For two years the story of corporate AI has been a puzzle: enormous spending, genuine enthusiasm from the people using the tools, and almost nothing to show for it on the bottom line. This week a cluster of new research moved the conversation past simply noticing the gap and toward explaining it. The emerging verdict is unflattering to executives and reassuring to engineers. The models are mostly doing their job. The organisations around them are not.
Start with the numbers everyone keeps citing. A preliminary report from MIT's Project NANDA found that roughly 95 percent of integrated enterprise generative-AI pilots showed no measurable effect on profit and loss, with only about 5 percent extracting real value. It is one unreviewed study, drawn from more than 300 disclosed initiatives and interviews through mid-2025, and worth holding loosely. But its diagnosis is the interesting part. The researchers concluded that "the core barrier to scaling is not infrastructure, regulation, or talent. It is learning." The tools get dropped into a workflow and stall, because most of them do not retain feedback, adapt to context or improve over time. Impressive in the demo, forgotten by the second week.
Gallup's 2026 workplace data points to the same place from a different angle. Inside companies that have adopted AI, 65 percent of employees say it has improved their productivity, yet only 12 percent strongly agree it has transformed how work actually gets done. Just 13 percent of US workers use AI daily; nearly half never touch it. What separates the organisations getting results is not better software but active management. Where a manager visibly supports AI use, 79 percent of employees use it frequently; where they do not, that figure falls to 46 percent. Only 21 percent of employees strongly agree their manager offers that support, and only a quarter say their company has communicated any clear AI strategy at all.
The economists arrive at the same corner. Writing for the American Enterprise Institute, analysts note the widening distance between what AI can do in a lab and what it delivers in a workplace, pointing out that if AI were genuinely lifting productivity at the frontier, it would show up in total factor productivity. It has not: by the San Francisco Fed's estimates, that measure has hovered near zero over the past year. This, they argue, is exactly what economic history predicts. New general-purpose technologies pay off only after firms reorganise themselves around them, and that reorganisation takes time.
Vendors, unsurprisingly, offer a more purchasable explanation. Research from the enterprise software firm Aptean found 77 percent of leaders saying general-purpose AI simply cannot handle the complexity of their operations, and 82 percent reporting that integrating AI with core systems is harder than the AI technology itself. Their prescription is industry-specific models tuned to a company's data and rules. A separate survey summarised by MarketScale adds a twist: enterprises are seeing returns, just not where they budgeted for them, with gains landing in insight generation and customer engagement rather than the headcount savings that justified the spend. When the payoff shows up in the wrong column, the KPIs, not the technology, may be what needs rewriting.
Put these together and the finger stops pointing at the model. The recurring culprits are workflows never redesigned, managers never enlisted, strategies never communicated, and budgets aimed at the exciting use cases rather than the dull, back-office ones where MIT found the clearest returns. That is genuinely good news, in a sense, because management problems are fixable without waiting for the next frontier release. It is also harder news, because writing another cheque is easy and rewiring how an organisation works is not. The companies in that lonely 5 percent appear to have done the second thing. Most of the rest are still shopping.