Working copy
Draft text, rebuilt from the manuscript on every change. Stubs and source passages are still here, and none of it is final. Back to the working copy contents
Drafted The chapter is written and still being revised.
The vendors: players, not referees
My first reaction to a survey of what construction's software vendors are building with AI was that this is sort of good news. The technology is being used, which is good, but the vendors may be able to outcompete an open approach thanks to high levels of service and pockets deep enough for a user to sue for product malfunctions. Still, there seems to be room for an open solution to sprout. The rest of this chapter is what the survey found, sorted by who sells what to whom.
The big platforms hold the model and the project record. Their bots read PDF drawings, count and measure from them, and spot what changed between revisions. Their plugs let an outside AI work inside a model or a project file. Their newest agents compare subcontractors' quotes for a General Contractor. To them, "procurement" means tracking materials after the contract is signed. Nothing they sell checks a price against the market.
The estimating specialists sell to bidders. Their agents draft a first estimate from the firm's own past jobs, flag every number the AI produced, and leave the final call to a person. One sells itself on the line that nothing finalizes without your approval. This is the bidder's private cost check, and a public market check would sit beside it, not replace it.
The startups price straight from PDF plans with a general model and a supplier's price list. This is very close to our approach. If done cheaply and well, bidders could use it without needing a new standard. None of them publishes how far its prices miss, which is the sentence this chapter keeps coming back to.
The bid platform is a species of one. In the United States, a single company sits on both sides of nearly every state highway letting: the agency's estimate, the bid file, the encrypted bid, the results and the analysis afterwards. It sells the states' own bid data back to the industry under licence. It does not price a tender. It owns the memory.
Germany's vendors live in a different climate. Public tenders there have travelled as data files since the 1980s, under a standard kept by a public committee since 1966, and a dozen programs compete to read and write them. The programs are certified, and the country's highest civil court has upheld throwing out a cheapest bid that arrived only as a PDF. Nobody there publishes a price error either, but the referee's whistle, at least, exists.
The newest species is not a vendor at all. It is the general model with a chosen level of effort. The pace of the compute build-out makes me think the specialized tools' edge is temporary. The option to pick a model plus a level of effort should resolve the misses over time faster than the specialized tools can update. Alternatively, AI will eat the costs and margins of the specialized tool. The one head-to-head found so far is from another profession, legal research. In 2025 a general model scored level with the specialized tools on accuracy and both beat the lawyers, while the specialists kept a small edge on citing the right authority, from their own databases. The data was the moat. The writing was not.
Who is liable today
Every major vendor's terms for its AI say the same thing in different words. Liability is capped at a year's fees or a thousand dollars. The outputs are recommendations only. Verification is the user's job. One platform's own AI suppliers must indemnify it for their errors, which is the same rule facing the other way. So the deep pockets a user might sue are, by contract, empty for exactly the loss this book cares about: a number the bot got wrong and the user relied on. Today the user pays, and the vendor sells the shovel.
What they need
A target to build to. The internet's engineers learned that when a format tolerates malformed input, its flaws become rules that every new program must copy, which raises the cost of entry. A strict, public format lets a new tool compete on equal terms, and a published check does the same for a bot: pass the test and you are in the market, whoever you are.
Demand. The software vendors offering estimation are not actual competition. They are a source of demand. Each tool sold is a user who now relies on numbers under terms that leave the risk with them, and who will want a check they can show a bonding company or an owner. Email is the precedent: when the big mail providers required public authentication, about two million domains adopted the standard in a year and most still had not, which is a standing market for the firms that help them.
Protection from the general models. Insurance on the outputs would generate a market for a premium. Being able to use numbers without being solely responsible for them would protect the specialized tool from general AI programs. The market for sign-offs will pressure AI to become treated as licensed professionals. Standards would provide the licensing, and insurance the insurance. The evidence so far says the licence goes to the firm, not the machine. England's solicitors' regulator authorised the first AI-driven law firm in 2025 and said named, regulated solicitors remain accountable, with insurance as a condition. Several American states have barred legal personhood for AI. That still gives the vendor what it wants. A tool that passes a public test, made by a firm that carries the insurance, is licensed and insured like a professional, and liable through its maker.
Their moat, kept. The vendors' real moat is data, service, habit and compatibility, and a standard touches none of those. What it removes is the moat made of friction, the lock-in and the private format, and they will notice.
What a standard does for them, and what it costs them
It gives them a certified market. The project's answer is a certification like Germany's, hosted by a standards body, ISO, ANSI or CSA, and checked by the normal auditing process. Many tools compete on a public format and a public test. The vendors become test-takers, not targets.
It gives them the referee nobody is building. No vendor, startup, platform or German program publishes how far its prices miss. A public test with a measured miss rate is what lets a buyer, a bonding company or an insurer tell a good bot from a cheap one, and it is what lets a cheap one prove itself. That cuts both ways. A standard that makes cheap code credible squeezes the margins of expensive code. What if AI eats the other programs? If it does, code gets cheap and trust does not, and the rulebook the cheap tools run under is the scarce thing left. AI plus a standard would hit them in the profits. I had hoped this would be a greenfield exploration, but it looks like my goals will be contested: on the data, where open bid results undercut a licensed dataset, and on the standard, where every incumbent wants a hand on the pen. On the referee itself, the field is still empty.
And it gives them the premium market. Insurance on outputs needs a measured error rate, and nobody has one for construction prices. The standard's test is the table an underwriter prices from. Insurers already back vendors' performance guarantees in other fields when the error is measured. The vendor that wants protection from the general models gets it the same day someone measures the miss.
Vendors join public tests they expect to do well on, and they notice a standard when a buyer writes it into a tender, not when a book describes it. The pitch to them is the one made to every player: the referee is not your competitor. Chapter 1 said the vendors would like to own the sun. A standard is the industry saying, in writing, that the sun was never for sale.