Most small businesses are using AI now. Almost none of it is sticking.
Both halves are true at once, and the gap between them is the whole story. The U.S. Chamber of Commerce put small-business generative-AI use at 58% in 2025, up from 23% two years earlier. Broader surveys run higher still: McKinsey's 2025 State of AI report puts company AI use at 88%, up from 78% the year before. Then MIT's 2025 NANDA study — a research initiative tracking real corporate generative-AI pilots — found that 95% saw no measurable financial return. Not small return. None. Only about one in twenty reached production with a result you could point to on a P&L.
So the question is not whether to use AI. You already are. The question is why the tool gets opened, used for a week, and quietly dropped.
The models are not the problem. They are good and getting better on a schedule you cannot influence. MIT's own read on the failures was blunt: what breaks is rarely the model, it is the data readiness and workflow integration around it. The AI ends up sitting next to the work in a separate chat window instead of inside the system where the work actually happens. Gartner projects that organizations will abandon 60% of AI projects that lack AI-ready data through 2026.
I have written this before in a different shape. The hardest part of running a business was never which tool. It is the thing under the tool. The rhythm. The rules. The data the tool reads from. The decisions that have to still be true six months from now. AI does not change that. It raises the cost of not having it.
Here is what the dead pilots have in common.
No business problem underneath. The pilot starts as "let's try AI" instead of "drafting a proposal takes three hours and we do six a week." When nobody wrote down what success looks like before the build, there is no way to call it a win, so it gets called a failure and dropped.
The floor was never poured. Messy data in five systems that do not talk to each other. No agreed rule for what the AI can touch. No one who owns it. The tool works in the demo and breaks on the second real input.
It was bought, not learned. A license got handed out. Nobody taught the team how to brief the thing, check its output, or feed corrections back in. Curiosity turns to distrust in about two weeks, and weekly active use goes to zero.
Tool sprawl. Six AI gadgets, none wired into the actual workflow, all giving slightly different answers. It is the oldest failure in technology adoption, and the cause never changes: the tool got bought before anyone understood the process it was supposed to fix.
I felt this from the other side early. About eighteen months ago, when I was still learning to build with AI, the forums were full of people complaining that their context bloated and the model's answers went sideways after a while. Mine mostly did not. It took me a minute to understand why. Without naming it, I had already built the thing the experts were just starting to teach: lasting context — the notes and rules the AI keeps between sessions instead of starting blank every time — plus basic frameworks saved so they carried over. A floor.
Where I did not have the floor, it showed. Memory was the weak spot for a long time. Session after session, I was reminding the system about something, sometimes trivial, sometimes a specific deploy that was already done. It did not get reliable until we fixed the mechanics underneath it. How files get tracked and updated. Where each kind of memory lives, the hot stuff loaded every time and the cold stuff filed away until it is needed. Once that floor was poured, the whole thing got sharper and more precise. Same model. Different floor.
Now the other side. The 5% where it sticks are not smarter and they are not better funded. They do the same boring things.
They start from a problem they can measure, not a tool they want to try. They pick one internal workflow, write down the number it should move in 60 to 90 days, and name who owns it. They keep the first pilot small and low-risk and out of the customer's view. Worth noting: around half of AI budgets get spent on sales and marketing, but MIT found the steadiest return in back-office automation, the unglamorous work nobody films a demo about. They put a thin layer of governance in early, one page on what data is allowed and where a human signs off. And they are willing to kill a pilot that does not earn its place, which is the part most people skip.
None of that is AI work. It is floor work, the same floor that made the last tool you bought useful or useless, and the one before that.
This is why I keep saying pour the floor first. A model dropped onto a business with no rhythm, no clean data, and no owner produces a 95% statistic. The same model dropped onto a business that already knows its process, already trusts its numbers, and already decides things the same way twice produces the boring, durable wins that show up in the 5%.
The tool is not the leverage. The floor under the tool is the leverage. AI just made that more expensive to ignore.