Forty-three minutes. That is how much of an average residential service call, in the shop modeled below, goes to something other than turning a wrench. That number — not a feature list, not a demo — is what should decide which ai tools for field technicians you actually put on a truck. Every minute in that 43 belongs to one of four parts of the tech's day: drive, diagnose, document, collect. Anything that does not remove minutes from one of those four is another icon your crew will ignore by week three.
So let's work one shop all the way through, with the arithmetic showing.
The shop we are modeling
This is an illustration built from typical numbers, not a named client. Six trucks, mixed HVAC and plumbing residential service, running on a mainstream field service platform. Assume:
- 6 technicians, 250 working days per year
- 4 completed calls per tech per day — 6,000 calls a year
- Average ticket around $420, roughly $2.5M in annual revenue
- 9-hour paid day, with a target of 6 billable hours
- Fully burdened technician cost — base wage plus payroll tax, workers' comp, benefits, phone, and vehicle allocation, which in most trades shops lands roughly a third above the base wage itself
Hold those five numbers. Every claim below runs through them. Note that the savings math in this article is deliberately expressed in hours rather than dollars, because the burdened cost of an hour varies enormously by market and by trade. Convert to dollars using your own payroll register, not mine.
Baseline: where the 43 minutes go
I sat with the shape of a residential call and split it into the six administrative blocks that exist on nearly every one. These are minute counts for the non-productive portion only — not the actual repair.
| Step | What the tech is doing | Minutes per call |
|---|---|---|
| Pre-arrival | Reading dispatch notes, prior job history, checking the address and gate/dog notes | 4 |
| Equipment ID | Finding the data plate, transcribing model and serial, pulling specs, charge, and wiring | 7 |
| Parts check | Calling the office or the supply house to confirm a part is on the shelf | 6 |
| Documentation | Typing findings, uploading photos, recording refrigerant added or recovered | 11 |
| Options build | Assembling repair-versus-replace options into something presentable | 9 |
| Closeout | Invoice, signature, payment attempt, notes to the office | 6 |
| Total | 43 |
Four calls a day means 172 minutes — 2 hours 52 minutes — of a nine-hour day spent not fixing anything. The 6-billable-hour target was never realistic. The tech was running at about 5.2.
Here is the unit of measure worth memorizing: in this shop, one minute saved per call equals 100 technician-hours per year (1 min × 4 calls × 6 techs × 250 days ÷ 60). When a salesperson says a tool saves "a couple minutes," you now know whether that is 200 hours or noise. Run the same formula against your own truck count before any demo, and you will walk into the call with a number the vendor has to beat.
Drive: the four minutes before the truck stops
Routing is the wrong place to look first. Most six-truck shops are density-limited, not algorithm-limited, and I have written elsewhere about auditing scheduling and dispatch before buying anything. The drive-time win for the technician is smaller and more specific: the pre-arrival block.
A tech pulling up to a house has to reconstruct context from a dispatch note, a customer note, and possibly three prior visits. That is reading, and reading requires the truck to be stopped. An in-platform assistant that produces a spoken 30-second brief — last visit date, what was done, the equipment on file, the open balance, the gate code — turns four minutes of parked reading into listening on the drive.
The edge case worth naming: this only works if the prior visits were documented well enough to summarize. A summary of three thin notes is a thin summary. Fix documentation first, and the pre-arrival brief gets better on its own six months later.
Realistic recovery: 4 minutes to 1. Three minutes back.
One regulatory note while we're on drive time: most service vans stay inside FMCSA's 150 air-mile short-haul exception, which is why almost no residential trades shop needs an ELD. Track drive time through the GPS already in your field service platform. Buying a compliance product you are exempt from is a common and expensive mistake.
Diagnose: the data plate is the real bottleneck
This is the block where current AI is genuinely, unambiguously better than what the tech was doing. Multimodal models read a photo. A data plate is a photo.
The old sequence: crouch behind a condenser, hold a flashlight in your teeth, squint at a sun-faded plate, type a 17-character serial into a phone with gloves on, get it wrong, retype it, then search a manufacturer portal for the spec sheet. Seven minutes, and the transcription error rate on faded plates is not small.
The new sequence: photograph the plate. The model returns manufacturer, model, serial, manufacture date from the serial pattern, factory charge, and refrigerant type — and drops it into the equipment record. Then the tech asks a follow-up in plain language: "What's the subcooling target on this at 95 outdoor?"
Two things make this more than a convenience:
- Refrigerant recordkeeping. EPA Section 608 requires service records on appliances containing 50 pounds or more of refrigerant, with leak-repair triggers at 20% annual leak rate for comfort cooling and 30% for commercial refrigeration. A photographed, parsed, timestamped equipment record with the charge logged is a compliance artifact. A scribbled note on a work order is not.
- The A2L transition. Equipment manufactured on or after January 1, 2025 moved to A2L refrigerants under the AIM Act technology transitions rule. Your techs now roll up on R-410A and R-454B systems in the same week, with different gauges, different recovery equipment, and different leak detectors. Reading the plate correctly the first time is the difference between the right tools coming off the truck and a second trip.
Realistic recovery: 7 minutes to 2. Five minutes back.
The parts-check block is the honest counterexample. Confirming a capacitor or a flapper is on the shelf is an integration problem — your inventory record talking to the supply house — not an intelligence problem. Calling it AI is marketing. Real fix: get truck stock counts accurate and let the tech query them. 6 minutes to 4. Two minutes back, and none of it from a model.
Document: dictation only works with a template behind it
Eleven minutes of typing on a phone, usually done in the driveway after the job, sometimes done at 8 p.m. in the driveway at home. The write-up is thin, the photos are unlabeled, and the office calls the tech two days later asking what actually happened.
Voice dictation alone does not solve this. Raw transcription produces a wall of text that is worse than a short typed note. What works is dictation plus a structured template: the assistant takes 90 seconds of the tech talking through the call and returns a formatted record with findings, work performed, parts used, readings taken, refrigerant added or recovered, and a customer-facing summary written in plain English rather than trade shorthand.
That customer-facing summary is worth more than the time savings. It is what the homeowner forwards to their spouse, and it is the difference between an approved recommendation and "let me think about it."
Two failure modes to watch. First, noise: a running condenser fan or a compressor in a tight closet will wreck transcription accuracy, so the habit has to be dictating in the truck, not standing at the unit. Second, drift: models will occasionally smooth a tech's hedged language into something more definitive than what was said. Require the tech to read the generated record before it posts. Ten seconds of review protects you on a warranty dispute a year later.
Realistic recovery: 11 minutes to 5. Six minutes back.
Quote: build options at the equipment, not at the tailgate
Nine minutes assembling good-better-best options is nine minutes the customer spends unsupervised, deciding they'd like to get another bid. Pricebook-driven option building with an assistant that pulls the right task codes for the equipment on record compresses this hard — and consistency across six techs improves more than speed does. I covered the accuracy tradeoffs in detail in how contractors quote jobs faster without losing accuracy; the short version is that the model should assemble from your pricebook, never invent a number.
That constraint is not a nicety. A model allowed to estimate freely will produce a confident, plausible, wrong figure, and your tech will present it to a homeowner. Wire the tool so it can only select from task codes that already exist in your system, and it can only be as wrong as your pricebook already is.
Realistic recovery: 9 minutes to 5. Four minutes back.
Collect: the 34% that leaves without paying
In our model shop, 34% of residential tickets ended the visit unpaid — the tech left, the office invoiced, and the money arrived on average 19 days later. Annual write-off ran 2.1%.
Run that: 2.1% of $2.52M is $52,920 a year that was earned and never collected. That is more than most owners spend on their entire marketing budget, and it is invisible because it never shows up as a line item — it just quietly is not there.
The fix here is barely AI. It is a card reader in every truck, text-to-pay as the fallback, and an automated reminder sequence that stops the moment payment lands. The assistant's contribution is small and specific: it drafts the reminder in the customer's language and it knows when to stop. There is also a margin point people miss — a card keyed in over the phone by the office is a card-not-present transaction and carries higher interchange than a tapped card at the customer's door. Collecting on site is cheaper per dollar, not just faster.
Tightening this took unpaid-at-departure from 34% to 12% and write-off from 2.1% to 0.9%. Recovered: about $30,240 a year, plus roughly 13 days off the collection cycle on the affected portion. Closeout time also dropped from 6 minutes to 3. Three minutes back.
Add it up — then subtract what the crew ignored
Theoretical savings: 3 + 5 + 2 + 6 + 4 + 3 = 23 minutes per call. From 43 down to 20.
That is the number a vendor would put on a slide. It is not what happens.
In the first 90 days of a rollout like this, two of the six techs use everything, three use the dictation and the data-plate lookup only, and one keeps typing notes at 8 p.m. because that is how he has done it for eleven years. Weight it honestly and the shop captures roughly 11 minutes per call, not 23.
Now the math that matters:
- 11 min × 4 calls × 6 techs = 264 minutes per day = 4.4 hours
- × 250 days = 1,100 technician-hours per year
- A productive tech delivers roughly 1,500 billable hours a year
- 1,100 ÷ 1,500 = 0.73 of a full-time technician
Here is where owners get it wrong. Those 1,100 hours are capacity, not revenue. They only become revenue if you have demand to fill them. If your board is thin, the tools bought you a crew that goes home on time — which is worth something for retention, but it is not a return.
If your board is full and you were about to add a seventh truck, the arithmetic is different and much better: you deferred a hire, a vehicle, a wrap, a tablet, insurance, and a training ramp — while adding $30,240 in recovered collections. That is the real answer to whether AI is worth it for a small service business: it depends entirely on whether you have demand waiting on the capacity.
The rule that made it stick: no new icon
Every tool above lives inside something the tech already opens. That was not a preference. It was the condition.
Technicians open their field service app — Jobber, Housecall Pro, ServiceTitan, Workiz, FieldPulse — dozens of times a day because the schedule lives there. That app is the only surface with guaranteed adoption. Jobber Copilot, ServiceTitan's Sidekick, and Workiz's assistant all exist for exactly this reason: the highest-adoption AI is the one behind an icon that is already open. Before you buy anything standalone, check which AI tools actually connect to Jobber, Housecall Pro, and ServiceTitan, because a tool that requires a second login is a tool your crew will use for nine days.
The one exception in the list above is the data-plate lookup. That one earns its place on the phone's home screen because the win is large enough (five minutes a call, 500 hours a year across six trucks) to survive the friction of a separate app. Nothing else on this list clears that bar.
If you want to test any of this before committing, pick one step — I would pick documentation — and run it on two techs for three weeks. Time twenty calls before and twenty after with a stopwatch, not a survey. Pick your two most willing techs for the trial and your most reluctant tech for the second wave; if it survives him, it survives the shop. Then multiply by 100 hours per minute per year and decide with a real number in front of you.
That is the whole method, and it is the approach we take at Turnkey AI: count the minutes in the actual day first, then buy only what removes them.
Questions owners actually ask
Will my older techs use this?
Some will not, and you should budget for that in the arithmetic rather than pretending otherwise. In the model above, weighting for one full non-adopter and three partial adopters cut the projected savings from 23 minutes a call to 11. That is still 1,100 hours. Plan for 50% capture and be pleasantly surprised.
Can AI diagnose the equipment for my technician?
No, and you should be suspicious of anyone selling that. Current tools are excellent at retrieval — reading a data plate, pulling a spec, finding a subcooling target, surfacing a service bulletin. Diagnosis is pattern recognition against a physical system with sound, smell, and touch involved. The tool hands your tech the manual faster. Your tech still makes the call.
Does dictated documentation hold up for EPA recordkeeping?
The dictation is a capture method; the record is what matters. Section 608 recordkeeping for appliances with 50 pounds or more of refrigerant requires the servicing details, dates, and amounts added or recovered. A structured, timestamped entry in your field service platform with a photographed data plate attached is a stronger record than a handwritten work order. Confirm your platform stores refrigerant fields as structured data, not as free text buried in a note.
What about A2L systems — do the lookup tools know the difference?
The good ones read refrigerant type straight off the plate, which is exactly the information that determines whether your tech needs A2L-rated gauges, recovery equipment, and leak detection on that call. Verify this on your own equipment before trusting it. Photograph five plates across a mix of R-410A and R-454B units and check every field the tool returns against the plate.
Should I start with the technician's day or with the phones?
The phones, almost always. A missed call is a job that never enters the system at all, and no amount of technician efficiency recovers it — run the 12-point answering audit first. Fix intake, then come back to the four-block breakdown above once the board is full enough that recovered capacity turns into revenue instead of an early finish.
Frequently asked questions
Will my older technicians actually use AI tools?
Some will not, and you should build that into the math. In the shop modeled here, one full non-adopter and three partial adopters cut projected savings from 23 minutes per call to 11 — still about 1,100 technician-hours a year. Plan for roughly 50% capture.
Can AI diagnose equipment for my technician?
No. Current tools are strong at retrieval — reading a data plate, pulling a spec, finding a subcooling target, surfacing a service bulletin. Diagnosis involves sound, smell, and touch against a physical system. The tool hands your tech the manual faster; your tech still makes the call.
Does dictated documentation hold up for EPA recordkeeping?
Dictation is a capture method; the stored record is what counts. Section 608 recordkeeping for appliances with 50 pounds or more of refrigerant requires servicing details, dates, and amounts added or recovered. Confirm your platform stores refrigerant fields as structured data, not free text.
Do data-plate lookup tools handle A2L refrigerants correctly?
The good ones read refrigerant type straight off the plate, which determines whether the tech needs A2L-rated gauges, recovery equipment, and leak detection. Verify it yourself: photograph five plates across a mix of R-410A and R-454B units and check every returned field against the plate.
Should I start with the technician's day or with the phones?
The phones, almost always. A missed call is a job that never enters the system, and no amount of field efficiency recovers it. Fix intake first, then return to the drive, diagnose, document, and collect blocks once the board is full enough that recovered capacity becomes revenue.