We mapped 172 companies selling AI to construction. Here is the map.

Three shapes, and one empty corner

Knowing this market is our job, so we have done the groundwork properly: 172 companies, each analysed against three questions. Does the product remove hours from the work? Does it free people? Does it get the work provably right? The companies range from tender analysis in Stockholm and Oslo and drawing review from Tel Aviv to the AI programmes of the design-software incumbents and the general-purpose agents of the AI labs. Sorted by what they do with the work, they fall into three shapes.

101 assist a task: they answer questions, draft text, find documents. 44 check a fragment: a clause, a drawing set, a code requirement. 27 take on whole jobs, and do so in other disciplines than the deliverables that engineering consultancies and contractors live on. The crowd, in other words, is at one end of the work, and the emptiness at the other. The rest of this piece walks the map: what is worth buying, what the industry builds for itself, where whole jobs are already automated, why the empty corner looks the way it does, and where we ourselves stand.

The copilots and the review tools are worth buying, with open eyes

Buy them. A good copilot takes friction out of the working day, and a good review tool pays for itself with the first error it catches. But set the expectations right at the same time. At firms that have rolled out assistants broadly, project timelines have so far not moved; the help shows up in the working day, not in the delivery plan. The deliverable stays on your desk: the copilot speeds up minutes inside the work, the review tool checks what it was built to check, and every hour of the deliverable still has an engineer in it.

Watch the acquisitions, too. Bluebeam bought the drawing-review tool Firmus AI, and Trimble is buying the contract-review tool Document Crunch: the review tools are becoming platform features. That sounds convenient, and it is, but it has a price. The feature only sees what the platform sees, your data gets bound harder to one format and one vendor, and which checks you have access to gets decided by whose license you happen to own. What looked like a tool choice becomes a platform choice.

The point solutions sit in the same part of the map. Generative 3D design, quantity takeoff, schedule analysis: tools that do one bounded technical task completely, often impressively well. But the task is a cutout of the deliverable, and the result is still carried into the whole by an engineer, together with everything else.

The industry builds for itself. Count on the numbers being the companies' own.

Alongside the 172 runs a wave that shows up in no sales demo: the consultancies' own builds. Tyréns states that its industry assistant SvenAI is used by over 1,300 people; Sweco that a majority of employees use SwecoGPT regularly; WSP's Microsoft partnership of over a billion dollars is a contract, not a measured effect. What the announcements share: the numbers are the companies' own, about their own internal products, without independent evaluation.

What has demonstrably been built are assistants: chat, search and drafting help for the person doing the work, the map's first shape. Tyréns has also launched A.Engineer, which according to the company automates calculations and reports; no independent evaluation exists there either. And none of the initiatives has so far reported the thing that would genuinely impress: a moved timeline in a project. We have written before about why it turns out this way: building an assistant is nowadays weeks of work, while verification that holds for a signed deliverable is a discipline of its own, and rarely core business even for the industry's largest.

The end-to-end systems exist, in bounded disciplines

The 27 end-to-end systems prove the shape works. Consigli designs a building's mechanical, electrical and plumbing systems to tender-ready level, and was acquired in November 2025 by the engineering group AECOM, which was buying capacity rather than software. Stockholm's Endra automates building services design in Revit. Others produce permit-ready document sets or whole housing layouts. What they have in common: each takes on a whole job in one bounded discipline, mostly design and geometry.

The AI labs' general agents belong to the same class: they can carry a long document workflow with a person overseeing, but lack the rulebook, the verification and the traceability. So the capability to take over whole jobs exists, and the acquirers price it as production capacity. What remains untouched is the form of work that dominates the industry's everyday: the written deliverables.

Products gather where responsibility is thinnest

Lay the shapes side by side and the logic appears. Helping a person with a task requires no deep connection to the rules of the work; being useful is enough. Checking a fragment requires more, but only for the fragment. Every step toward the finished deliverable raises the bar: more formats, more sources, more rules, and in the end a responsibility someone has to dare to carry. Products gather where responsibility is thinnest, and thin out where it is heaviest.

The heaviest responsibility lives in the written deliverables: the investigations, technical descriptions, reviews and tender documents that consultancies and contractors answer to the rulebooks with, in building projects as in infrastructure. A deliverable like that is finished only when it survives review by another engineer, and it carries a named engineer's signature. AI is not deterministic, and a system that produces deliverables cannot take over that responsibility. The signature stays. What a system has to do is hand over work worth signing, and that is no feature to bolt onto a chatbot. It is an engineering discipline of its own: hence the empty corner.

The analysis shows the same picture from another angle. On hours and people, the market is climbing fast. On the third question, proven ability to catch engineering errors, the map is nearly empty. And the quality claims that do exist are the vendors' own.

What Yesper does, no one else on the map does

Yesper is the AI civil engineer for construction and infrastructure. It takes the full project material and does the whole job: reads every appendix, checks the content against the rulebooks, calculates, writes and reviews, and hands back a finished deliverable where every claim can be traced to its source and every figure can be recomputed. Investigations, technical descriptions, reviews, tender documents; in building projects as in infrastructure. The engineer reviews and signs, exactly as before.

The evidence is not our own testing but customers' findings in live projects: their reviewers have seen the system catch errors that experienced eyes read past, and their deliverables, the kind that used to take weeks, have come back the same day. No names here; our customers' projects are theirs. But that is the difference from the map's self-reported numbers: our evidence comes from customers' projects, not from our own demos. The tools on the map improve the working day. Yesper does the work.