AI Cut $1.2 Million From a Construction Bid. She Put It Back.

The total at the bottom of the cost plan was $18,742,600.

The estimator had spent three years building totals like that from drawings, subcontractor prices and notes scattered across hundreds of pages. This one had been produced in two days by a system that read construction documents, counted materials and matched the results with prices from earlier projects.

She had checked enough of its work to know it could be useful. On one recent job, it found concrete curbs she had missed during her first pass. It could count doors without losing its place and revise a total when the architect issued a new drawing set. Work that once kept her occupied for more than a week could reach a review meeting in three days.

That speed had become part of her value to the company. She was in her late 20s and being considered for a senior estimator role, a promotion that would put her in front of clients instead of leaving her to prepare the numbers behind the meeting.

Executives had already shared the $18,742,600 total with the owner of a planned office renovation. The number was below the owner’s budget and below an earlier estimate from another builder. By the time the cost plan reached her for final review, the lower figure had been described inside the company as a win.

She printed the plan and began marking it with a pencil.

A complete-looking number

The system did more than transfer figures into a spreadsheet. It used computer vision to identify walls, doors and structural members in drawing files. It read specification text, assigned the detected work to cost categories, then generated quantities and suggested unit prices from the company’s past estimates.

That combination mattered. A spreadsheet can add the numbers it receives, but this tool decided which numbers should exist. It inferred that a drawing symbol represented a door, that a note belonged to structural steel, or that a marked area required demolition rather than new work. The finished cost plan arrived with itemized rows and confidence indicators, which made it look as if every part of the building had been considered.

The estimator started with structural steel because the framing total seemed low. The tonnage looked reasonable. Erection labor was present. Fireproofing was not.

She searched the cost plan again. There was no allowance for spray-applied fire-resistive material, which would cover designated steel members so they could withstand fire for the period required by the design. The architectural drawings showed rated assemblies, and a note in the structural set directed the contractor to coordinate fireproofing at protected members.

The AI had read that note. Its document viewer displayed the text among the passages linked to the steel category. But the generated cost plan treated fireproofing as included in another contractor’s scope, even though no subcontractor proposal had been attached and no line item covered the material or labor.

The confidence indicator beside structural steel was high.

That score described the system’s confidence in the quantities and category it had produced. It did not measure whether the category was complete. The model could be confident about 430 tons of steel while remaining silent about the coating applied to some of it. There was no empty row to draw attention to the missing work.

She wrote “fireproofing” beside the $18,742,600 total and kept reading.

The second omission

The accessibility work was harder to see.

The project involved an older building whose entrances, restroom layouts and door clearances did not all match the renovation drawings. Some changes appeared on demolition sheets, while the replacement work appeared elsewhere. The AI had recognized several existing doors as objects and counted them, but it had classified the associated notes as descriptions of current conditions rather than work included in the renovation.

That distinction removed automatic door operators and changes to restroom partitions from the estimate. It also left out labor around widened openings. The drawings still contained the information, but the model had separated related instructions that a human estimator would normally read together.

This was a known weakness in systems that generate takeoffs from construction documents. Drawings do not describe a project in one continuous account. A requirement may begin in a general note, appear as a symbol on another sheet and gain its cost in a specification section. Models can extract each fragment while failing to connect them, especially when existing conditions and new work share the same area.

Her manual additions came to $746,000 for fireproofing and $412,000 for accessibility work. The revised total was $19,900,600.

The difference was $1,158,000.

She took both totals to her manager. The lower number had already reached the client, and replacing it would mean explaining that the company’s new estimating system had produced a plan that looked complete but was not. The manager asked whether the missing work could be carried later as an allowance, after the company had advanced further in negotiations.

That would preserve the promised figure for the next meeting. It would also move known costs out of view, leaving the project team to recover them through changes, savings elsewhere or a dispute over who was supposed to include them.

The estimator would still have to approve the cost plan.

“If I signed it, it became my number,” she said later.

She declined to release the $18,742,600 version. She attached the drawing references supporting both additions and marked the original plan as incomplete. Her manager warned that executives were watching how employees handled the new system, and that senior estimators were expected to solve problems without slowing a pursuit.

For four days, the promotion felt less likely.

What changed after the bid

The company submitted the higher total with a short explanation of the added scope. The client did not reject it. Instead, the two sides reviewed the fireproofing assumptions and adjusted part of the accessibility design before choosing which work would proceed.

The final proposal remained above $19 million.

Inside the estimating group, the cost plan became a test case. The company stopped treating the system’s confidence indicators as a check on completeness. Reviewers began comparing generated categories against a human-prepared scope outline, with special attention to work that appears across several drawing sets.

The estimator also raised a second concern. To produce the plan, the tool had processed project drawings on a vendor’s servers. Those files included building access details and tenant layouts. The estimating team knew the documents had been uploaded, but several members could not say how long copies remained available to the vendor or whether user corrections could be retained to improve the system.

The company had approved the tool. That did not settle what each client had agreed to share. A later internal review limited which project documents could be uploaded and required the estimating team to confirm retention settings before using the system on restricted drawings.

Six months after the bid, the estimator received the senior title. Her manager did not say that the disputed total had helped. The promotion came with more client meetings and responsibility for reviewing estimates produced by other employees using the same tool.

She kept using it.

On a later renovation, the system compared a new drawing set with the previous issue and identified changed wall lengths in less than an hour. She accepted some quantities, corrected others and added scope the model had not created. The speed was real, even after the review work returned.

The original cost plan stayed in her project folder. At the bottom, $18,742,600 remained crossed out in pencil, with $19,900,600 written beside it.

Questions people ask

How can an

AI construction estimate miss required safety work?

These systems identify objects and text, then infer which cost categories belong in a plan. In this case, the tool found the steel and read a fireproofing note, but it treated the coating as included elsewhere. Its high confidence applied to the items it generated, not to the completeness of the estimate.

Who is responsible when an AI-produced bid is wrong?

In the estimator’s workplace, a person still had to approve the cost plan before it went to the client. The system produced quantities, categories and prices, but the estimator’s name carried the internal responsibility. Refusing approval forced the company to replace the promised total rather than leave known costs for later.

Did the estimator stop using AI after finding the omissions?

No. She continued using it for drawing comparisons and quantity takeoffs, where it could reduce days of work. The company added human scope checks because the system could produce accurate counts inside an incomplete plan. Her job shifted toward testing what the machine had included and noticing what never appeared.

What privacy issue did the construction drawings create?

The uploaded files contained tenant layouts and building access details, while employees were unsure how the vendor retained documents and corrections. The company later restricted which drawings could be processed and checked retention settings before uploads. Some sensitive files stayed in the project folder instead of going to the tool.