AI receptionist for HVAC after-hours calls: intake and safety boundaries
A practical after-hours HVAC call flow that separates emergency direction, urgent human callbacks, and routine scheduling without asking AI to diagnose equipment.
Field note
By Vault Zero
An HVAC AI receptionist should handle after-hours calls in three lanes: direct life-safety concerns to emergency instructions, escalate business-defined urgent requests to a human, and capture routine service requests for follow-up. It should never diagnose a furnace, declare a property safe, or promise that a technician has been dispatched.
That boundary matters because carbon monoxide is not a routine repair symptom. The CDC describes CO as an odorless, colorless gas that can cause sudden illness and death, and lists headache, dizziness, weakness, nausea, chest pain, and confusion among common symptoms. A voice agent must not turn those facts into a remote diagnosis.
Define the three lanes before writing the prompt
| Lane | Example caller statement | Approved AI action | Prohibited AI action |
|---|---|---|---|
| Emergency direction | “The CO alarm is sounding,” “I smell gas,” or “There is smoke” | Stop routine intake and deliver the company's counsel-approved emergency instruction | Troubleshoot, diagnose, minimize risk, or keep the caller answering sales questions |
| Urgent human review | No heat during dangerous weather, a vulnerable occupant, active water near equipment, or another business-defined trigger | Capture only necessary details, attempt the approved transfer or urgent page, and state exactly what happens next | Guarantee response time or say a technician is on the way before dispatch confirms |
| Routine follow-up | Maintenance, estimate request, noisy unit without a safety trigger, thermostat question, or future replacement | Complete structured intake and offer only verified scheduling options | Invent availability, price, rebates, diagnosis, or warranty coverage |
The HVAC company—not the model—must decide which conditions enter each lane. A licensed professional and legal counsel should review the policy for the jurisdictions served.
Put emergency direction before normal intake
When a caller reports a CO alarm or possible exposure, the flow should interrupt immediately. The U.S. Consumer Product Safety Commission tells consumers whose CO alarm sounds to move outside to fresh air, call emergency services or 911, account for everyone, and not re-enter until emergency responders authorize it. Your exact wording should be approved, localized, and short.
A safe pattern is:
“I can't determine whether the property is safe. Please leave the area and contact 911 or your local emergency service now. Do not wait for an HVAC callback. If you are already in a safe place, I can record your name and callback number for the service team.”
This is a design example, not legal or medical advice. Do not bury it behind address collection, equipment questions, or lead qualification. Do not ask the caller to inspect, reset, relight, ventilate, or operate equipment.
Use a short, structured HVAC intake
For non-emergency calls, collect only information the next person can act on:
- caller name and confirmed callback number;
- service address, city, and ZIP code;
- whether the caller is a current customer;
- plain-language issue in the caller's own words;
- system type only if the caller already knows it;
- when the issue began;
- business-approved urgency triggers;
- preferred appointment window or callback timing; and
- access notes that are appropriate to store.
Keep reported_issue separate from agent_summary. Preserve phrases such as “alarm is chirping” or “smells unusual” rather than converting them into a confident technical conclusion.
A sample after-hours conversation
Receptionist: “Thanks for calling Northside Heating. I'm the automated receptionist helping with after-hours intake. Are you calling about smoke, a gas smell, a carbon-monoxide alarm, or someone who feels ill?”
Caller: “No. The heat stopped about an hour ago.”
Receptionist: “I can take the details for the on-call team. I can't diagnose the system or guarantee a dispatch time. What is the service address?”
Caller: “14 Oak Street in Barrington.”
Receptionist: “Thank you. Is anyone at the property especially vulnerable to loss of heat, and what is the best callback number?”
The opening identifies the automation and checks the safety gate. The second response explains the boundary. The final question applies a business-approved escalation factor without offering technical advice.
Make transfer outcomes explicit
An attempted transfer is not the same as a completed handoff. Define what the caller hears when:
- the on-call person answers;
- the destination is busy or does not answer;
- the transfer drops;
- the escalation service is unavailable; or
- the caller declines to wait.
The fallback should confirm the next action in plain language: “I recorded this for an urgent callback” is different from “you are scheduled.” If the workflow uses Vapi, its official documentation supports predefined transfer destinations and caller-facing transfer messages, but configuration still needs end-to-end test calls.
Do not let the calendar create false promises
After-hours scheduling should use live, read-before-write availability with an explicit timezone and service area. If the calendar cannot be reached, the receptionist should capture a preference and say that the office must confirm it.
Use three separate states in downstream data:
| State | Meaning |
|---|---|
requested | Caller stated a preferred time; no availability check succeeded |
held | A temporary slot is reserved but requires another step |
confirmed | The system of record accepted the appointment and the business's rules permit confirmation |
Never turn requested into “booked” in the spoken response, text message, dashboard, or follow-up email.
Review privacy and outbound rules separately
Decide what is recorded or transcribed, how disclosure and consent work, where data is stored, who can access it, and when it is deleted. Recording and transcription requirements can vary by jurisdiction; have counsel review the operating locations and call flow.
Inbound intake also does not create blanket permission for future automated marketing. The FCC has confirmed that calls using AI-generated human voices fall within the TCPA's artificial or prerecorded voice rules for covered outbound calls, including applicable consent and identification requirements. See the FCC's declaratory ruling and obtain legal advice for the actual outreach use case.
Test the exceptions, not just the happy path
Before launch, place test calls for:
- CO alarm, gas smell, smoke, and a caller reporting physical symptoms;
- ambiguous phrases such as “the detector is beeping”;
- no heat with and without an approved vulnerability trigger;
- a routine maintenance request;
- unavailable transfer destinations;
- a calendar timeout and a double-booked slot;
- background noise, a weak connection, accented speech, and caller corrections; and
- a caller who refuses recording or asks for a person.
Use the full voice-agent failure-mode test plan, then review the HVAC front-desk system to align the website promise and phone path.
Put one call path under pressure before changing the whole phone system.
Tell me where calls get missed, what a useful handoff contains, and which situations still need a person. I'll map the narrowest pilot that can answer those questions honestly.
