The one number AI didn't move

29 July 2026 · Carlos Greblo

Your data is now a cost line

Ask anyone in this industry whether they use AI and you will get a slightly impatient yes. The question stopped being interesting some time last year. Everyone uses it, most of them daily, and the surveys have caught up: the number of firms using none of it has fallen to almost nobody.

So the interesting thing is no longer who adopted what. It is that after two years of this, most firms can tell you their work moves faster and almost none of them can point at the money.

Faster, better, no cheaper

The pattern is remarkably consistent, and you can see it in your own business without reading a report. Documents get drafted quicker. Reviews that used to eat a week get done in an afternoon. People produce more, and often produce better. Ask the same people what it did to the cost of delivering a project and the room goes quiet.

Autodesk's survey this year found exactly that shape. Every measure of benefit climbed except cost, which sat where it was while spending on AI went up across the board.

There is a charitable reading. Savings lag investment, the tooling is still being paid for, the return is coming. Fine. But there is a harder reading that anyone who has written a business case will recognise. AI is sold on cost and delivers on throughput, and those are different promises to different people. The tool you justified on headcount is now defended on cycle time, in a meeting where someone has kept the original slide.

If you are about to run a pilot, decide now what would count as it working, and make it a cycle-time or rework claim. That is the one the evidence will support when you are asked.

The thing nobody can buy

Here is why the money hasn't moved, and it is not a technology story.

Every serious attempt to put AI to work in a construction business runs into the same wall, and it is never the model. It is that the information the model needs is spread across four systems that do not speak, named three different ways by three different people, and stored in a folder structure that made sense to someone who left in 2023. The specification does not match the model. The drawing register has duplicates nobody has had time to resolve. Somewhere there is a file called Final_v7_draft_final_signed and it is not final.

None of that is new. What is new is that it now has a price.

For a long time, poor data hygiene was an irritation. It cost you an afternoon of hunting, an argument at a coordination meeting, the occasional rework. It was absorbed, the way everything gets absorbed. Three things happening right now stop it being absorbable.

What is arriving, and what it charges

The agent that only works if it knows where to look. The word is being used for at least three different things, and the gap between them matters commercially. There is the assistant that runs a fixed sequence of steps. There is the coordinator that moves work between tools and systems. And there is the thing in the keynote, autonomous, understanding intent, adapting as it goes. Autodesk's own research lead calls that third one aspirational and not realised in most practical applications, which is a striking thing to find in a vendor's report and worth holding in mind at your next demo.

The middle version is real and deployable. Haskoning has hundreds of agents doing genuine work, tender reviews and contract analysis, the reading nobody volunteers for, and has trained thousands of its people to use them. That is not a science project. But look at what makes it possible: a bounded job, a defined scope of data, an output someone can check. The scope is doing the work. An agent pointed at a well-kept system is useful within a week. An agent pointed at your shared drive is a support ticket.

The machine that builds what your model actually says. At CES this year Caterpillar showed five autonomous machines: wheel loader, dozer, haul truck, excavator, compactor. Sensing, deciding, working without anyone in the cab. Mining has done this for years at scale. The jobsite version is new, and it will not stay a Caterpillar exclusive for long.

The instinct is to read this as machines coming for jobs, and on a site where you already cannot fill the vacancies, that instinct is wrong. Everybody in the industry knows the real shape of it: the experienced people are leaving faster than anyone is arriving, half the site knowledge lives in the heads of men within a few years of retiring, and there is nobody queued up behind them. Autonomy is walking into that hole. Judge it against the crew you cannot hire.

Then think about what these machines take their instructions from. Your model. The one with the clash everyone agreed to sort out on site. A grading tolerance stops being a drawing-office concern the moment a forty-tonne machine treats it as an instruction, and it will build your mistake perfectly, at speed, without once looking up and thinking that seems wrong.

The client who reads your submission with software. In June the government put a tool called Extract into the hands of every council in England. It takes planning documents, including handwritten and historic ones, and turns them into structured data. In the councils that trialled it, work measured in hours per document came back measured in minutes. The plan is to cover every planning document type by the end of the year.

This is the quiet one and it is the one that changes the relationship. Until now, AI in construction has been something firms pointed at their own work, on their own terms, at a pace they chose. This is the other party pointing one at yours. Put it next to a Building Safety Regulator that already judges applications on whether the information in them is adequate, and something shifts underneath you. A submission used to be a document written for a person, a person who might read the covering letter, understand the intent, give you the benefit of the doubt. It is becoming a dataset handed to a process that will do none of those things.

A structured submission survives that intact. Consistent identifiers, real classification, metadata that a machine can read without guessing. A scanned drawing with the important information sitting in the annotations does not, and there is nobody on the other end to explain yourself to.

Three arrivals, one bill. Each one converts the state of your data into money or delay, and none of them cares how good the model is.

The deadline that did not move

Europe gave the industry some breathing room this year. The omnibus deal pushed the AI Act's high-risk obligations back, standalone systems to late 2027 and AI embedded in regulated products to 2028. If you built a compliance plan around the original dates, you have time you did not expect.

Less than the headlines suggest. Most of the Article 50 transparency obligations were not deferred and apply from 2 August 2026. And none of it touches the thing that actually reaches you first, which is a client asking AI governance questions in a procurement process, whenever they feel like it, deadline or no deadline.

What has changed more than the law is the character of the anxiety. Two years ago the worry was that AI would take the work. Now the worry is that it will get the work wrong, and quietly, and in a document with your name on it. That is the more serious fear and the harder one to design around.

Use the delay for classification work, not as permission to defer governance. Know which of your uses fall under Article 50. Be able to state, in writing, what a given model was given and what it produced. Which is, once again, a question about your data wearing a compliance lanyard.

What this is built on

ReqTwin assumes the constraint instead of the model.

Its AI panel answers against one project's own requirements, inspection plans, programme and bill of materials, retrieved by meaning rather than keyword, so the answer lands on rows you can open and check rather than a paragraph that sounds right. Classification suggestions carry a confidence figure and a threshold, so the obvious matches go through in bulk and the ambiguous ones stay in front of someone who knows the job.

Neither is impressive as a model demonstration, and that is the point. Both work because the data underneath was in order before anyone pointed a model at it.

The firms getting real value out of AI are not the ones with the best model. They are the ones whose data was worth asking a question of.