AI voice agent booking into field-service software: HVAC, plumbing, electrical

What it is

This is the core job every other home-services use case depends on: a voice agent that takes a booking request and turns it into a job on the dispatch board that a dispatcher can send to a truck without touching it. That means reading the things a human dispatcher reads (technicians, skills, zones, job types with durations, arrival windows, customer and equipment history, membership status) and writing the things a human dispatcher writes (a job with the right type, duration, window, address, notes and lead source, on the right customer record).

The word "integration" hides three very different products. Lead creation posts a name and a phone number into the system and someone calls back. Calendar sync sees busy and free blocks and creates an event with a title. Native write reads availability and writes a complete job. Only the third one changes how many people you need in the office. The rest of this page is about how to tell which one you are buying.

Small contractors buy this as an AI receptionist that books into their named platform. Groups specify dispatch-board write-back with field-level acceptance criteria.

Who buys it

• Owner-operators on Housecall Pro, Jobber or Service Fusion who already run the business from the app and want the phone to feed it directly.
• Mid-size HVAC and plumbing companies on ServiceTitan or FieldEdge with a dispatcher who spends the morning rebuilding jobs from messages.
• Multi-branch groups and franchises where zone logic, technician skills and membership tiers decide which window the agent may offer, and where a wrong booking costs a truck roll.
• Call-centre leads in larger groups who want containment on the highest-volume intent before they touch anything harder.

Budget owner: the owner or operations lead; the field-service system administrator owns the sandbox and the acceptance test.

KPIs

Reconcile every day for the first month: bookings the agent confirmed to a caller against jobs that exist on the board. Then track dispatcher corrections by field (type, window, zone, duration), duplicate customers created, booking containment, and latency on the turns that query the schedule.

Two traps. Containment only counts when "contained" means a job on the board with no dispatcher edit. And latency must be measured on tool-backed turns, because a platform that is fast on greetings and slow on availability checks will feel slow to every caller who actually books.

Demo script

Bring a sandbox of your own account, with real job types, two zones, three technicians and one who is on holiday. Then:

1. Greeting. Listen for the AI disclosure and the recording notice.
2. Returning customer. Call from a number that exists in the sandbox. Pass: it greets you by name after a light confirmation and knows the equipment on file. Fail: it asks you to spell your name.
3. Intent with interruption. Ask for a furnace tune-up, then cut in mid-offer with "actually the hot water is also lukewarm". Pass: it stops, adds the second problem, and asks whether it should be one visit or two. Fail: it books the tune-up and forgets the second item.
4. Constraint negotiation. Ask for the earliest morning window with the technician who came last time, not Friday. Pass: two or three real windows in one turn, honouring both constraints. Fail: a window with a different technician or on Friday.
5. Ambiguous time. Say "first thing next week". Pass: it names the date and the earliest window the board has. Fail: it books "8 a.m." when the earliest window is eight to twelve.
6. Address in the wrong zone. Give an address in the zone the on-holiday technician covers. Pass: it offers windows from the covering technician or says when the zone is next served. Fail: it books the technician who is away.
7. Phone read-back. Pass: digit by digit, then a yes.
8. Eight seconds of silence after an offer. Pass: a short prompt, then a hold or message option. Fail: hang-up or a full repeat.
9. Emergency injection. Say there is a smell of gas near the furnace. Pass: it stops booking, tells you to leave the house and call the gas emergency line, and alerts on-call. Fail: it keeps offering windows.
10. Write and verify. Accept a window and open the sandbox board. Check job type, duration, window, zone, technician, address with unit, notes and lead source. Then disconnect the sandbox and call again. Pass: it takes details and never says "booked".

Score each trap pass or fail. A vendor who wants to run the demo from their own audio or their own account has not passed the demo.

Compliance notes

Booking calls are inbound and customer-initiated, which places them outside the TCPA consent rule in the United States. Two things the agent does afterwards are not: an automated SMS confirmation and an automated arrival-window callback are outbound contacts that need prior express consent, and any call that could count as a solicitation sits inside the 8 a.m. to 9 p.m. window in the customer's local time. Announce recording at the start; roughly a dozen states require all-party consent. In the United Kingdom, a recording is processing of personal data under UK GDPR and the agent should say so; if the agent is later used to call the same customers with offers, that is an automated marketing call under PECR Regulation 19 and needs specific prior consent. In Australia, announce recording because state surveillance-devices laws vary, and keep any outbound follow-up inside the telemarketing hours of Monday to Friday 9 a.m. to 8 p.m. and Saturday 9 a.m. to 5 p.m. with Do Not Call Register checks. Field-service records hold names, addresses and equipment details, so the vendor's data-processing terms should cover every sub-processor in the audio path. Informational, not legal advice.

Build or buy

Buy if you run one location on a mainstream platform and the vendor can show a native write in your sandbox; the integration is the whole product and the platform's partner programme has done the hard part. Consider a platform or a build if you run custom zone or membership logic that no packaged product exposes, or if you need the agent to reason over equipment history before offering a window. In both cases keep the tool surface small: one call to read availability, one to write the job, with the rules living in the field-service system rather than in a long prompt.