Construction has been sold intelligence it can't yet use. The reason is worth understanding, because it also points to the fix.
Many of the contractors we talk to have heard the same pitch. AI will price your jobs, flag your risks, watch your cash, catch the compliance lapse before it costs you. Some of you have bought a version of it. And a lot of you have quietly noticed that the thing you were sold doesn't do what the demo did.
You are not imagining it, and you are not behind. In 2025, MIT's NANDA initiative studied 300 real enterprise AI deployments and found that roughly 95% of generative-AI pilots delivered no measurable business impact. Not 95% of the careless ones. Ninety-five percent of everyone. Gartner, surveying the broader market, places AI squarely in what it calls the "trough of disillusionment" — the point in any hype cycle where reality collects its debts.
I have two jobs that both point at this problem. At ConStrat, I'm the one accountable for whether our AI actually earns its keep for you: protecting your margins, speeding up your pay apps, and putting real numbers behind the decisions you used to make on gut feel. But I also teach it. I teach AI courses at the University of Florida, which means I spend a lot of my week explaining to smart people what these systems genuinely do and, just as often, what they don't. From that vantage point, the usual explanation for the 95%, "the technology isn't ready", is simply wrong. The technology is remarkable, but it's being pointed at the wrong problem, in the wrong order.
First: what an AI model actually does
The first thing I tell people who want to understand AI is to ignore the marketing and look at the mechanism. Strip it down and a modern AI model is a prediction engine. It was trained by reading a large amount of text and learning to predict what comes next. That's it, and it turns out that's enough to produce something that reads as fluent, knowledgeable, even reasoned. Do not get caught in the description of these systems as smart or mimicking the human brain. It's a real breakthrough, but its nature has two consequences that decide everything downstream.
One: the model knows the world in general. It does not know you. When you ask it a question, it can draw on only two things, the general patterns it absorbed in training, and whatever context you place in front of it at that moment. Now retrieval tools, and features like Claude's Projects and ChatGPT's memory, try to address this by providing memory components to ingest personal information. However, they typically have no standing knowledge of your backlog, your payment history, your certifications, or the job that went sideways last spring. If that information isn't handed to it, in a form it can read, it does not exist as far as the model is concerned.
Two: when the model lacks the context, it doesn't stop. It guesses. It's built to always produce the most plausible-sounding continuation, so a gap in what it knows doesn't return an error. It returns a confident, well-worded answer that may be wrong. The industry calls this "hallucination," and it is not a defect that can be fixed in the next version. Hallucination is a structural property of these prediction systems. These systems' main job is to guess the likeliest next word and when they lack the facts, they guess fluently and can be wrong with total confidence.
Then there's the arithmetic. Real decisions aren't one step; they're chains. Read the financials, weigh the backlog, check the exposure, reach a call. Say the model is 95% reliable at each step. Across ten steps, 0.95 multiplied by itself ten times is about 0.60. A tool that looks nearly flawless in a demo becomes barely better than a coin flip on anything with real steps in it. In a bonding or financial decision, a wrong "yes" is a default someone has to pay for.
So the model isn't the weak link. The weak link is everything you didn't or couldn't put in front of it.
Why builders get the worst of this
Here's where a general problem becomes a specifically detrimental one for our industry.
An AI model is only as good as the context it's given. In construction, that context lives in fragmented places. Finance is in QuickBooks, a payroll tool, and a spreadsheet that don't reconcile. Operations lives in a superintendent's head and a text thread. Compliance is a folder of certifications that only surfaces when one has lapsed. The real current picture of a contractor's business is scattered across disconnected tools and human memory, and almost none of it is in a form an AI can read.
This is exactly what MIT found underneath the 95%. The pilots didn't fail because the models were dumb. They failed because the tools couldn't hold context, couldn't reconcile with real workflows, and collided with disconnected, low-quality data. Point the smartest model in the world at a contractor whose data lives in a shoebox, and you don't get much value from it. You get an articulate guess. That's a foundation problem and not an AI problem.
The cost compounds
The instability isn't contained to the contractor who has it. Construction is a chain, and each link prices in the uncertainty of the one beneath it as a margin for error, a padded number, a tighter term. A trade contractor carries some ambiguity in their books. The general contractor, unable to fully verify it, prices in a cushion. The surety, sitting at the top of the stack, absorbs the accumulated fog of everyone below and prices that. A small slip at the bottom marks up at every level until a minor inconsistency at the trade level is a material question mark on a bond. One weak data point costs everyone downstream, and it costs more each time it lands.
You cannot bolt AI onto that chain and expect intelligence to come out. You have to make the chain legible first.
Foundation before intelligence
So the sequence almost everyone runs — buy AI, then figure out the data — is backwards. The order that works is the reverse. Give a contractor one place where finance, operations, and compliance actually reconcile — a stable, machine-readable picture of the business — and two things happen at once. First, the busywork falls away, and second, there's something clean enough for AI to reason over. This is precisely what separated MIT's 5% from the 95%: the systems that delivered weren't the flashiest models. They were the ones integrated deeply into real workflows and built to hold context and learn over time. The winners fixed the foundation and the losers bought a demo.
There's a part of this I care about as an educator especially. For years I've studied how tacit, gatekept knowledge — the unwritten rules that decide who gets in — locks capable people out of rooms they've earned. Bonding and credit run on that kind of hidden curriculum: a contractor is deemed "bondable" or not partly on signals that insiders read fluently and outsiders never see. When you make a business legible — reconciled, complete, readable — you do more than feed an algorithm. You hand a contractor who was never taught the code the same clear picture the incumbents have always had. Done right, this is intelligence that widens the gate to others.
The part almost no one is building for: context compounds
Here's the piece that turns a foundation into something far larger than a cleaner spreadsheet.
Once contractors operate on a shared, reconciled foundation instead of a thousand private shoeboxes, context stops being trapped inside one company. A payment pattern, a backlog signal, the early shape of a project going wrong, things invisible inside a single contractor's files become legible across the whole trade-to-GC-to-surety chain. The system doesn't just see one business more clearly; it begins to understand how businesses like that one behave, and what tends to come next.
Which means the intelligence improves as more of the industry runs on it and the advantage compounds. Researchers studying the enterprise divide make the same point from the outside: early adopters accumulate a data advantage that steadily widens the gap for everyone else. Insight built on shared, reconciled context isn't something a competitor closes by buying a better model next quarter. It's earned, it grows, and it belongs to whoever built the foundation first. Foundation-first isn't only risk management. It's the moat.
What to ask the next time someone sells you AI
You don't need to teach the course to avoid the 95%. You need to ask the questions the failed pilots didn't:
- Where does it get its context? Is it reasoning over a reconciled, connected source of truth — or a screenshot of one? If it assumes clean data you don't actually have, it will guess, confidently.
- Can it show its work? For any decision with money on the line, an answer you can't trace is an answer you can't defend to an owner, an auditor, or a surety.
- Does it get better with use, or is it frozen? Static tools stalled. Systems that accumulate context and learn are the ones that last.
- What does it do when it isn't sure? If the answer is that it's never unsure, it's guessing — fluently, and with total confidence.
Where this leaves us
The promise you were sold isn't a lie. AI genuinely will price better, see risk sooner, and widen access to capital for contractors who've been shut out of it. I've spent my career on these systems, in the lab and in the market, and I'm not talking anyone out of the destination.
However, I'm telling you the road runs through a place nobody put on the brochure. Intelligence can't act on a picture it can't see. Construction's picture is scattered, and every day it stays scattered, the cost compounds down the chain. Fix that foundation — make the business legible, reconciled, and shared — and the same compounding starts working for you instead of against you.
That's the order. Foundation, then intelligence. Everything real starts there.
Emmanuel Dorley is Co-Founder and Head of AI at ConStrat, where he builds AI systems for the built economy. He is also an Assistant Professor of Computer Science at the University of Florida, where he teaches and researches artificial intelligence.