When the MIT figure first went around, the one from the NANDA initiative's State of AI in Business report saying roughly 95 percent of enterprise generative AI pilots delivered no measurable impact on the bottom line, a lot of people read it as a verdict on the technology, as if AI had been weighed and found wanting. We read it differently, and the more time we spend with it the more convinced we are that it's actually a verdict on how people went about the work, which is a very different and much more hopeful thing. Because if AI itself were the problem, there'd be nothing to do but wait. If the approach is the problem, then the approach is something any of us can change.
What's striking, once you sit with the research, is how rarely the failures had anything to do with the model underperforming. The pilots usually worked fine in the demo. They hit their technical marks. What killed them was everything that came after, the integration into systems that were never designed to talk to each other, the data that turned out to be a mess the moment it left the sandbox, the absence of any agreed definition of success so that nobody could even say whether the thing had worked. MIT's researchers estimated that the large majority of the effort in a successful deployment is not the AI at all, it's the unglamorous plumbing around it, and most projects simply never did that part.
There's a detail in the MIT work that we find quietly important for small businesses specifically, which is that purchased solutions tended to outperform the ones companies tried to build themselves. The instinct inside big organizations is often to construct a custom system, partly out of pride and partly out of a belief that their needs are uniquely complex, and that instinct produced a lot of the wreckage. The teams that bought a focused tool aimed at one clear job did better than the teams that set out to engineer their own everything-machine, which is worth remembering the next time someone suggests your business should roll its own.
Here's the part that we think genuinely flips the story in favor of the little guy. The reasons enterprise AI fails are almost all reasons that get worse with size. Sprawling legacy systems, fractured data scattered across decades of acquisitions, procurement processes that take a year, a dozen departments each lobbying for their own pet project, internal politics that turn a simple deployment into a turf war. A small business has approximately none of that. You might have your whole operation in your head and a few tools you can name on one hand, and that simplicity, which sometimes feels like a disadvantage, is exactly what lets a focused AI effort actually land.
We've watched this difference play out from up close, and it's almost funny how consistent it is. A small operator who decides to fix one specific, irritating, measurable problem can go from idea to result in a matter of weeks, because there's nobody to convince, nothing to integrate across, and no committee to satisfy. The enterprise version of the same idea would still be in a planning meeting. The owner can look at one number, before and after, and know with their own eyes whether it worked, and that direct line of sight is something the big company has to spend millions trying to recreate with dashboards and consultants.
The catch, and there's always a catch, is that the small business has to resist the same temptation that sank the big ones, which is the temptation to be grand. The 95 percent failed in large part because they reached for transformation, for the sweeping change, for "let's reinvent how we work," and that ambition is precisely what has no measurable edges. The 5 percent that succeeded tended to be almost boringly narrow, one process, one bottleneck, one painful recurring task, with a clear before and a clear after. Narrowness isn't a compromise here, it's the actual mechanism of success.
We think a lot about why narrowness works so well, and our best understanding is that a narrow problem is a measurable problem, and a measurable problem is one you can prove you solved. "Make my business better with AI" can never be evaluated, it has no finish line, so it drifts and dies. "Stop letting calls go unanswered at lunch" has a number attached to it, you can count the missed calls before and after, and a thing you can count is a thing you can manage and defend and build on. The whole difference between the two camps, as far as we can tell, is whether the goal had a number.
It's also worth mentioning that some of the 5 percent probably got a little lucky, and that not every focused project succeeds just because it's focused. We've seen narrow efforts fail too, usually because the underlying information was a mess or the goal turned out to be the wrong one. Focus is necessary but it isn't magic, and anyone telling you that simply being small guarantees AI success is selling something. What being small gives you is not a guarantee, it's a fair shot, the ability to try something specific quickly and find out cheaply whether it worked, which is more than most enterprises can actually claim.
Where we think this goes is that the failure statistics, scary as they sound, are going to quietly become a small business advantage as the dust settles. The big companies are entering their budget-review reckonings now, cancelling the projects that couldn't prove themselves, growing cautious and slow. Meanwhile the operator who picked one problem and fixed it is just quietly better off, with no fanfare and no press release, having sidestepped the entire failure pattern by refusing to be grand about it. The headlines will keep saying AI is failing, and in a narrow sense they'll be right, and the small businesses that ignored the headlines and got specific will keep winning anyway.
So if the 95 percent number worried you, we'd gently suggest reading it the other way. It isn't telling you AI doesn't work, it's telling you that vagueness doesn't work, that complexity doesn't work, that hoping for transformation without naming a single concrete target doesn't work. None of those are problems you have to share just because the giants did. Pick the one thing that's genuinely costing you, give it a number, aim something at it, and check the number in a month. That's the whole recipe for the 5 percent, and it happens to be a recipe that fits a small business better than almost anyone.
