AI in Construction Estimating: What Is Real, What Is Still a Demo, and How Contractors Can Tell

Construction runs on drawings. Before a contractor can bid on a building, someone on their team does a takeoff. They go through the plan set, the bundle of PDF drawings the architect and engineers issue for the project, and count and measure everything the contractor will supply and install: outlets, doors, sinks, linear feet of wall, square feet of flooring. Those quantities go into the estimate, the priced document that becomes the bid. If the bid wins, the same numbers come back in change orders, the priced and signed amendments to the contract when the design moves.

Takeoff is slow, repetitive and expensive to get wrong, which makes it an obvious target for AI. I build software that reads construction drawings, and the contractors I talk to are trying to decide which AI claims to believe. Here is how I separate the parts that work from the parts that are still a demo.

Optimism Is Running Ahead of Adoption

Two recent surveys put numbers on the gap. In a Dodge Construction Network study with CMiC, published in December 2025, 87% of contractors said they expect AI to transform the industry, and 19% said they had adapted their workflows for it.

RICS, the professional body for surveyors, polled more than 2,200 construction professionals for its 2025 report on AI in construction. About 45% said their organizations had no AI implementation, and fewer than 1% described AI use as fully embedded across the organization. Construction Dive covered the findings under the headline “Builders slow to adopt AI despite perceived benefits.”

Optimism and adoption in two recent surveys. Data: Dodge Construction Network and CMiC (December 2025), RICS (2025). Image: Bildrix

Preconstruction, everything a contractor does before ground breaks (bids, document review, scheduling), is a natural place to start, because the inputs are already digital documents.

What Works Today

Reading a plan set is now a working product category. Software can sort and name the sheets in a set, find each drawing’s scale, count symbols such as outlets, light fixtures and plumbing fixtures against the legend the engineer drew for that project, and measure rooms and walls. Symbol recognition on drawings has been an active research area for years. The FloorPlanCAD dataset, released in 2021, annotated more than 10,000 floor plans across 30 object categories so researchers could train and test models on this task.

Two other capabilities hold up in practice. One is question answering over drawings and specifications, where the answer cites the sheet it came from, so a person can check it in seconds. The other is carrying reviewed quantities into an estimate, as long as the contractor’s own price book covers the items.

What Is Still a Demo

Three claims deserve more suspicion.

Pricing that nobody reviews. Material and labor prices are regional, and they move. An AI-suggested price can be a reasonable starting point, but it should be visibly tagged as a suggestion and checked against a supplier quote before it goes into a bid.

Every trade from any input. Clean vector PDFs, the kind exported straight from design software, are the easy case. Scanned sheets, phone photos of drawings and partial sets still break things, and coverage varies a lot by trade. A tool that reads electrical drawings well may do little with another trade’s sheets.

Fully automatic bids. Much of what decides a bid never appears on the drawings: scope gaps between trades, means and methods, existing conditions on a renovation, the schedule, and how much risk the contractor is willing to carry. The estimator signs the number and answers for it.

What holds up in practice today, and what is still a demo. Image: Bildrix

Four Signs You Are Watching a Demo

1. It only runs on the vendor’s drawings. Ask to see it on your own plan set, ideally the messiest one you bid last quarter.
2. It quotes a single accuracy figure with no definition. Accuracy at counting what, on which drawings, measured how? A number with no per-item breakdown tells you very little.
3. Nothing is flagged. Every model is unsure about something. A tool that never shows doubt is hiding it.
4. The output ends in a spreadsheet export that someone retypes into the estimate. Every retyped number is a new chance for a typo.

Questions to Ask Any Vendor

• Which inputs does it accept, which trades does it cover, and what does it refuse to do?
• How does it show uncertainty, and how long does review take on one of my sets?
• Can I edit the output, and do my edits carry into the estimate and into later changes?
• Does my data train anyone’s models, and where is it stored?

The NIST AI Risk Management Framework is a useful reference for the last group. It describes trustworthy AI as, among other things, valid and reliable, secure and resilient, accountable and transparent, and privacy-enhanced. Each of those turns into a question a contractor can put to a vendor.

Why Doubt Is a Feature

Models are often more confident than they should be. Guo and colleagues showed in 2017 that modern neural networks tend to be poorly calibrated. The confidence a network reports often does not match how often it is right. Calibration can be improved, and once it is, a confidence score becomes a work queue. Low-confidence items go to a person, and high-confidence items get a quick look.

That property matters more to me than any headline feature. An estimator does not have time to re-check every count on a set of a few hundred sheets. They do have time to check the few dozen items the software marks as uncertain. The estimator still signs the bid, and good tools make that review fast.

The software reads and counts, a person reviews what it doubts, and the estimator signs the bid. Image: Bildrix

Unite.AI recently ran a piece by Baber Javed on designing AI for payroll, admissions and other domains that punish mistakes. He argues that deterministic rules should control anything with financial or legal consequences, with the model interpreting and explaining, and with human review where errors are costly. A construction bid fits that description.

What to Watch Next

I expect the next year of progress to come after the first takeoff: comparing drawing revisions, and carrying changed quantities through to change orders and invoices without retyping. Revisions are where contractors lose money without noticing, because a change nobody priced is work done for free.

If you run a contracting business and want to try any of this, pilot it on your own drawings, keep a person reviewing every output, and write down your own numbers: hours spent, errors caught and errors missed. You will learn more from them than from any vendor’s figures.

What would you need to see before you trusted a machine-read quantity in a bid with your name on it?