AI’s blind spot is the building industry
This is general information, not legal or financial advice. Your counsel reviews any license before you sign.
AI models are tested on how well they do professional work. For the people who design and build the physical world, those tests barely exist. The gap is measurable, and it points to where the missing material sits.
What the benchmark map shows
In October 2026, Mercor published its Benchmark Map, which checks which occupations have meaningful AI benchmark coverage. Of the 61 Construction and Extraction occupations, 2 are covered.
The coverage inside those occupations is thin too. According to the same map, Architects have 2 of 24 tasks covered, Civil Engineers have 2 of 16, and Construction Managers have 1 of 25.
The uncovered tasks are the ones that carry the most judgment. The map finds zero coverage for resolving construction problems with owners and contractors, negotiating contract revisions, and adjusting estimates for changes. Those are decisions, and nobody has built a public test for them.
Benchmarks test reading, and projects run on deciding
The benchmark that does exist for this industry tests reading. AEC-Bench, from Nomic AI and Aurecon (March 2026), measures how models handle construction documents. On submittal review, frontier models scored between 6.5 and 23.1 out of 100.
A submittal review is a bounded task. The reviewer compares a product submission to the specification and the drawings, and marks it approved, revised or rejected. If the best models struggle with that, they are a long way from the harder questions that follow: whether a revised detail is worth a change order, who carries the cost, and what to tell the owner.
Those questions are decided through RFIs, change orders, meeting notes and email. The decision is written down at the time, by the people who made it. A benchmark built from drawings alone cannot see any of that, which is why a model can score well on reading and still have nothing to say about what happens next on site.
The decision traffic is large and costly
A Navigant study from 2013 found about 796 RFIs on a typical project, at about $1,080 each to process, with a median response time of 9.7 days.
Rework adds to it. Hwang and colleagues, writing for ASCE in 2009, put rework at about 5% of construction cost.
The wider picture is slow progress. A 2025 Richmond Fed analysis found that US construction labor productivity in 2023 sat at its 1948 level. An industry with that record has a real use for tools that help with coordination and review.
Where AI already does well
Some applications work when the task is narrow and the answer can be checked. In 2026, the City of Seattle reported that its AI plan review agreed with human reviewers 92% of the time.
Plan review checks a submission against written code. Project decisions have no single rulebook. How an RFI gets resolved depends on the contract and the people involved, and the reasoning lives in the archive of whoever handled it.
The decision record lives in firm archives
A model can learn to read a drawing from public examples. It cannot learn how a project team resolved a conflict between a structural detail and a mechanical run from public text, because that exchange was never published.
It sits in the RFI logs, submittal registers and change-order files of architecture, engineering and construction firms. Each thread shows what was asked, how long the answer took and what was decided. Records that link a thread to what it cost and how it ended are the scarce ones.
Builders of AI tools for this industry need that material to train and to test their models. The public benchmark map shows how little of it exists outside firm systems.
What a decision record looks like
Take a single RFI. A contractor asks whether a beam can be lowered to clear a duct. The engineer answers, the architect checks the ceiling height, and the owner learns that the fix adds cost. A change order follows, and the estimate is adjusted.
That chain touches three of the tasks the Benchmark Map lists as uncovered: resolving a construction problem with the owner and contractor, negotiating a contract revision, and adjusting an estimate for a change. A firm that kept the thread holds a worked example of all three.
Multiply that by the 796 RFIs on a typical project, and a firm with ten years of closed projects holds a large set of worked examples that no benchmark author has been able to collect.
Licensing is how the gap closes
Firms can supply that record on their own terms. A license can limit the use, set a term, require deletion of raw files and keep the firm’s ownership intact. Client-confidential material, federal work and personal data stay out.
Done that way, the firm is paid for work it already did, and the buyers get records that show how decisions were made. Better tools for RFIs, submittals and change orders benefit the firms that use them.
A firm that wants to take part should start with an inventory. Which systems hold the threads, how many years they cover, and which contracts restrict their use. Counsel reads the contracts, and anything with a restriction is set aside before a buyer sees a summary.
The alternative is a gap that stays open. Models built without this record have little to learn from, and the firms that hold it have no say in the tools their staff end up using.
Archeon represents AEC firms in licensing their records, with a disclosed success fee and nothing upfront. The firm keeps ownership.