Skip to content
·6 min read

how a four-person roofing company stopped writing quotes at 10pm

aitradescontractorsquotingautomationottawaworked examplesmall businessai adoption

this is an invented example. no client of ours is described here. we have written it as a full project (scope, build, timeline, cost, result) because "ai can help with quoting" is useless and a specific build you can argue with is not.

the research behind the time savings is on our ai adoption page, with the papers.

the situation

ridgeline roofing. four people. dave owns it and still goes up ladders. two crew. his sister-in-law does the books two days a week.

dave quotes everything himself. a typical tuesday: four site visits between 8am and 4pm, drive home, dinner, then from about 8:30pm he sits at the kitchen table with his phone, scrolling back through photos trying to remember which roof had the rotten fascia and which one had the awkward access.

each quote takes him about 25 minutes. forty quotes a month. that is 17 hours, every month, at the kitchen table after a full day on a roof.

the quotes go out the next evening at best. usually two days. he loses roughly one job in five to somebody who quoted faster, he knows this because customers tell him when they call to say thanks but no thanks.

what was actually built

three things. the first one is not ai and it is the one that matters most.

1. the rate card, written down properly (week one)

dave's pricing lived in three places: a spreadsheet from 2023, a laminated card in the truck, and his head. the head version was the accurate one.

we sat with him for half a day and wrote it out as structured data:

  • labour rates by crew size and job type
  • material costs per square, by product line, with his actual supplier pricing
  • standard assemblies, "tear-off and replace, asphalt, single storey, simple access" as one priced unit
  • the modifiers he applies without thinking: steep pitch, two-storey, bad access, disposal, winter

this is 60% of the project and it contains no ai whatsoever. it is also the deliverable he would keep if he fired us in week two.

2. capture on site (week two)

a form on dave's phone. at the end of each visit, standing in the driveway, he:

  • takes photos, as he already does
  • taps the job type from a list
  • enters measurements into four fields
  • holds the mic button and talks for thirty seconds about anything unusual

that is 90 seconds, on site, while he is still looking at the roof. not from memory at 9pm.

3. the draft (week three)

overnight, each captured job becomes a draft quote:

  • the voice note is transcribed
  • the job type and measurements select line items from dave's rate card
  • the photos attach to the right sections
  • the model writes the covering paragraphs in dave's own phrasing, taken from forty of his past quotes
  • anything the rate card does not cover is left blank and flagged in red at the top

dave opens his laptop at 7am with four drafts waiting.

how long it took

weekwhat happened
1half a day on site with dave. rate card written down and checked against his last 20 invoices.
2capture form built and tested on three real visits. dave broke it twice; we fixed it.
3drafting built against his past quotes. first drafts reviewed side by side with how he would have written them.
4live. dave sends real quotes from drafts, we watch what he edits.
5–6tuning. the model was too formal and kept writing "we would be pleased to"; dave does not talk like that.

live in four weeks. useful by week six.

what changed

dave's tuesday, after:

four site visits, 90 seconds of capture at each. drives home. has dinner. does not open the laptop.

wednesday 7am, coffee, four drafts. he reads each one, fixes the bits that are wrong, sends. eight minutes each.

beforeafter
minutes per quote258
hours a month175
quote goes out1–2 days latersame morning
written at10pm, kitchen table7am, with coffee

12 hours a month back, and quotes land before the competition's.

the hours are not really the win. the win is that dave stopped doing unpaid admin after a nine-hour day, and that his quote is now usually the first one the customer sees.

what it would cost

honest numbers for a build this size:

  • the project: $6,000–9,000, most of it week one and three
  • running it: roughly $40–70 a month in model and hosting costs at forty quotes
  • payback: if one extra job a month closes at an average roofing ticket, this pays for itself in the first month and keeps paying

in ontario, the planning portion may be eligible for a matched grant through the digital competence centre, which changes the arithmetic further.

the one rule that makes it safe

prices come from dave's rate card. the model never invents a number.

a language model asked to price a roof will produce a confident price. it will be wrong in both directions, and the direction that hurts is winning a job at a number that loses money.

so the model assembles and describes. it does not price. and dave approves every figure before it reaches a customer, which takes him about fifteen seconds and is the reason this is safe to run.

what we would not build

  • auto-send. the draft waits for dave. always.
  • pricing anything unusual. if the job has something the rate card does not cover, the draft stops and says so rather than guessing.
  • the follow-up call. the quote does not win the job, the call does, and no software fixes that.

when this is not worth doing

if you write five quotes a month, do not build this. the arithmetic needs volume, somewhere above about twenty quotes a month is where it starts paying, and at forty it is comfortable.

if your pricing genuinely lives in one clean spreadsheet already, week one gets much shorter and the project gets cheaper.


if you want this worked out with your real numbers rather than dave's invented ones, that is what a readiness review is: a written assessment, fixed fee, with the arithmetic done on your actual quote volume.

other worked examples: funder reporting in a charity · intake and referrals in a clinic