Why the AI wave skipped construction, and why that is about to change

Construction is one of the largest sectors of the economy. Depending on how you count, it’s somewhere around 4% of GDP in the United States and employs close to eight million people. It is also, by almost any measure anyone has tried, the least digitized major industry in the country. Studies going back a decade put it at the bottom of the list, next to agriculture and hunting.

Both of those facts have been true for a long time. The interesting question is why the last few years of AI investment, which touched nearly everything, barely touched us.

AI money went where the data already was

Here’s my read, after twenty years watching how this industry actually runs.

Machine learning needs something to learn from. The industries that got transformed first were the ones that already had clean, structured, digital records of what they did. Finance had transaction histories going back decades. Retail had every purchase logged. Software companies were instrumented from birth; every click was a data point. Healthcare had electronic records, however imperfect.

Those industries didn’t have to be digitized before AI could help them. They already were. The models had something to eat.

Construction had none of that. The information exists, in enormous quantity, but it lives in text threads, in truck notebooks, in a superintendent’s memory, in a PDF someone photographed on a tailgate. Every job is a one-off. Every set of plans is different. The schedule is a whiteboard. The budget is a spreadsheet that was accurate in March.

You cannot train a model on that. So the capital went elsewhere, and the gap between construction and everyone else got wider.

The thirty-year pitch that never worked

The industry’s answer, for about three decades, was to ask construction to digitize itself first.

Get the plans into the system. Get the schedule into the system. Standardize your process. Train the crew. Then, at the far end of all that work, software can help you.

It’s a reasonable pitch. It almost never worked, and I don’t think that’s because contractors are stubborn or backward. It’s because the payoff was years out and the cost was immediate. If you’re running five people across three jobs, you do not have a spare quarter to restructure how you operate. You have Tuesday.

The large firms did it, partly. They had the staff. That’s why the tooling gap between large and small contractors is so wide, and it’s not really a technology gap at all. It’s a ‘who has time to feed the system’ gap.

What changed

The thing that changed in the last two years is that the models got good enough to work with the mess.

They can read a plan set. Not a clean CAD file, an actual photographed sheet with a coffee ring on it. They can pull dimensions off it, read the schedules, and count what’s there. They can take a job described out loud in a driveway, in ordinary language, and turn it into a material list with quantities.

In other words, they no longer need the industry to clean itself up first. Which was always the thing that was never going to happen.

That’s a bigger deal than it sounds. For thirty years, the barrier to construction tech wasn’t the software. It was the data entry. Every product that failed, failed because somebody had to feed it. Remove that requirement and the whole calculation changes.

Why this decade looks different

I’ve been skeptical of construction tech for most of my career, so let me be precise about what I’m claiming.

I’m not saying construction is about to be transformed by AI. Most of what happens on a jobsite is physical work and human judgment, and that isn’t changing.

I’m saying the specific barrier that kept software out of the small contractor’s truck, the requirement that someone sit down and type everything in, has been removed. And that’s the barrier that mattered. Not because construction finally decided to modernize. Because the technology finally met the industry where it actually is.

What nobody’s talking about yet

There’s a second-order effect to all this that I haven’t seen anyone write about.

If the tools no longer need clean data going in, then the whole idea of what a ‘construction problem’ is starts to shift. Most of what goes wrong on a jobsite isn’t a building problem at all. It’s something else wearing a costume.

That’s the next post.