It is 9:10 p.m. on a Tuesday in January. The overnight low is 11 degrees. A homeowner calls and says the heat quit an hour ago. Your office closed at 5. The on-call tech is at a dinner and has his phone on vibrate.
That one call tells you more about an AI receptionist than any demo. A product that handles routine tune-up bookings well can still fail it, because the right answer depends on rules only you can write. This post is the test, and the rules to write before you run it. The general buyer's guide, with the other five tests, is AI receptionist for contractors. Here the focus is a single call.
Why no-heat is the HVAC test
Most missed HVAC calls are not emergencies. A no-heat call in deep winter is the exception, and it combines everything hard about the phone line at once: urgency, risk, a caller who may not know what is wrong, a limited pool of technicians and a pricing conversation (an after-hours rate) nobody wants to have.
It is also the call your competitors answer. ServiceTitan's June 2022 data covering 3,000+ trade businesses showed large shops booking 61% of calls at peak and 21% after 6 p.m., and small shops going from 26% to 9% (ServiceTitan). That is one vendor's customer data from 2022, but the direction is the point: the evening call is the one that slips.
Write the triage ladder first
Before you call the vendor, write a one-page ladder. The AI should follow your rules, not a template. Here is a structure to adapt. The thresholds are examples, so set your own with your techs and, where the rules involve safety, with counsel and your insurer.
Level 0: Safety. Not a booking.
- The caller smells gas, hears a carbon monoxide alarm or reports symptoms such as headache, dizziness or nausea with a fuel-burning system running.
- The correct response is to tell the caller to leave the home and call emergency services or the gas utility, and then alert your on-call person. The agent does not offer an appointment slot as the first response.
- The CDC lists gas- and oil-burning furnaces and water heaters among common household sources of carbon monoxide and describes it as "odorless, colorless" and deadly without warning (CDC).
Level 1: Emergency dispatch.
- No heat and an outdoor temperature at or below your cutoff, or any vulnerable occupant: an infant, an elderly person, someone medically fragile.
- Triggers an immediate call to the on-call tech, with the caller kept informed.
Level 2: Urgent, next available.
- No heat but mild weather, or a partial failure (one zone out, weak airflow).
- The agent offers the first slot tomorrow and a clear after-hours option if you offer one.
Level 3: Routine.
- Noise, uneven temperatures, a tune-up request.
- Books normally.
If the AI vendor cannot encode a ladder like this, including a temperature input, it is not ready for HVAC. If your product depends on the caller deciding what is an emergency, you will get callers who label everything an emergency.
The call, minute by minute
Run this with a coworker playing the homeowner. Use your own staff voice, not an actor reading a script. Record it.
0:00 to 0:20, answer and greet. Does it answer within a ring or two, identify your company by name and disclose what it is? Whatever your counsel decides on disclosure belongs in the greeting. Our post on call recording and AI disclosure laws covers the state-by-state questions.
0:20 to 1:00, the safety screen. The caller says: "My heat's out. It's freezing in here." The agent should ask whether anyone smells gas, whether a carbon monoxide alarm has gone off and whether anyone feels ill, before anything else. If it launches into "what's your address and what day works for you," it failed.
1:00 to 2:00, severity. Does it ask what the thermostat reads and what the weather is, or whether anyone vulnerable is in the house? Does it identify the system type (gas furnace, heat pump, boiler) and whether anything has been tried? The thermostat reading and weather are your triggers for Level 1.
2:00 to 3:00, identify the customer. Does it match the caller's number to an existing record and recognize a maintenance agreement member? Members often get priority and a waived diagnostic fee. Does it behave differently for members, and does that match your policy?
3:00 to 4:00, handle the price question. The caller asks: "How much is this going to cost?" The agent should state your after-hours diagnostic or dispatch fee if you have one, say it cannot quote the repair, and get consent to continue. It must not invent a price, promise "no extra charge" or guess at a part. Get your office to approve this line verbatim.
4:00 to 5:00, dispatch or book. For a Level 1 call, the agent should say it is contacting the on-call technician and give an honest expectation. It should not promise an arrival time unless it can see real technician locations. Ask what it said. "Someone will be there within 30 minutes" without a basis is a failure.
5:00 to 8:00, the handoff. Does your on-call tech get a text or call within a minute with the caller's name, address, number, system type, the temperature at the house, who is in it and the recording link? Time it. Then the tech does not pick up. What happens next? A second number, an escalation, a callback to the homeowner explaining the delay? An agent that says "we'll call you back" and does nothing is the worst outcome of the night.
After the call. Is a record in your field software with the right job type and priority? What does the caller see: a confirmation text, an update when the tech is on the way? If texting is involved, the registration and consent rules in missed call text back for contractors apply.
Failure modes specific to HVAC
Every call becomes an emergency. A caller with a lukewarm house in 50-degree weather says "emergency." If the agent accepts it, your on-call tech runs a night of low-value calls and starts ignoring alerts. The cure is a rule tied to temperature and occupants, not the caller's adjectives.
Troubleshooting that creates liability. An agent that walks a caller through resetting a unit or opening a panel is taking on risk. Decide whether it may offer anything at all, such as checking the thermostat batteries or the breaker, or whether the rule is no instructions at all. Many operators allow only the simplest checks and require a human for anything touching the equipment.
Heat pump confusion. Frost on an outdoor unit and a defrost cycle can look like a failure to a caller. If you service heat pumps, test one.
Multi-location mix-ups. A caller in a ZIP code you serve from the other branch gets booked at the wrong one. If you have more than one location, test a call from each service area, plus one just outside.
The January surge. A real cold snap means ten no-heat calls in one hour. Ask the vendor what limits the number of simultaneous calls, then check your plan, because the limit can be a pricing tier. Our AI receptionist cost guide covers how overage and concurrency show up on the invoice.
Platform-specific booking gaps. If your agent books into ServiceTitan, check that it respects real technician capacity. ServiceTitan's own AI Virtual Agent page describes booking by job type, location and required skills against real-time availability, and says it can transfer to a human, while saying nothing about emergency calls (ServiceTitan). A third-party agent has to match that depth through its integration. We compare the two routes in ServiceTitan AI voice agents vs third-party AI.
Score it
Give each call a score against rules you wrote before testing. A simple sheet:
| Check | Pass | Fail |
|---|---|---|
| Safety screen asked before booking | Yes | No |
| Correct level assigned | Matches ladder | Mismatch |
| Price handled per script | Yes | Quoted or promised |
| On-call alerted within 60 seconds | Yes | No |
| Handoff includes full context | Yes | Caller repeats |
| No-answer fallback works | Yes | Dead end |
| Record in field software correct | Yes | Missing or wrong |
Any failure on the safety screen or the no-answer fallback ends the trial. Fix and retest; do not argue it away.
What to do next
Run the test twice: once on a cold night, with the settings you plan to deploy, and once as a second-tier scenario such as a mild-weather no-heat call, to make sure it is not treating everything as an emergency. Then route a slice of real after-hours calls and compare to what your team would have done.
If you want the rules written, the integration built and the weekly booked-job numbers reported, that is the work of the Booked-Jobs Recovery System from Di-Hy, and our AI customer support service covers the broader setup. For the earlier question of why calls go unanswered at all, see why contractors miss calls.
Frequently asked questions
Should an AI receptionist dispatch an HVAC emergency on its own?
It should alert your on-call process and keep the caller informed, using rules you wrote. Whether it can also commit a technician depends on whether it sees real availability. Many shops keep a human in the dispatch decision for emergencies.
What if the caller smells gas?
Your rule should send the caller to emergency services or the gas utility, not to the schedule, and alert your on-call person. Test this explicitly. It is the one scenario where booking is the wrong response.
Can it tell a real emergency from a caller exaggerating?
Only if your rules rely on facts the agent can collect, such as the temperature, who is in the home and the type of failure, instead of the caller's word for it.