Enterprises are pouring money into AI, and most of them still can't say whether it paid off. We read through this year's big AI surveys, and they're strikingly consistent on that point. They're even more consistent about why.
The measurement gap
Almost no one can prove it
The numbers are uncomfortable. Forbes Research found that fewer than 1% of companies report significant ROI from their AI. McKinsey's latest State of AI has only 39% able to attribute any EBIT impact to it at all. Nearly half of organizations (46%, in Wavestone's count) have no structured way to calculate ROI in the first place. And MIT's widely cited GenAI Divide put it most bluntly: for most companies, corporate AI has so far delivered no return they can measure.
When we dug into why, the reason turned out to be more mundane than the technology. The problem usually isn't the models. It's that almost nobody is keeping score.
Most companies never measure whether their AI works
Among professional-services firms, only 18% formally track what their AI delivers. Source: Thomson Reuters, 2026 AI in Professional Services.
And the firms that do measure tend to count the easy things. Cost savings and raw usage sit at the top of the list, mostly because they're the simplest numbers to pull. The metrics that actually stand for value to the business, like client satisfaction, revenue, and new business won, sit at the bottom. Fewer than one in four track them at all.
The firms that do measure count the easy things
% of measuring firms that track each metric
What measuring firms track, by metric. The value-bearing outcomes are measured least. Source: Thomson Reuters, 2026 AI in Professional Services.
You can't show value you don't measure.
A small group does better. McKinsey calls them high performers, and they're about 6 percent of the field. What sets them apart isn't fancier technology. It's that they actually keep track. They set targets, watch the numbers, and change course when the numbers tell them to. For these companies, measuring isn't something they do after a win to write it up later. It's how they find the wins in the first place, and how they catch the failures early enough to kill them.
The missing middle
The missing middle
So why is the value so hard to count? Because nobody connects the dots. A company can see what it spends, and it can see its sales at the end of the quarter, but nothing links the two. The steps in between are where the answer actually lives: whether people use an AI feature at all, and whether it helps them in the moment that matters. Those are exactly the steps that go unmeasured.
The chain most companies measure only at the ends
A working model of the chain. Most teams record spend and, sometimes, a far-off result, then guess at everything in between. Meilynx's analysis.
Close that gap and the abstraction goes away. The work turns into plumbing instead of philosophy. You connect each AI action to the event that follows it (a suggestion to the click it earned, a recommendation to the purchase it produced), then put a price and a clock on the result. "We think the assistant helps" turns into "a shortlisted candidate costs six cents and lands in forty-five seconds." That's no longer an opinion about AI. It's an account of what it does.
Buy, don't build
Buy the instruments
Companies have already figured out how to get their AI: buy it, don't build it. In a single year, internally built solutions fell from nearly half of all deployments to under a quarter. Roughly three-quarters are now bought, and most of some $37 billion went to outside platforms. The math is hard to argue with. A custom build runs from hundreds of thousands of dollars into the millions, plus maintenance, while a bought platform starts in the low thousands.
Enterprises are buying AI, not building it
Share of enterprise AI solutions, built vs. purchased. Source: Menlo Ventures, 2025: The State of Generative AI in the Enterprise.
The same logic applies to the layer that measures AI. Outcome tracking and governance are exactly the kind of horizontal tooling you'd want to buy rather than rebuild in-house. The instrumentation barely changes whether you're scoring a coding assistant or a customer-facing recommendation, so every team that builds it from scratch ends up paying twice for the same plumbing.
The skepticism about AI is fair, but it's aimed in the wrong direction. The companies that come out ahead won't be the ones that spent the most. They'll be the ones that can finally say, in dollars and seconds, what the spending actually bought.