Start with the number that has been quietly haunting boardrooms for the better part of a year. According to MIT’s GenAI Divide: State of AI in Business 2025, a report out of the university’s NANDA initiative, roughly 95 percent of enterprise generative-AI pilots have delivered no measurable impact on profit and loss. Only about 5 percent produced rapid revenue acceleration. The study drew on 150 interviews with business leaders, a survey of 350 employees, and an analysis of 300 public AI deployments — not a small sample, and not a fringe source.
Sit with that for a second, because it cuts against almost everything else you have read about AI in the last eighteen months. The dominant narrative has been one of relentless, near-universal adoption: surveys showing 80 or 90 percent of companies “using AI,” budgets ballooning, vendors reporting record demand. And that narrative is not wrong, exactly. It is just measuring the wrong thing. Adoption and transformation turn out to be two very different numbers, and the space between them is where most of this year’s software spending is quietly disappearing.
The gap between “using AI” and “getting anything from it”
Here is the tension that makes this story worth telling rather than just repeating. If you read the headline adoption figures, AI has already won. If you read MIT’s ROI figures, AI has barely started. Both can be true at once, and understanding why is more useful than picking a side.
The MIT authors call the divide a “learning gap,” and their explanation is refreshingly unglamorous. The problem, they found, is not that the underlying models are bad. It is that generic tools like ChatGPT shine for individuals precisely because they are flexible, but they stall inside enterprises because they do not learn from or adapt to a company’s actual workflows. A chatbot that dazzles in a demo turns brittle the moment it has to remember context, follow a real process, and improve over time. The pilot looks great to the executives who greenlit it and useless to the people who have to work with it.
That framing matters because it reframes the failure. It is not a technology failure. It is an integration failure — an organizational one. And organizational problems do not get solved by waiting for a smarter model to ship next quarter.
Three findings that should change how you spend
Buried underneath the viral 95 percent figure are three findings from the same report that are, frankly, more actionable than the headline.
The first is a misallocation of money. MIT found that more than half of generative-AI budgets are being poured into sales and marketing tools, while the biggest returns actually showed up in back-office automation — the unglamorous work of cutting outsourcing costs, trimming agency spend, and streamlining operations. Companies are spending where the excitement is, not where the payback is.
The second is a buy-versus-build gap that is hard to ignore. Purchasing tools from specialized vendors and forming partnerships succeeded about 67 percent of the time, according to the report. Internal builds succeeded at roughly a third of that rate. That is a striking spread, and it runs directly counter to the instinct in many regulated industries — finance especially — to build everything in-house for control. The data suggests that going solo is, more often than not, the more expensive way to fail.
The third is the quiet rise of what the researchers call “shadow AI.” Even where official pilots stalled, employees across most firms were already using consumer AI tools on their own, unsanctioned, to get real work done. In other words, the technology was delivering value — just not through the sanctioned, budgeted, boardroom-approved channel that companies were measuring. The grassroots adoption was working better than the top-down rollout.
Now the pushback, because it matters
A responsible read of this story has to include the counterargument, and there is a real one. When the MIT figure went viral, a number of analysts pointed out that a “95 percent failure” headline is doing a lot of rhetorical work. “Failure” here means no measurable P&L impact yet — a high bar, applied early, to a technology most companies have been deploying in earnest for only a year or two. By that standard, the early years of enterprise cloud, or the web itself, would also have looked like mass failure. Skeptics have argued the number says more about the difficulty of measuring value than about the absence of it.
That critique is fair, and it should temper any doom-laden reading. But notice that it does not actually rescue the optimistic narrative either. Whether you call it “95 percent failed” or “95 percent haven’t proven value yet,” the practical takeaway for anyone choosing software is identical: adoption alone guarantees nothing, and the returns are concentrated in a small minority of deployments that were structured differently from the rest.
What this means if you actually run something
Strip away the hype and the anti-hype, and the signal underneath is genuinely useful. The organizations landing in MIT’s 5 percent were not the ones with the biggest AI budgets or the flashiest internal labs. They were the ones that picked a single painful, well-defined problem, chose a tool that could integrate deeply and adapt to their workflow, empowered the line managers who actually own the process, and were willing to buy rather than build when buying was the surer path.
That is not a description of a technology bet. It is a description of disciplined operations — the same discipline that separates good software adoption from bad in any category, AI or not. Which is, perhaps, the least surprising and most reassuring conclusion in the whole report: the tools have gotten radically more capable, and the thing that still determines whether they pay off is how thoughtfully a human being decides to use them.
If 2025 was the year everyone adopted AI, 2026 is shaping up to be the year the bill comes due on how carelessly a lot of that adoption happened. The winners will not be the companies that used the most AI. They will be the ones that were honest about what was actually working.
Related reading: Open vs. Closed AI Models: The Gap Closed, the Trade-Offs Didn’t · Free Mac SEO Software Worth Your Time
