How AI dispatch works for HVAC after-hours: the full playbook

An AI dispatcher for HVAC after-hours has one job: pick up every call in under a second, decide whether it's an emergency, and either page the on-call technician within 30 seconds or book the job into the next available slot. Nothing else matters at 2am — not pricing chat, not warranty questions, not lead-magnet upsells. The call is either a fire or it isn't.

This post walks through how that triage actually works in 2026: the decision tree, the integrations, the failure modes, and what separates a real dispatcher from a generic voice agent with an HVAC system prompt.

The decision tree

Every after-hours HVAC call falls into one of four buckets. The dispatcher's job is to figure out which one in 45 seconds or less.

Bucket	Examples	Action
Emergency	No heat below 32°F outdoor, no AC above 95°F, gas smell, frozen pipes, CO alarm	Page on-call tech in 30 sec; tell caller "tech will call back within 15 min"
Urgent	System cycling on/off, water leaking from indoor unit, breaker tripping repeatedly	Book first-available morning slot; flag for dispatcher review
Routine	Tune-up, filter, noisy outdoor unit, thermostat questions	Book into next standard slot
Quote / install	New system, replacement, upgrade, second unit	Book a 30-min in-home assessment

The taxonomy looks obvious on paper. The hard part is that callers don't describe their problems in the dispatcher's categories — they describe them in their kitchens at 11pm in February. A dispatcher that asks one clarifying question well is worth ten that ask none.

The 30-second emergency clock

For the emergency bucket, the dispatcher has a 30-second budget from "yes, this is an emergency" to "tech has been paged." Anything slower and the caller hangs up to find the next contractor.

The actual sequence:

1. Confirm the emergency. "You're saying you smell gas in the basement — is that right?" Single yes/no, no menu.
2. Capture the essentials. Address, contact number, ZIP. Hi Agent's standard prompt asks these three at once: "Can I grab your address, the best number to reach you, and your ZIP?"
3. Give the caller permission to act. For gas smells and CO alarms specifically: "Please get out of the house and call 911 from outside if you haven't already." The dispatcher is not the safety system, but it's the first system on the line.
4. End the call. "Your on-call tech will call you in the next 15 minutes from a 412 area code. Stay safe." Don't keep the caller on hold — the tech needs to be paged, and the caller needs to be free.
5. Page the tech. SMS + voice call to the on-call rotation, with the full call summary attached. Standard format: URGENT: Lisa Henderson, 412-555-0188, 14 Oak St, "smells gas in basement, lights still on". Hi Agent ended call at 23:47. ETA target: 00:02.

If any of those five steps fails, the chain breaks. Most generic voice AI agents fail at step 5 — they capture the data but don't page the human.

What "after-hours" actually covers

There's no single after-hours definition. In practice, four windows matter for HVAC dispatch:

• Evenings (6pm–11pm). ~22% of weekly call volume. Mostly routine but with a long emergency tail (post-dinner thermostat failures).
• Overnight (11pm–6am). ~7% of volume but ~40% of emergencies. Highest-stakes window.
• Weekends. ~18% of weekly call volume. The bucket that traditional shops underprice most — emergency premiums collapse on Sunday despite costs being higher.
• Holidays. Low volume but the highest emergency density. Thanksgiving and Christmas in cold-weather markets are reliably the busiest emergency days of the year.

A dispatcher that handles weekdays cleanly but fumbles weekends or holidays isn't a working dispatcher. The whole point is the time-of-day reliability.

Why generic voice AI fails at this

The voice-AI category in 2026 has three layers: infrastructure (Vapi, Bland, Retell, Synthflow), generalist assistants (Goodcall, generic ChatGPT phone), and verticalized receptionists (Hi Agent, Avoca, Smith.ai for legal). Generic systems fail at HVAC dispatch in predictable ways.

Failure 1: They miss the triage cues. A caller saying "my upstairs is wet" might be a routine leak or a burst supply line. Generic agents book the routine slot. A trained dispatcher recognizes the ambiguity and asks one qualifying question ("Is water actively flowing?").

Failure 2: They over-promise. Generic prompts encourage agents to be "helpful" — so they commit to "next-day service" because the prompt told them to. A real dispatcher knows the truck schedule and books accurately.

Failure 3: They handle the no-callback voicemail like a regular call. If a caller leaves a 4-second message and hangs up, a generic agent just files it. A dispatcher recognizes the pattern (probably hit voicemail by mistake or got cut off) and triggers a callback within 60 seconds, while the customer is still near the phone.

Failure 4: They don't page the human. This is the most common failure. The voice AI captures the emergency cleanly and dumps the transcript into a Slack channel that nobody is watching at 2am. The page-to-tech integration is the difference between an answering service and a dispatcher.

The integrations that actually matter

For an AI dispatcher to be useful in an HVAC shop, it needs to talk to four systems:

System	What for	Common providers
CRM / dispatch	Book the job, log the call, attach the recording	ServiceTitan, Housecall Pro, Jobber, FieldEdge
Phone system	Receive the call, hand off to humans when needed	RingCentral, OpenPhone, Twilio, native SIP
On-call paging	SMS + voice page to the tech rotation	Twilio, PagerDuty, custom on-call rotations
Calendar	Confirm slot availability before committing	Calendly, Google Calendar, embedded in CRM

The integration depth varies by vendor. The pattern to look for: does the booking land in the dispatch software while the customer is still on the call, or does it batch overnight? Real-time integration is the difference between a dispatcher and a glorified voicemail-to-text.

What this looks like in production

A typical HVAC shop running Hi Agent overnight sees the following pattern over a 30-day window:

• 230 after-hours calls answered (vs ~38 with previous voicemail + callback system)
• 19 emergencies dispatched, average page time 22 seconds
• 41 urgent/routine jobs booked into the next morning
• 28 install/quote leads captured (typically routed to the daytime sales rep)
• 142 routine routine routine — tune-up bookings, filter questions, thermostat troubleshoot, hang-ups

The 19 emergencies are the headline number. At an average residential HVAC emergency ticket of $1,400 (ServiceTitan 2026), that's $26,600 in revenue from one month of after-hours pickup that would otherwise have hit voicemail.

How to evaluate a dispatcher before deploying

If you're shopping AI dispatch for an HVAC shop, the right way to evaluate is a 14-day pilot on a single phone line, with three measurements:

1. Pickup rate by hour-of-day. Should be 100% across all hours after deployment. Anything else is a configuration bug.
2. Emergency dispatch latency. Time from "this is an emergency" to "tech has been paged." Should be under 45 seconds 95% of the time.
3. Booking accuracy. % of bookings that landed correctly in the dispatch software with the right slot, address and customer info. Should be above 96%. Below that and you're paying a CSR to clean up after the AI.

These three numbers tell you whether the system is working. Anything else — voice quality, "naturalness", brand fit — is secondary.

FAQ

Does this replace the dispatcher?

Usually not. The standard 2026 pattern is: AI handles 100% of after-hours and lunch-hour pickup, plus overflow during the day. Your dispatcher handles complex daytime calls and the AI escalations. The combined cost is lower than two dispatchers and the coverage is meaningfully better.

What happens if the AI gets the emergency wrong?

Two safety nets. First, the on-call tech callback within 15 minutes catches misjudgments — if the caller insists this is urgent and the AI dispatched as routine, the tech rebalances on the callback. Second, all calls are recorded and reviewed daily for the first 30 days of deployment; the prompt is tuned based on what's missed.

Does it work for emergencies outside of weather extremes?

Yes. Gas smells, water damage, CO alarms, frozen pipes, electrical smoke, no-heat in cold-weather markets — all standard emergencies in the triage taxonomy.

Can it handle Spanish-speaking callers?

Modern voice models (Hi Agent uses ElevenLabs Conversational AI with auto language detection) handle English, Spanish, French and Mandarin out of the box. Accuracy is highest in English and Spanish for US contractor deployments.

What's the latency for the actual page to the tech?

SMS pages typically deliver in 4-8 seconds. Voice pages place the call within 15-25 seconds of the dispatch decision. The end-to-end "caller hangs up" to "tech's phone rings" is consistently under 45 seconds in production.