How to Build an AI Dispatcher with Claude for Your HVAC Company
Missing 27% of your calls costs the average home service shop $45,000 to $120,000 a year. Here is a realistic build plan for an AI dispatcher on Anthropic's Claude, the shortcuts that work, and the production pitfalls nobody warns you about.
Key takeaways
- Home service shops miss 27% of inbound calls per Invoca, and 85% of callers who hit voicemail never call back.
- Even a 5% improvement in booking rate lifts an average HVAC shop by roughly $100,000 in annual revenue per ServiceTitan.
- Claude Haiku 4.5 at $1 per million input tokens keeps raw model costs below $300 a month for most $2-5M shops.
Home service shops miss 27% of inbound calls, per Invoca's research. Each missed call costs about $1,200 in foregone revenue.
The kicker: 85% of those callers never try you again. They dial the next shop on the Google results page.
This post walks through what it actually takes to build an AI dispatcher on Anthropic's Claude, where the hard parts are, and where you are better off buying instead of building.
What an AI dispatcher actually does
An AI dispatcher is not a phone tree. It is a system that answers a call, holds a conversation, gathers the information a human CSR would, and writes a job into your CRM.
A minimum viable build needs to:
- Answer the phone within two rings
- Ask qualifying questions (equipment type, symptom, warranty, address, preferred window)
- Pull the customer's history if they already exist in your CRM
- Check real availability against your dispatch board
- Book the job or warm-transfer to a human
- Write every interaction back to the CRM with a recording and transcript
ServiceTitan reported in their 2025 Residential Services Report that a 5% booking rate improvement adds roughly $100,000 in annual revenue for a typical HVAC shop (ServiceTitan). That is the prize.
Why Claude specifically
Claude is built by Anthropic, the company founded by former OpenAI research leads. The model family (Opus, Sonnet, Haiku) covers a range of cost and latency points.
For a dispatcher, the most useful traits are:
- 200K token context window, per Tech Insider's 2026 comparison. That is enough to hold a customer's full job history in a single prompt.
- Lower hallucination rate on document tasks, which matters when the model is reading work orders and extracting part numbers.
- Native tool use, documented in Anthropic's tool use docs. Your dispatcher does not need a separate orchestration framework to call ServiceTitan or Jobber.
Claude Opus 4.7 added self-verification of outputs before reporting back, per MarkTechPost. For an AI that is booking real appointments on your calendar, that check saves you from the "ChatGPT said it was the capacitor" problem profiled by ACHR News.
The build plan
Here is the realistic architecture for a Claude-based dispatcher.
1. Voice layer
Claude is a text model. You need a voice pipeline in front of it.
The usual stack is Twilio for the phone number, a speech-to-text service (Deepgram or Whisper), Claude for reasoning, and a text-to-speech service (ElevenLabs) for the reply. You can also use an all-in-one like Vapi, Retell, or LiveKit.
Rough monthly cost for 500 calls: $150-$400 in voice infrastructure before you add the model.
2. The Claude layer
You send the call transcript to Claude with:
- A system prompt that defines the dispatcher's role, tone, and boundaries
- A tool definition for each CRM action (get_customer, check_availability, book_job, transfer_to_human)
- The live conversation history
Claude returns either a response to speak back or a tool call. Your code executes the tool, returns the result, and loops.
Anthropic's tool use docs cover the mechanics. For HVAC specifically, your tools need to handle nuances the docs do not cover: emergency routing, warranty lookups, maintenance plan members getting priority.
3. CRM integration
This is where most builds stall. Jobber, Housecall Pro, ServiceTitan, and Workiz all expose APIs, but the rate limits and webhook reliability vary.
You need:
- Real-time availability (not yesterday's calendar cache)
- Two-way sync so bookings show up for your techs within 30 seconds
- Conflict handling when the dispatcher books a slot a human just grabbed
Jack Carr of Owned and Operated and Rapid Response Heating & Cooling has been public about his shop's AI experiments. He has emphasized that AI in dispatch works best when it assists rather than replaces, coaching human dispatchers from when calls come in and suggesting better assignments. That pattern (AI plus human) is the safer path for a $2M-$5M shop.
4. Model cost
Anthropic's pricing page lists Claude Haiku 4.5 at $1 per million input tokens and $5 per million output tokens. A typical 4-minute dispatcher call is about 2,000 input tokens and 500 output tokens. That works out to roughly $0.005 per call, or $2.50 per 500 calls.
Even at Sonnet 4.6 pricing ($3 in, $15 out), 500 calls costs under $15 in raw model cost. The money is in the voice infrastructure and engineering time, not the model.
5. Prompt caching
Anthropic supports prompt caching, which cuts cost dramatically on the repeated system prompt portion of every call. Without caching, you are resending your 3,000-token dispatcher prompt on every turn. With caching, those tokens are 10x cheaper on cache hits.
Turn prompt caching on before you go live. It is the difference between a $400/month bill and a $40/month bill at 500 calls.
The hard parts nobody talks about
Compliance
If the dispatcher takes credit cards over the phone, you are in PCI scope. Most shops route payment to a human.
If the dispatcher sends follow-up SMS, you are in TCPA territory. You need explicit opt-in, STOP/START handling, and quiet-hours enforcement. The Hatch 2024 report found that 80% of sales take 5-12 touches but most shops follow up once or twice. Automating that gap is worth real money but the compliance layer is table stakes.
Evals
Every prompt change is a risk. You need a golden set of 50-100 historical calls with the correct booking outcome labeled. Replay them through your dispatcher after every change.
Anthropic's pricing includes tooling for this but you are writing the evals yourself.
Human fallback
Real calls go off script. A customer speaks Spanish. A line is bad. A dog is barking. Your dispatcher needs a warm-transfer path to a human with full context handed over.
One contractor quoted in AI receptionist coverage said it plainly: "I didn't even know I was missing that many calls until I saw the data. I just thought business was slow." The same contractor learned that when AI takes the call, about 15-20% still need a human. Plan for it.
Latency
A 1.2-second response delay feels fine to you on a chatbot. On a phone call, it reads as awkward silence.
Aim for sub-800ms time-to-first-audio on every turn. That means Haiku for latency-sensitive turns and Sonnet only when you genuinely need more reasoning.
Build vs buy math
A senior engineer plus voice infrastructure, CRM integration, eval rig, and compliance layer runs 4-8 months of work and $30K-$80K in first-year costs. Ongoing maintenance adds $2K-$5K a month as models update and your CRM API shifts.
Contrast with a vertical product built for contractors. Avoca reports that one customer's booking rate went from 55% to 90% after switching, enough that they had to hire more techs to keep up, per the Owned and Operated interview.
If your time is worth more than $150/hr and your CRM is Jobber, Housecall Pro, ServiceTitan, Workiz, or GoHighLevel, buying beats building for most $1M-$10M shops.
Where Sully fits
Sully is a vertical AI platform built specifically for $1M-$10M home service contractors. It ships pre-built agents for missed-call follow-up, lead qualification, quote follow-up, AI chat trained on your company data, and a morning brief.
Sully runs on Claude under the hood. You get the Claude reliability, the tool use, and the large context window without writing a line of prompt code. The integrations (Jobber, Housecall Pro, ServiceTitan, Workiz, GoHighLevel, Gmail, Google Calendar, Slack, QuickBooks, HubSpot) are already built and maintained.
Claude is the engine. Sully is the truck you actually drive.
Recap
An AI dispatcher on Claude is technically buildable. The model cost is low, the tool use API is production-ready, and the 200K context window covers a full customer history.
The hard work is voice latency, CRM integration, compliance, evals, and the human fallback path. A realistic first-year total is $30K-$80K and 4-8 months of engineering time.
If that math does not fit your shop, buy the vertical product and let it run on Claude under the hood.
Sources: Invoca on missed calls, ServiceTitan HVAC statistics, ServiceTitan Residential Services Report 2025, Anthropic tool use docs, Anthropic pricing, MarkTechPost on Opus 4.7, Tech Insider Claude vs ChatGPT, Owned and Operated on AI, Owned and Operated on Avoca, Hatch State of the Home Improvement Industry 2024, ACHR News on ChatGPT diagnoses, RevSquared contractor AI receptionist stories.
See Sully in action
Sully is the pre-built AI for home service shops. Connect your CRM, email, and phone system in minutes and the agents run on your real data.