Estimating, rebuilt around AI.

01 / The job

One number. A dozen sources.

An estimator reconciles drawings, specs, addenda, sub quotes, the last job’s numbers and their own experience into a single price.

And when the spec says “see drawings” and the drawings say “see spec”, someone has to chase it down.

02 / Under deadline

No time to read it all. No time to ask.

Addendum 4 lands two days before bid day, and the RFI cutoff has already passed. Somewhere in 1,186 spec pages, hardware set 4 calls for closers the door schedule leaves out.

So questions turn into assumptions, and assumptions into risk.

03 / The human limit

Even the best estimator is human.

Under deadline, judgment takes shortcuts in predictable ways:

• FatiguePage 900 doesn’t get the attention page 9 did. A set with a “ground floor” and a “first floor” loses one of them.
• AnchoringA number that looks like last time goes in without a second look.
• OptimismEstimates drift toward what the project hopes to spend.

None of it is carelessness. It’s what deadlines do to people.

04 / How we work

One finding, start to finish.

Specialist AI agents, one per trade, read the set and cross-check it. A cost engineer directs them, challenges what they find and decides what goes into the estimate.

1. 1Specialist agents read every sheet, one per trade. Nothing goes unread.
2. 2The spec and the door schedule disagree.
3. 3The openings agent flags it, with both sources.
4. 4The cost engineer challenges it and asks for a check.
5. 5The answer changes the quantity.
6. 6The estimate keeps the decision and its basis.

The agents do the reading. The engineer changes the number.

05 / What carries forward

The next project starts with this one’s lessons.

Every quantity, question and decision is logged with its source. This project added one check to every set we read after it: compare hardware sets against rated doors on the life-safety plan.