response is a scripted safety instruction and an immediate handoff. If it offers you an appointment slot, you have your answer and you can stop the evaluation.
Then try an ambiguous one: no heat, mild weather, elderly resident. A good system asks who is in the building. A generic one classifies on the words "no heat" alone and either wakes a technician or misses a real emergency.
Then try a caller who does not know the words. Most people do not say "the condenser fan is not spinning"; they say it is making a noise and blowing warm. A system trained on tidy phrasing falls apart here, and real calls are never tidy.
The short versionThree test calls: gas, ambiguous no-heat, and a caller with no vocabulary. They separate trade knowledge from a script.
When the honest answer is a human
If your call mix is mostly complex commercial diagnostics, or your work is unusual enough that no two calls resemble each other, the triage layer has little to work with and a person is the better answer.
The same is true if your volume is low and you or a family member genuinely answer every call already. There is no system that improves on that, and we would tell you so rather than sell you one.
The case for an agent is strongest where volume is high, the pattern is repetitive, and the misses cluster outside working hours, which describes most residential service businesses and very few specialist contractors.
The short versionHigh volume, repetitive pattern, after-hours misses. If that is not you, an agent is probably the wrong purchase.
