A bad estimate costs you twice. Quote too low and you eat the difference — on a $9,000 bathroom, a 20% miss is $1,800 out of your pocket. Quote too slow and the customer signs with the guy who got back to them Tuesday instead of Friday.
So the pitch for AI is obvious: paste in your notes, get a clean scope of work in 90 seconds. That part mostly works. The part that doesn’t work is the part with dollars in it.
I ran the same three jobs through a general-purpose AI chatbot and through the AI features built into estimating software. Here’s what came back, what was useful, and what would have cost real money if I’d sent it.
The test
Three jobs, described exactly the way a contractor would scribble them in a truck:
- Bathroom remodel — “full gut, 5x8, new tub/surround, tile floor, vanity, toilet, exhaust fan, move no walls, house built 1961”
- Service panel upgrade — “100A fuse box to 200A, detached garage feeder, 1950s SoCal single story, meter on side wall”
- Deck repair — “replace 14 joists and all decking on 12x20 deck, keep framing posts, composite, stairs stay”
For each, I asked for a scope of work, a materials list, labor hours, and a price. Same prompt every time.
What AI got right
The scope language was genuinely good. Better than what most contractors write by hand at 9pm.
For the bathroom, it produced a 14-line scope that included demo and haul-off, moisture-resistant backer behind the surround, blocking for the grab bar, caulk/grout schedule, and a line about protecting adjacent flooring. It flagged that a 1961 house may contain asbestos in floor tile or mastic and said testing should happen before demo — that’s a real exclusion most people forget to write down and then argue about later.
The materials list was about 85% complete. It missed the small stuff — mixing buckets, shims, a replacement closet flange, thinset quantity math — but it got the big line items and it got the sequence right.
AI is good at structure and language. It is not good at prices.
Here’s roughly where the time went on the bathroom:
| Task | By hand | With AI | Saved |
|---|---|---|---|
| Write scope of work | 35 min | 8 min (incl. editing) | 27 min |
| Draft materials list | 25 min | 15 min (incl. fixing) | 10 min |
| Price it out | 30 min | 30 min (did it myself) | 0 |
| Format the proposal | 20 min | 5 min | 15 min |
| Total | 110 min | 58 min | 52 min |
Call it a 45% time cut on paperwork. Across eight estimates a month that’s close to seven hours back. That’s real.
Where it quietly got numbers wrong
Now the expensive part.
Labor hours came back low on every single job. The bathroom got 48 labor hours. A 5x8 full gut with tile, done right, with a one- or two-man crew and real-world dry times, runs closer to 70–90 hours in most markets. The panel upgrade came back at 8 hours — no mention of the utility disconnect/reconnect coordination, no permit inspection wait, no allowance for the feeder to the detached garage being undersized or the service entrance cable needing replacement.
Why does this happen? The model is averaging across everything ever written about bathroom remodels on the internet, including a lot of confident content written by people who have never swung a hammer. Smooth, ideal-conditions numbers.
Material prices were stale and regionally blind. It quoted composite decking at a per-square-foot number that was plausible two or three years ago, and nowhere near what my local yard charges this week. It had no idea whether I’m in a market with $19/hr helpers or $34/hr helpers.
It did not compute overhead at all. When I asked for a price, it multiplied hours by an hourly rate it invented and added materials with a round markup. There was no fully-loaded labor cost, no overhead per billable hour, no margin divisor. If you don’t already know the overhead math behind your hourly rate, AI will happily hand you a number that pays your wage and loses your business money.
It invented code requirements. On the panel upgrade it stated a specific working-clearance dimension and an AFCI requirement as if they were universal. Those things vary by code cycle and by what your AHJ has actually adopted. It sounded authoritative. That’s the dangerous part — it does not say “check with your inspector,” it says a number.
Here’s what the deck quote looked like side by side:
| Line | AI’s version | Actual |
|---|---|---|
| Labor hours | 24 | 38 |
| Decking material | $1,450 | $2,180 |
| Fasteners/hardware | $180 | $340 (hidden fasteners, joist tape, hangers) |
| Dump fees | not included | $220 |
| Permit | not mentioned | $165 |
| Overhead | not applied | $885 |
| Price | $3,900 | $6,450 |
Send the AI version and you lose roughly $2,500 and work two weeks for free.
The one-sentence rule
Let AI write the words. You write the numbers.
Everything on the left side of that line — scope narrative, exclusions, assumptions, the customer-facing description, the cover note, the follow-up email — AI handles well and fast. Everything on the right side — hours, rates, material costs, markup, margin, code requirements, permit fees — comes from your own numbers and your own supplier.
How to actually use it
Here’s the workflow that works, in order:
- Dictate your site notes into your phone right after the walkthrough, while you’re still in the driveway. Rough, ugly, complete. “Joists 16 OC, two rotted at the ledger, access is tight on the north side, homeowner wants gray composite, dog in the yard.”
- Paste those notes in and ask for a scope of work only. Explicitly tell it: no prices, no labor hours. You’re getting language, not math.
- Ask it for a list of what you might have missed. This is the single most useful prompt. “What does a contractor commonly forget to include when quoting this job?” It’ll surface dump fees, permits, trim transitions, paint touch-up, the closet flange.
- Price it yourself using your fully-loaded labor cost and overhead rate. If you haven’t built those yet, the pricing walkthrough has the arithmetic.
- Call your supplier for anything over $500 in a single material category. Current price, current lead time. Not the AI’s guess, not last year’s invoice.
- Have AI tighten the final proposal — tone, clarity, formatting. Then read every line yourself before it goes out.
Step 6 matters more than it sounds. The scope is the contract. If AI writes “replace damaged subfloor as needed” and you don’t catch it, you just agreed to an unlimited subfloor allowance. Vague language in a scope always gets interpreted in the customer’s favor when there’s an argument. Our guide to estimates that actually win jobs covers what belongs in the document regardless of who drafted it.
What about the AI features in estimating software?
Different animal, and generally better — because they’re pulling from your data.
The AI features showing up in job management platforms mostly do one of three things:
| Feature | What it does | Worth it? |
|---|---|---|
| Notes → scope draft | Turns dictation into a line-itemed scope | Yes. Biggest time saver. |
| Price book suggestions | Pulls your past line items and prices | Yes, if your price book is clean |
| Auto follow-up emails | Drafts and schedules quote follow-ups | Yes — follow-up is where quotes get won |
| “Suggested price” | Guesses a total | No. Turn it off or ignore it. |
The second one is the key distinction. When software suggests $4,200 because that’s what you charged for the last three similar jobs, that’s useful — it’s your own history talking. When a chatbot suggests $4,200 because that’s the internet average, it’s noise.
If your price book is a mess, the AI inherits the mess. Garbage in, confident garbage out. Fix the price book first. If you’re still deciding on a platform, the software comparison walks through which tools have usable price books versus which ones just have a notes field.
Things to never hand off
Short list. Keep it somewhere you’ll see it.
- Code and permit requirements. Verify with your AHJ. Every time. AI will state a dimension or a requirement with total confidence and be wrong for your jurisdiction.
- Anything structural. If it holds the building up, that’s an engineer, not a chatbot.
- Change order pricing. This is where margin lives or dies. Do it yourself.
- Contract terms. Payment schedule, lien language, termination, warranty period. Get these reviewed by someone licensed in your state once, then reuse your own template.
- Final labor hours. Your crew’s pace is your number. Nobody else has it.
One more honest caveat
Everything above is based on a handful of jobs run through tools that update constantly. The specific weaknesses I hit — low hours, stale prices, invented code — may shrink over the next couple of years. The structural problem probably won’t: a model that doesn’t know your overhead rate, your supplier’s Tuesday price, or how fast your helper tiles cannot produce a number you should sign your name to.
Test it yourself before you trust it. Take an estimate you already completed and know the actuals on. Feed it the same notes. Compare line by line. You’ll learn more in 20 minutes than from any article about AI, including this one.
Your next move
Pick your last three completed jobs — ones where you know the real hours and real material cost. Run the notes through whatever AI tool you have access to and compare the output to what actually happened.
If the hours come back more than 15% low, you now know exactly how much to adjust every future AI draft before you look at it. That single correction factor is the whole trick.