AI Voice Agent for an HVAC Company — 3-Month Results and Revenue Impact

here’s a case study from the field that shows what happens when you apply AI voice agents to boring real-world businesses instead of chasing the latest hype cycle.

background: an HVAC company in the DFW area. 8 trucks, 12 employees, roughly $2M a year. decent operation. but one problem kept coming up — the owner was spending hours every day on the phone answering the same 5 questions on repeat:

“what’s your rate for a new AC unit?”
“do you service [random suburb]?”
“how fast can someone come out?”
“AC is blowing warm air — what to do?”
“can a quote be sent?”

every single day, the exact same conversations, sometimes 30-40 times a day. a receptionist quit because she was bored out of her mind. her words: “she said she felt like a broken record.”

a developer in the family offered to fix it with an AI agent. the owner laughed at the idea. the developer built it anyway.

the build (3 weekends, ~30 hours total)

full disclosure: a developer built this, so there was a technical advantage. but most of the work was just wiring existing APIs together. nothing built from scratch.

Vapi for the voice agent — handles telephony, speech-to-text, and TTS. $0.12/minute inbound, $0.07/minute outbound. totally worth it for how much it handles out of the box.

a custom knowledge base was written with all the company’s pricing, service areas, common troubleshooting steps, and scheduling policies. fed to the agent as the context prompt. the trick: structure it as Q&A pairs, not paragraphs. the agent retrieves faster and hallucinates less.

hooked up to the company’s Google Calendar through a simple webhook. when someone asks “can you come Thursday at 2pm?”, the agent books the slot directly. no human needed.

a fallback was added: if the customer asks something the agent can’t handle (like a complex technical question about a specific unit model), it transfers to the owner’s cell with full conversation transcript. happens maybe twice a day. the owner is fine with that because the other 30-40 calls are handled completely by the agent.

cost per call: about $0.50-$0.80 depending on length. average call is 3-4 minutes. most are scheduling or pricing questions that take less than 2 minutes.

how it works in practice

customer calls the company’s main number. the agent picks up with “Hi, you’ve reached [company name], this is [agent name], how can [company] help you today?” — sounds completely natural. elevenlabs voice, not robot TTS. the first time the owner heard it the reaction was “wait, that’s the AI?”

agent handles the call, answers questions, books appointments if needed. if booking, it sends an automated confirmation text with the details. if it’s an emergency situation (AC completely dead in summer, which is a crisis in Texas), it flags as urgent and the owner gets an immediate notification.

at the end of the day, the agent generates a summary of all calls, categorized by type: new quotes requested, service booked, troubleshooting advice given, missed/transferred. the owner gets it as a text message at 8pm.

the numbers (3 months in)

calls handled by the agent: 1,247
calls transferred to human: 186 (about 15%)
appointments booked automatically: 342
quotes sent via text follow-up: 218

that’s 1,247 calls the business didn’t have to pay someone to answer. at $15/hour for a receptionist handling 30 calls per day, that’s roughly $150/week in saved labor. but the real win is the appointments. the agent books about 12 appointments per day automatically. the booking rate went up about 30% because the agent can handle calls 24/7 — people call at 9pm, get an appointment for next morning, and the agent books it. before, they’d call at 9pm, get voicemail, and maybe call back the next day. or more likely call a competitor who picked up.

estimated revenue impact: roughly $15-20K incremental per month from the extra bookings. the owner credits the agent with about 20% of that — the rest is just natural growth. but even 20% of $15K is $3K/month.

monthly cost to run the agent: ~$350 (vapi + twilio + the cheap server it runs on). so $3,000 incremental revenue on $350 cost. that’s a 8.5x return.

this case study has since been used to pitch to other local service businesses.

the pitch template that’s working

so far the pitch has been made to:

– a plumbing company in Austin. signed. $600/month setup + $300/month maintenance. agent went live last week.

– a pest control company in Houston. demo next week.

– a lawn care company in San Antonio. said “not right now” but asked to check back in 3 months (winter = slow season for them).

– an electrician in Dallas. interested but wants to see the plumbing company results first.

the pitch is stupid simple. call the business owner, say “hey, there’s a system that answers your phone for you, books appointments automatically, and runs 24/7. it costs less than a part-time employee. want to see it on a free trial?”

the free trial is key. set it up on a separate phone number and forward their after-hours calls to it for 2 weeks. they see the results before paying. nobody has said no after the trial.

lessons learned building ai agents for the real world

1. small business owners don’t care about AI. they care about “does the phone ring less” and “do bookings increase”. don’t sell the technology. sell the outcome. the pitch never mentions AI, agents, or automation. just “a system that answers your phone.”

2. the voice quality has to be indistinguishable from human. elevenlabs is good enough now. but the settings need tuning — the default voice sounds okay but not natural enough. about 10 hours was spent testing different voice combinations until ones were found that pass the “mom test” (the developer’s mom thought it was a real person).

3. fallback handling is the hardest part. knowing when the agent can’t handle something and passing it gracefully is crucial. nothing worse than an AI agent confidently saying “yes, help is available for that” and then giving wrong information. the agent was trained to say “transferring to a specialist” for anything outside its scope. clients appreciate the honesty.

4. every business is different. the HVAC agent knows pricing, service areas, and troubleshooting. the plumbing agent needs different knowledge — pipe materials, water heater specs, emergency protocols. can’t copy-paste. a knowledge base has to be built from scratch for each one. takes about 4-5 hours per business.

5. the most common failure mode is pricing questions. customers ask “how much for a new furnace?” and there are 15 variables that affect the answer. the safe approach: have the agent give a range (“between $2,500 and $4,500 depending on the unit size”) and schedule a call with a human for a precise quote. ranges are safe. specific numbers when you don’t have full info are dangerous.

6. the booking confirmation text is crucial. about 20% of people who book forget they booked. the automated text reminder saves so many no-shows. the company’s no-show rate dropped from about 12% to 4% after the agent started sending confirmations.

what’s next

a simple onboarding tool is being built that lets the business owner fill out a form (pricing, service areas, hours, common questions) and generates the agent configuration automatically. should cut the setup time from 4-5 hours to about 1 hour. that’s the difference between a consulting business and a scalable product.

also exploring inbound-only vs. outbound. inbound is easier because the customer already wants something. outbound (agent makes calls to warm leads) is harder but the unit economics are way better. the HVAC agent could call past customers for maintenance reminders. hasn’t been tried yet. might be the next experiment.

if you’re a dev looking for a real-world AI agent use case: go talk to local service businesses. plumbers, electricians, HVAC, pest control, lawn care. they all have the exact same problem — too many repetitive phone calls, not enough time. the solution is the same for all of them with minor tweaks. and they have money. they’ll pay for stuff that directly drives bookings.

forget building AI agents for crypto trading or content generation. go build one that answers “how much for a new water heater?” and you’ll have more customers than you can handle.