I built an AI that quotes construction jobs. Here's what it gets right — and where a human still wins.

Receipts first, opinions second.

A twenty-two-square reroof — materials and labor, itemized line by line — quoted at $20,994. A 50-gallon water heater swap at $3,376. An EV charger circuit at $2,775. Real quotes out of a live beta, each one built in minutes from a job typed in plain English. Not mockups, not a concept video. You can go run one yourself at gcquote.com.

I built the thing, so you should read this the way you’d read any builder talking about his own work: with a raised eyebrow. Fair. That’s exactly why this post spends as much time on where the AI loses as where it wins.

Where this started: my own shop’s quoting nights

I’m a metal fabricator in Chicago — twenty years on the floor. The first version of this engine was CreateQuote, built for my own trade because quoting was eating my nights: describe a fence, a gate, a staircase in plain English, get back a line-item quote with materials, cut list, build sequence, and labor at my real rate.

Then it turned out the bid is the bottleneck in every trade, not just mine. So the same engine became GC Quote — electrical, plumbing, HVAC, roofing, concrete, and on down the list. Thirteen trades, sixty job types, live in beta with self-serve signups open. That’s where the reroof number above came from.

What it gets right

It reads the job like an estimator, not a form. You type the job the way you’d text it to a buddy. It pulls the scope out of your words and shows you what it understood before it prices anything. No dropdowns pretending to be intelligence.

It asks the questions a seasoned estimator would. I once fed the fab version a sloppy fence description on purpose. It caught that a cantilever gate needs two heavy carriage posts plus a catch post — not standard fence posts. That’s a whole post the customer should pay for, and a miss I’d have eaten. Catching the expensive miss is worth more than the speed.

It knows a hinge isn’t a MIG spool. Hardware gets quoted whole — nobody needs 25% of a hinge. Consumables get quoted by what you’ll actually burn — maybe 90% of a spool of wire — and you can bump it to a full spool and watch the quote update. That’s the level of detail where most estimating software quietly gives up.

It prices at your numbers. Your hourly rate, your markup, your consumables — not a national average. The average doesn’t know what the market charges in your city, and it doesn’t know what you’re worth. Averages are how guys end up at $60 an hour while the shop across town bids $145 and gets it.

It hands you the paperwork, not just a number. One click and there’s a branded client proposal with a plain-English scope of work — and on the fab side, the shop sheet with the cut list, plus a materials PDF you can send straight to whoever’s buying the steel. The quote isn’t done when the math is done; it’s done when everyone on the job has their paper.

It’s fast where slow was costing you. The bid that used to take an evening takes minutes. In a small operation, that’s the difference between bidding five jobs this week or two — and between quoting at the kitchen table at 11pm or not.

Where a human still wins

Here’s the part the AI-hype crowd skips, and I won’t — I wrote the code, I know exactly where its blind spots live.

Site conditions. The AI has never walked your job. It doesn’t know what’s behind the drywall, that the wall is out of plumb, or that the “easy access” in the customer’s description is a gangway two feet wide. You walk the site; the machine doesn’t.

Permits. Sometimes it won’t ask about permit fees — or the four hours a GC burns pulling them. I know because that’s my time getting burned. When it doesn’t ask, I add it to the job description myself, and the quote picks it up.

Allowances and judgment calls. Some numbers are ballparks and need to stay ballparks until you’ve got eyes on the job. A tool that pretends otherwise is dangerous. This one tells you it won’t catch every question there is — I built it to say that, because a quoting tool you can’t trust about its own limits isn’t a tool, it’s a liability.

The smell test. After twenty years, a wrong number smells wrong before you can say why. No model has your scar tissue. The AI builds the draft; the last look before the customer sees it is yours, and it should stay yours.

The honest workflow

So no, this isn’t “AI replaces the estimator.” It’s simpler and more useful than that:

1. You describe the job in plain English, with everything you know about it.
2. The AI drafts the bid — scope, materials, labor, line items, in minutes, at your rates.
3. You judge it — walk the site, add the permits, set the allowances, run the smell test, bump the number to your margin.
4. It produces the paperwork — the client proposal and the shop-side documents — and you go build the thing.

The machine does the typing and the arithmetic and the remembering. You do the judgment. That split is why it works — and every “AI estimator” pitch that promises you won’t need the judgment part is BS. You will. That’s the job.