AI agents for plumbers, HVAC and contractors: what to subscribe to and what to build

Why do home-service businesses lose money on missed calls?

A plumbing or HVAC call is rarely a browsing call. The caller has water on the floor, no heat, or a quote deadline, and if nobody answers they ring the next listing. Unlike a retail enquiry, a missed trade call is usually not deferred, it is lost, together with the ad spend that produced it. The structure of a trade shop makes this worse: the person who answers the phone is often the owner, a tech on a job, or one dispatcher who cannot take two calls at once, and demand peaks are weather-driven, so the busiest hour of the year is also the hour the phone is least likely to be picked up.

The answer rates are lower than most owners assume. In the Invoca Home Services Lead Conversion Benchmarks Report 2026 (July 2026), drawn from over 70 million calls across the Invoca customer base and covering nine home-services sub-industries including HVAC and plumbing, 52% of callers to home-services businesses reached a person; filtering out misdials and quick hang-ups by counting only calls over 15 seconds, the answer rate was 65%, and sub-industry answer rates ranged from 32% to 74%. Read the other way: roughly one call in three that lasted long enough to be a real enquiry was not answered by anyone.

That is the gap an inbound agent closes: it picks up the calls that currently ring out, at 7 pm, on Sunday, and when three calls arrive in the same minute.

What can an AI agent do for a plumbing or HVAC company?

The useful jobs are the ones that sit on the phone, the calendar and the invoice. Each of these is a separate agent or a separate skill inside one agent, and the first three carry most of the value.

• Inbound voice answering and booking: the agent greets the caller, captures name, callback number, service address and the problem, asks the qualifying questions your office asks (gas or electric, tenant or owner, under warranty), checks live availability and books the slot, then sends a confirmation text.
• After-hours and overflow: the same agent takes the line when the office is closed or every human line is busy, applies your emergency rules, and either books the next available slot or transfers to the on-call tech.
• Scheduling and dispatch: the agent reads technician skills, zones and the current calendar, proposes the slot that keeps drive time short, and writes the job into the field-service system with the notes the tech needs.
• Quoting: for flat-rate work, the agent reads the price book and gives the published range; for anything else it collects photos and details and creates a quote task for a person. It should never improvise a price.
• Estimate follow-up: an outbound text and call sequence that checks in on open estimates, answers simple questions, and books the install when the customer says yes.
• Review requests: after a job is closed as complete and paid, a text asks for a review and routes complaints to the owner instead of the public listing.
• Invoice reminders: polite, scheduled reminders on unpaid invoices with a payment link, with a hard stop before anything that could be read as a collections threat.

A large-shop example, with the caveat

The scheduling and dispatch side has the clearest published example, and it comes from a vendor. In Salesforce's customer story on The Grout Guy (Salesforce, Australia), a trades business that grew from 30 to 130 field technicians in three years reports that Agentforce Field Service saved more than 20 hours a week on work-order processing, cut work-order turnaround from up to two days to two minutes, lifted conversion from the low 30s to over 50%, and moved the dispatcher-to-technician ratio from one per 6 to 7 workers to one per 20 to 25. Treat that as what the platform can do at 130 technicians with the vendor writing the story, not as what a five-tech shop should expect.

How is an AI voice agent different from an IVR or an answering service?

An IVR routes; an answering service takes messages; an AI voice agent completes the task. The practical differences are whether the caller is booked by the end of the call, whether the job lands in your software with the right notes, and what happens when three calls arrive at once.

What does voice AI cost per call?

Voice platforms charge per minute of conversation, and the figures below are the vendors' own list prices in October 2026. On its pricing page, Retell AI lists voice infrastructure at $0.055 per minute, with the language model, text-to-speech and telephony billed on top (telephony at $0.015 per minute in most regions), an all-in range it states as roughly $0.07 to $0.31 per minute depending on the model and voice you choose, and phone numbers at $2 per month. On its pricing page, Bland AI lists a Start plan at $0.14 per minute with no platform fee and 10 concurrent calls, and a Build plan at $0.12 per minute plus $299 per month with 50 concurrent calls.

The worked example assumes 300 inbound calls a month averaging 3 minutes, which is 900 billable minutes. It covers platform usage only. It does not include an AI receptionist product's subscription if you buy one on top of the platform, the integration and setup work, or the human time for transfers and reviews, and the per-minute rate changes with the model and voice you pick. Run the same arithmetic on your own call log before you believe any number.

What the per-minute number leaves out

At 900 minutes a month the platform bill is small next to the value of a single recovered job, which is why the real decision is not the per-minute rate. It is the one-off cost of integration and testing, the monthly cost of whoever maintains the prompts and rules, and the subscription of any packaged product you put in front of the platform. Our AI agent cost guide for small business walks through those layers for a general business; the sections below apply them to a trade shop.

Subscribe or build, by shop size?

There are three routes. AI receptionist products aimed at trades and local services bundle the voice platform, a trained script and a basic calendar connection into a monthly subscription; you configure, you do not build. Voice platforms such as Retell, Bland or Vapi give you the raw agent and expect you, or someone you hire, to write the conversation design and wire it to Jobber, Housecall Pro or ServiceTitan. Custom agents are built on your field-service platform's own agent tooling or on a framework, with your dispatch logic, price book and escalation rules encoded in software you own. We have not verified the subscription prices of individual receptionist products for this guide, so check each vendor's page rather than a review site.

The honest middle option

Most shops between 5 and 15 technicians do not need a custom build, but they do need someone who can make a product or platform talk to their software properly. That is a few weeks of contract engineering, not a department. Our comparison of an AI automation agency, an in-house hire and a freelancer lays out the trade-offs; the broader AI agents for small business guide covers the same decision for other industries.

How do you connect it to Jobber, Housecall Pro or ServiceTitan?

Integration means four things, and a product that claims to integrate may do only the first. Calendar and job creation: the agent reads real availability (by technician, zone and job type) and creates the job in your software with the address, problem description, arrival window and notes, rather than emailing a summary for the office to re-key. Customer records: the agent matches the caller to an existing customer by phone number or address before creating a new one, so you do not end up with three records for the same house. Dispatch rules: the rules your dispatcher applies in their head (gas work only to licensed techs, no installs after 3 pm, tenants need landlord approval) are written down and enforced by the agent. Write-back and logging: every call leaves a recording, a transcript and a structured summary attached to the job, so the tech knows what the customer said and the owner can audit the agent.

Who does the work depends on the route. Receptionist products ship their own connectors, usually for the calendar and basic job creation; you configure them during onboarding. For a voice platform, an engineer builds the connection: the platform calls a webhook or function during the conversation, a small service checks availability and creates the job through the field-service platform's API, and the result is read back to the caller. The engineer also handles retries when the API is slow, what the agent says when the calendar is unreachable, and how a half-completed booking is recovered when the caller hangs up.

Scope the work before you hire for it. Write down the ten most common call types, the questions the office asks on each, the fields each one needs in your software, and the rules that decide who gets the job. That document is the specification for the agent and the test script for the pilot, and it takes an owner or office manager an afternoon. If you want the integration done by a vetted engineer rather than an agency, Hyparz matches pre-vetted AI agent developers and workflow automation specialists typically within days, with a one-week risk-free trial. See AI agent development services and AI integration services for the engagement shapes.

Can AI handle calls after hours and weekends, and when must it hand off to a human?

Yes, and after hours is where the agent earns its keep, because the alternative is voicemail. The rules need writing down first. Emergency rules: define what counts as an emergency for your trade (active leak, no heat below a temperature, gas smell, sewage backup) and what the agent does with each: book the first emergency slot, transfer to the on-call tech, or, for gas smell, tell the caller to leave the building and call the gas emergency line before anything else. Warm transfer: when the agent transfers, it should brief the human with the caller's name, address and problem rather than dropping a cold call, which is what the Retell test did for its 73 transferred calls. Escalation: anything involving a complaint, a price dispute, an injury, a request for the owner by name, or a caller who asks for a human twice goes to a person or, after hours, to a tracked callback within a stated window.

Two disclosure points belong in the rules. First, tell callers they are speaking with an AI assistant at the start of the call; callers do not mind an assistant that solves their problem, but they resent discovering it afterwards. Second, know the rules that apply where you operate. For businesses serving customers in the EU, the transparency obligations in Article 50 of the EU AI Act apply from 2 August 2026, including the duty to tell people clearly that they are interacting with an AI system unless it is obvious, as summarised for small businesses by Tair Khamitov on dev.to (July 2026). In the United States, call recording is governed state by state and some states require the consent of all parties on the call, which matters because voice agents record by default. Check both points with your counsel before go-live; this guide describes what the rules say and is not legal advice.

What goes wrong, and how do you test before going live?

The failures are predictable, and almost all of them show up in the first fifty recordings if you listen to them.

• Address capture: unit numbers, rural routes, and street names that sound like other street names. Fix: validate the address against a lookup service during the call, read it back, and ask for the nearest cross street when confidence is low.
• Double bookings: two agent sessions, or the agent and the office, booking the same slot. Fix: book through the field-service API with a hold, never from a cached copy of the calendar, and have the agent confirm after the write succeeds.
• Pricing promises: the agent quotes a number that is not in the price book, or quotes a diagnostic fee as the job price. Fix: the agent may only read prices from the book, must label them as starting prices where they are, and otherwise says a technician will confirm on site.
• Accents, noise and names: speech recognition degrades on a job site or in a car. Fix: confirm names by spelling, confirm phone numbers digit by digit, and set a transfer rule when the agent has asked the same question twice.
• Wrong trade, wrong zone: the agent books an electrical job for a plumbing crew or a job forty miles outside the service area. Fix: zones and services are hard rules checked before any slot is offered.
• Silent failure: the integration breaks and the agent keeps taking bookings that go nowhere. Fix: alert on every failed write and have the agent take a message and promise a callback when the system is down.

A two-week pilot plan

Week one: put the agent on after-hours and overflow only, keep the main line human, and have one named person listen to every recording each morning, logging each call as booked correctly, booked wrong, transferred, or lost. Fix rules daily. Week two: add the main line during the busiest two hours of the day, keep listening, and compare the agent's booking rate and call outcomes against the office's over the same period. Go live on the full line only when a week passes with no pricing errors, no double bookings and no wrong-zone jobs, and keep the daily listening at a smaller sample forever.

How much time and money can it return?

Build the number from your own call log rather than from a vendor case study, and show the assumptions. The two verified inputs above give a starting shape. In the Retell AI test (12 October 2026, vendor-run), 167 of roughly 240 calls, about 70%, ended in a booked appointment without a human and 73, about 30%, were transferred. In the Invoca 2026 home-services benchmark (July 2026), 65% of calls lasting over 15 seconds reached a person, so about 35% did not.

Worked example, with every assumption labelled. Assume your shop takes 300 real inbound calls a month and, like the Invoca average, a person answers 65% of them, leaving 105 unanswered. Assume, conservatively, that only a third of those unanswered callers would have become jobs if answered, which is 35 jobs. Multiply by your own average ticket; at an assumed $250 average that is $8,750 of monthly work the phone currently drops. Against that sits a platform bill of roughly $65 to $407 a month in the pricing table above, plus whatever you pay for a product subscription, plus the one-off integration. The return rests almost entirely on two numbers you can measure in a week, your true answer rate and your average ticket, not on anything a vendor claims.

On time, the saving is the office hours spent answering, re-keying and chasing. The Grout Guy figure of 20-plus hours a week on work-order processing is a 130-technician operation reported by its vendor; a five-tech shop should count its own hours instead, typically the owner's evenings and the dispatcher's re-keying. Recompute this after the pilot with the real call log, and if the recovered jobs do not cover the platform, the subscription and a maintenance budget with room to spare, stop. Hyparz is a remote-talent marketplace that matches companies with pre-vetted remote software and AI engineers, typically within days, with a one-week risk-free trial; if the numbers work and you need the integration built, AI automation services is the place to start.