Everyone’s adopting AI. Almost no one is seeing the return.

There’s a number from MIT that should bother more people than it does. In its 2025 study of enterprise AI, the NANDA initiative found that 95% of organizations are getting zero measurable return on their generative AI investment, despite tens of billions of dollars spent. Adoption is nearly universal. Return is not. Most of the spend is real, most of the effort is genuine, and most of the payoff is still theoretical.

The easy explanation is that the technology isn’t ready. That’s mostly wrong, and it lets everyone off the hook. MIT was explicit about this. The blocker wasn’t the models and it wasn’t regulation. It was what they called a learning gap. Generic tools succeed at one-off tasks like drafting an email, because the value is immediate and self-contained. They fail inside real workflows, because a workflow needs context, memory, and the specific rules of how your business actually runs. The models are good enough. What they land on, in most companies, is a set of processes that were never legible enough to hand to a machine in the first place.

This is the systems blind spot again. Bolting AI onto an operation held together by a few people’s judgment and a spreadsheet only one of them can read doesn’t automate the work. It automates the confusion, faster. You get outputs nobody trusts, sitting next to the manual process everyone still runs in parallel, because the manual process is the only one anyone can actually follow.

The pattern is almost boringly consistent. A team runs a pilot, the demo is impressive, and then it quietly dies on the way to production. MIT put a number on that too. Only about 5% of custom enterprise AI tools ever graduate into real operational use. The other 95% stall at the starting line. The AI didn’t fail. The operation did, and the AI just made the stall visible.

There’s a detail in the same research that gives the whole thing away. Companies pour the majority of their AI budgets into sales and marketing, but the clearest returns MIT found were in the back office, in the unglamorous operational work of streamlining how things actually get done. The money chases the visible story. The payoff sits in the plumbing. That is not a coincidence. It is the tell.

The same research points at how the winners get there. Buying and partnering with specialized vendors succeeds about 67% of the time. Building the tool internally succeeds only a third as often. Read that the way an operator would. The companies that win aren’t the ones with the most engineers or the biggest ambition to build it all themselves. They’re the ones who are honest about what they’re good at, plug in a proven tool where it fits, and put their energy into the operation around it. Going solo feels like control. In the data, it mostly looks like a lower success rate.

AI is turning out to be an unusually honest diagnostic. It pays off where the underlying system is clear, and it stumbles where the operation is held together by heroics. The companies seeing returns aren’t the ones who bought the most tooling. They’re the ones whose operations were legible enough to absorb it. The tool didn’t create the advantage. It revealed who had done the boring work first.

None of this is a fixed state, and it’s shifting faster than the skeptics expect. The returns are already starting to concentrate. Recent PwC research found that a small cohort, about a fifth of the companies surveyed, is capturing roughly three quarters of the AI-driven value. Other studies are watching the share of initiatives that hit their expected return climb out of the basement as more of them cross from pilot into production. The winners keep sharing one habit. They build measurement and process discipline in from the start instead of bolting it on after the demo. The gap is stopping being about who adopted AI. It’s becoming about who was operationally ready when they did.

So the useful question was never whether to adopt AI. Everyone already is. The question is whether your operation is legible enough that AI has something to grab onto. If the honest answer is no, that isn’t a reason to wait on the technology. It’s a reason to do the systems work now, because the return was always going to depend on it, and the window where readiness is still a differentiator won’t stay open forever.

Based in Calgary, working across Canada and remotely. If you’re spending on AI and not seeing it land, the first step is a 30-minute conversation about what’s underneath it.