AI agents for HVAC: answer every after-hours call and book the no-heat jobs without waking the wrong technician
The money in HVAC service is lost at 7:40 p.m., when the office is closed, the caller has a cold house, and the next company on their list picks up. Here is what an agent can safely do about that — and the calls it must never handle alone.
An HVAC agent is most valuable as an after-hours intake and triage layer, not a dispatcher. It answers every call, reads a fixed safety script for gas and electrical concerns, books routine work into real availability, and escalates true emergencies to the on-call technician. Dispatch decisions, diagnosis, and pricing stay with people.
Key takeaways
- The first measurable win is capture: calls answered and jobs booked after 5 p.m., on weekends, and during the first cold snap.
- Triage must be scripted, not improvised. Gas, carbon monoxide, and electrical symptoms trigger a hard stop and a human transfer.
- Quote follow-up has the highest dollar value per conversation, because a replacement estimate is worth more than a service call.
- Success is measured in after-hours capture rate, estimate close rate, and callback rate — not in call volume.
Where the hours actually go
Walk into a four-truck residential shop during the first cold week of October and the phone pattern is obvious. Calls arrive in two spikes, roughly 7 to 10 a.m. and 4 to 7 p.m., with a Saturday morning shift that nobody staffs properly. A two-person office handles 60 to 120 calls a day in peak season. Hold time climbs past five minutes, and a meaningful share of callers hang up and dial the next result on the map.
Dispatch is the second sink. It looks like one decision — send a tech — but it is really five: which technician has the right license and skill, who is closest, who is on call tonight, which job is under warranty, and which customer has already been moved twice. Each change sends a ripple of texts to the customer and the technician. The dispatcher is not typing; they are reconciling.
The third sink is quiet and expensive. A replacement estimate goes out after a sales visit, and then nothing happens. The homeowner is comparing three bids and waiting for a financing answer. Most shops send one follow-up, maybe two, then move on. The install that goes dark is worth several thousand dollars, and the reason it went dark is almost never price — it is an unanswered question about rebates, warranty terms, or timing.
Four use cases that pay for themselves
1. After-hours intake and emergency triage
What the agent does: answers within two rings at 11 p.m., captures the address, system type, symptom, and whether anyone is home, then classifies the call into one of four buckets: no heat, no cooling, water leak, or a safety concern such as gas odor or a carbon monoxide alarm. Safety calls get a fixed script and an immediate transfer to the on-call technician per the published rotation. Routine calls get booked into the next real availability window.
What stays human: the technician diagnoses the equipment. The on-call manager decides whether to roll a truck at 2 a.m. The dispatcher can override any booking the next morning.
Metric to watch: after-hours capture rate — calls answered and either booked or escalated, divided by total after-hours calls. Track it against last year's voicemail count.
2. Dispatch support and schedule protection
What the agent does: confirms arrival windows, answers "where is my technician" with live status, reschedules when a job runs long, and protects blocks for warranty work. It can offer a slot only if the schedule genuinely holds it.
What stays human: the dispatcher owns exceptions, priority calls, and any change that would break a promise already made to a customer.
Metric to watch: completed jobs per technician per day. If the agent adds confirmations but not completions, it is generating activity rather than capacity.
3. Replacement quote follow-up
What the agent does: runs a four-touch sequence over roughly two weeks, answers financing, warranty, and rebate questions from an approved document set, offers a revisit, and records a reason when the homeowner says not now.
What stays human: the comfort advisor sets the price, handles negotiation, and closes. The agent never discounts and never improvises a coverage promise.
Metric to watch: estimate close rate on followed-up quotes versus quotes with no follow-up, plus days to decision. The second number is often the more revealing one.
4. Maintenance renewal and seasonal booking
What the agent does: contacts agreement holders before the season, books tune-ups, handles pause requests, and flags customers who have missed two consecutive visits.
What stays human: the office decides pricing changes and handles complaints about prior work.
Metric to watch: renewal rate and tune-ups booked per week during the shoulder season.
What not to automate
The boundary here is not technical. It is liability and safety, and it should be written into the agent's instructions rather than left to the model's judgment.
- Any gas, carbon monoxide, or electrical safety diagnosis. The agent may recognize keywords and escalate. It may not ask a homeowner to check a flame color, reset a breaker, or describe what "smells like rotten eggs" means.
- Combustion, refrigerant, and code advice. This is licensed technical judgment. An agent that sounds authoritative here creates real risk.
- Warranty coverage determinations. "That should be covered" is a promise the office may have to honor. The agent can read the policy and route to a human.
- Price exceptions and discounts. Negotiation is where margin disappears. Keep it human.
- Final dispatch decisions during an emergency. If the on-call tech is already on a job, the agent should escalate to the secondary contact, not invent a solution.
Guardrails and compliance
Safety triage is a script, not a conversation. Write the exact language for gas, CO, and electrical keywords, including the instruction to leave the building and call the utility or 911. Test it with adversarial phrasing: "my wife smells gas but says it's probably the stove." The agent should not debate.
Escalation needs a clock and a backup. Define a maximum time before a human acknowledges an emergency transfer — five minutes is a common target — and a second contact if the first does not answer. An escalation path that ends in a voicemail is not an escalation path.
Call recording disclosure. Several states require two-party consent for recording. The agent should disclose recording at the start of every call, in the same sentence as the greeting, and the vendor should let you change that line without a support ticket.
Text messaging is regulated. Appointment confirmations and follow-ups sent by SMS fall under TCPA consent rules. Collect prior express consent at booking, honor STOP immediately, and keep a record of when consent was given.
One source of truth for the on-call rotation. The agent must read the schedule the office actually uses. If the rotation lives in a spreadsheet that is updated by hand, the agent will eventually call the wrong technician at the worst moment. Fix the data before you buy the software.
What it costs
| Line item | Range | Notes |
|---|---|---|
| Voice agent software | $150–$600 / month | Per-location plans; higher tiers add concurrent calls and analytics |
| Speech and telephony usage | $0.10–$0.45 / minute | The dominant variable cost; a 3-minute average call is the planning assumption |
| SMS follow-up | $20–$80 / month | Carrier fees plus platform markup; volume-based |
| Implementation | $1,500–$8,000 one-time | Guided template at the low end; dispatch or CRM integration at the high end |
| Likely payback | 1–4 months | One recovered replacement quote or 8–15 recovered service calls usually covers a year of software |
Treat the subscription as the small number and the integration as the real decision. An agent that books into a calendar but cannot write a job into the field service system will save less than the estimate suggests, because someone still re-enters the work order. The cost guide breaks out the hidden lines, including per-seat fees and overage tiers.
A 30-day rollout
Week 1 — baseline and scripts. Pull last year's call log. Count after-hours calls, missed calls, and the share that never called back. Write the safety script with your service manager, not with the vendor. Decide the escalation clock and the backup contact.
Week 2 — build and shadow. Configure the agent to answer after hours only. Run it in listen-only mode on live calls or with a small test number. Compare its transcripts against what your best CSR would have done. Fix the top three mismatches before going live.
Week 3 — controlled live. Turn on after-hours answering with a transfer-to-human button that works. Review every transcript daily. Track the capture rate and any call where the agent promised something it could not deliver.
Week 4 — expand and review. If capture is solid and no safety call was mishandled, add Saturday and overflow. Start the quote follow-up sequence on estimates older than seven days. Set the weekly review meeting that will continue into month two.
The three metrics that prove it worked
- After-hours capture rate. Answered and booked or escalated, divided by total after-hours calls. Compare against the same weeks last year. This is the cleanest before-and-after in the whole project.
- Estimate close rate on followed-up quotes. Segment quotes that received the full sequence from those that did not. If the gap is small, the follow-up is polite but not persuasive — rewrite the second and third touches.
- Callback rate within 30 days. The share of completed jobs that generate a return visit. A rise here means the agent booked work the technician could not finish, or misclassified the symptom. It is the quality check that keeps the capture number honest.
Frequently asked questions
Can an AI agent handle gas leak or carbon monoxide calls?
It should not diagnose them. A voice agent can answer, read a fixed safety script that tells the caller to leave the building and call the gas utility or 911, then transfer to the on-call technician immediately. The agent's job is a fast, consistent handoff, not an assessment of the risk.
Will an AI receptionist work with ServiceTitan or Housecall Pro?
Most voice platforms connect through an API, a booking link, or a middleware layer, but the depth varies. Ask whether the agent can write a booked job into your field service system and read the real on-call rotation, not just push a web form. If it can only send an email, you have added a step rather than removed one.
How much does an AI agent cost for a small HVAC shop?
Expect roughly $150 to $600 per month in software for a voice agent, plus $0.10 to $0.45 per minute for speech and telephony usage. Implementation runs from about $1,500 for a guided template to $8,000 for a custom build with dispatch integration. A single recovered replacement quote often covers the first year.