Capstone Plumbing called us about a Tuesday-morning problem. A tech rolls up to a house in Cave Creek. The homeowner is the one with the recirc pump. Or is it? Was there a slab leak last August, or was that a different customer on the same street? The information exists somewhere. It is just not in front of the tech when it needs to be. That is the problem we built around, and this is what shipped.
Call recordings have a 60-day half-life
Capstone runs on ServiceTitan, like a lot of home services companies. ServiceTitan is good at the job in front of you: dispatch, invoicing, payment. But call recordings only stick around for about 60 days. After that, the audio of what the customer told you last time is gone. The notes on the record might mention a water heater, but not the conversation about why they held off on replacing it, or the offhand comment about the dog in the backyard that bites strangers.
Multiply that by years of history and you get a quiet, expensive problem: techs roll up to jobs without context that exists somewhere, but has been deleted on a rolling window. Off-the-shelf transcription doesn’t fix it. Generic models don’t know what a PRV is, can’t spell “recirc pump,” and mangle anything that sounds like trade vocabulary.
Meet the team where they already work
We built two production integrations for Capstone. Both run on top of ServiceTitan. Neither asks a tech to log into a new app, learn a new UI, or remember another password.
The first is a forever-archive of customer calls. Every call is pulled out of ServiceTitan automatically and transcribed locally with a Whisper distil-medium.en model fine-tuned for plumbing vocabulary. Transcripts and metadata land in SQLite with FTS5, so phrase search is instant across the full archive. On day one we backfilled the whole 60-day retention window in a single pass (about 1,150 recordings), and the cron job picks up new ones every 15 minutes. Search “slab leak” or “PRV replacement” and you get every call where those words came up, no matter how long ago.
The second is an AI summary that lives where techs already look: pinned as a CRM Note at the top of every customer record inside ServiceTitan. Four sections per customer — who they are, job history, equipment to watch, site-access notes (gates, dogs, parking, codes). No new tab. No new app. The summary is on the screen the tech was going to open anyway.
The cheapest, most reliable AI deployment is one your team doesn’t have to remember to use. Adoption stops being a training problem when the value shows up inside a workflow they already had.
Only do the expensive work when something changed
Generating an LLM summary for every Capstone customer every 15 minutes is technically possible and economically absurd. Capstone has thousands of customers. Most aren’t changing from one quarter-hour to the next. So the system is built in two tiers.
A cheap scanner sweeps ServiceTitan’s modifiedOn watermarks every 15 minutes and queues only the customers that have changed. A separate worker runs every 5 minutes, picks up the queue, and only invokes the LLM when the content fingerprint for that customer shifts. Most of the time, the system is quietly doing nothing, because nothing needed doing. That is the goal. Costs stay flat, summaries stay fresh, and the whole pipeline can run indefinitely without anyone watching a meter.
Get the dollar figures right, every time
The fastest way to lose a trades team is to put a wrong dollar figure in front of them right before they walk into a customer’s kitchen. So nothing financial gets quoted from a transcript. Source-confidence rules clean up the noisy plumbing vocabulary first. Any dollar amount that ends up in a summary is reconciled against the actual ServiceTitan invoice before it appears. The number on the screen matches the number on the invoice, every time, because it came from the invoice.
The membership upsell works the same way. Capstone’s Advantage Plan is $360/year dues, $89 dispatch waived, and 10% off services and installs. Instead of letting an AI guess whether a customer “would probably benefit,” deterministic math runs against the plan terms and the customer’s job history. It tells the tech the exact dollar amount they would have saved on recent work. The tech walks in with a data-backed pitch instead of a hunch.
Never overwrite what a human wrote
Every Foundation-written note carries a [CAPSTONE AI SUMMARY · …] marker. Refreshes patch only the marked sections. Notes a human wrote stay exactly as they were, every refresh, every time. This is the rule that makes the whole thing safe to leave running. A tech who jots “customer prefers afternoon appointments, works nights” at the bottom of a record can do so knowing the AI will never touch it. That trust is the asset. Without it, nobody leaves an automation enabled long enough for it to matter.
What we’d tell another home services team
If your business runs on ServiceTitan (or any field service platform with a real API), the most valuable AI you can deploy is not a chatbot. It is the patient kind that listens to every call, reads every note, watches every change, and surfaces the right context at the moment someone needs it.
Most of the work is not the model. It is the pipeline: idempotent jobs, fingerprint diffing, trade-aware vocabulary, invoice reconciliation, and the discipline of never touching what a human wrote. Done that way, the result doesn’t feel like “the AI thing.” It just feels like your business finally remembered everything you already knew.
Capstone Plumbing has been running this in production since May 2026. If you run a plumbing, HVAC, electrical, or other home services business on ServiceTitan and want to see what a dedicated AI agent could do for your team, read the full case study or start a conversation. We’d rather build you something useful than send you a slide deck.