Economists at the Federal Reserve Bank of St. Louis just finished one of the largest studies yet on how companies talk about AI versus what AI actually does for their output. They read close to 490,000 earnings call transcripts from more than 5,000 public companies, covering the years 2000 through 2025. The goal was simple: track every sentence about productivity and see how many mention AI, and whether the tone is about gains already happened or gains still coming.
The pattern is clear. Before ChatGPT launched in late 2022, AI barely came up in productivity discussions, but that changed fast through 2023, leveled off in 2024, then jumped again through 2025 until AI was part of nearly 15% of all productivity comments on earnings calls. Yet when the researchers looked closer, about 95% of those AI comments describe gains executives expect to get later, not gains they have already banked.
Meanwhile, the actual productivity data for the economy barely moved. A broad measure that adjusts for how hard factories and offices are actually working grew by just 0.07% over the year ending in early 2026, which is close to nothing. So the gap between what executives say on calls and what shows up in national statistics is wide open right now.
One of the study's authors compared this to how electricity changed factories a century ago. Businesses did not get faster the day power lines were installed; it took decades to redesign factory floors, retrain workers, and rebuild how work actually got done before electricity showed up in the productivity numbers. His argument is that AI may be following the same slow path, with a real payoff that is simply delayed.
There is another explanation worth taking seriously too. When AI makes something cheaper to produce, the price of that output often falls at the same time, and the two effects can cancel out in the statistics even if a company is genuinely more efficient. That means the official numbers might understate what AI is doing, not just because it takes time, but because of how gains and falling prices interact.
Inside companies, the picture is uneven in its own way. A recent survey found that most senior executives believe their company's job structures are ready for AI, but only a small share of HR leaders agree. HR teams sit closer to the actual work, so they see the gap between what leadership expects and what jobs can realistically absorb right now.
That mismatch already shows up in hiring. Close to a third of HR professionals surveyed said their companies are now hiring fewer junior employees and more people at the middle level, because AI is doing tasks that used to train new workers from the ground up. If that keeps happening while the productivity payoff is still mostly a promise, companies are moving ahead of the proof, not behind it.
For anyone managing a workforce or a budget, the number worth watching is not what executives say AI will do on the next earnings call. It is whether actual output per worker moves at all over the next year, because that is the only number that settles the argument.