This year, there’s a figure that’s being discussed in boardrooms and during earnings calls that no one really wants to say aloud. Approximately 90% of businesses actively utilizing AI claim the technology has had no discernible impact on their productivity over the last three years, according to a National Bureau of Economic Research survey of nearly six thousand executives. Not a minor impact. Not one.
When compared to everything else coming out of Silicon Valley at the moment, that figure is unsettling. According to Anthropic, over 80% of the code integrated into its own production systems is now written by Claude Code. What used to take a week can now be completed in a day, according to individual engineers. Iren Azra Zou, a software developer at a trucking logistics startup, has discussed how AI tools significantly reduce her task time. These are not fictional tales. It’s true that people are faster.
How, then, do both become true simultaneously? It’s possible that timing is the primary cause of the gap. According to a survey conducted by economists at the Federal Reserve banks in Atlanta and Richmond, CFOs stated that AI had increased their productivity roughly three times more than their actual revenue figures. This suggests that operational improvements come first, followed by financial gains, if any. Economists believe that this is not the end of the story, but rather the awkward beginning of a much longer one.
However, there is another, less comfortable explanation that relates to what transpires following an individual’s victory. At the task level, Gupta-style efficiency gains are genuine. When businesses attempt to spread that benefit over a thousand workers and refer to it as a transformation, problems arise. This has been dubbed the “gen AI paradox” by a McKinsey partner: impressive pilot results that just won’t translate into corporate numbers. It’s one thing to get people to use a tool. Redesigning an entire organization’s workflow around that tool is a completely different matter, and most businesses haven’t done it.
As this develops, it’s difficult to ignore the analogy to the Solow Paradox, the old economist’s joke that computers are present everywhere but in productivity figures. Before the IT boom eventually appeared in the data, that gap persisted for years in the 1980s and 1990s. Though this time the spending is larger, faster, and much more visible to investors who anticipated results sooner, AI may be following the same slow, frustrating arc.
Layoffs are another messier wrinkle. Eighty percent of companies testing AI had eliminated jobs, according to a Gartner study of 350 global executives. However, there was no correlation between workforce reduction and actual returns. Businesses with the highest return on investment weren’t the ones making the biggest cuts; rather, they were the ones utilizing AI to increase the capabilities of their current workforce rather than just reduce costs. Many executives still appear to be overlooking this distinction, which appears to be more important than nearly everything else in the data.
None of this means the AI buildout is a mirage. S&P 500 profit margins hit record highs in early 2026 even with oil near $98 a barrel, an environment that has historically squeezed margins, not expanded them. Something is working. It’s just not working evenly, and it’s not working as fast as the spending implies it should.
Investors seem to believe the payoff is coming. The CFOs surveyed expect their productivity gains to roughly triple in 2026. Whether that’s optimism or evidence is still genuinely unclear, and probably won’t be settled until the next earnings season, or the one after that.
