Three years into the AI boom, the productivity statistics still refuse to show it. New research from the Federal Reserve Bank of St. Louis, which scanned roughly 490,000 earnings-call transcripts from more than 5,000 US public companies between 2000 and 2025, confirms what the official data has said all along: artificial intelligence has not yet produced a measurable bump in aggregate output per worker. Executives talk about it constantly, but the talk is almost entirely promissory. About 95 percent of the AI-related productivity comments describe gains they expect in the future, not gains already banked, and that share has barely moved since 2023.
The conventional reading is patience. Serdar Ozkan, one of the paper's authors, reaches for Robert Solow's line from the 1980s that you could see the computer age everywhere except in the productivity statistics, and points to electrification, which took several decades to show up in the numbers after the first power stations were built. Transformative technologies arrive slowly in the data because firms have to rewire how they work around them, not just plug them in. By this account AI is early, not overrated, and the payoff is a matter of when.
But Ozkan floats a second, stranger possibility, and it is the one worth sitting with. The gains may already be here and structurally invisible, because AI is destroying the value of the very things it makes abundant. The logic is almost arithmetical. When a task becomes radically cheaper to perform, its output becomes cheaper too. Anyone can now generate marketing copy, an animation, or a passable news article with a keystroke. But if everyone can, none of it commands what it used to. The work got easier and the product got cheaper, and somewhere in that trade a real gain slipped out of the statistics without ever registering as a loss.
This is the abundance problem, and it scrambles the usual optimism. Productivity is measured in value, not effort, so a technology that floods the world with cheap output can lift how much we make while flattening what it is worth. The two can cancel. That does not mean the work was pointless; a firm that fires half its copywriters and keeps the same output is genuinely more efficient. It means the surplus shows up as lower prices and thinner margins rather than as a visible jump in measured productivity, and it accrues to whoever captures the savings, which may be customers rather than the firms doing the boasting.
Both explanations can hold at once, which is what makes the finding unsettling rather than merely disappointing. The J-curve story says wait; the abundance story says some of the reward may never appear in the headline figure at all. What the earnings calls reveal is a peculiar gap between sentiment and evidence: executives describe AI as raising productivity in 95 percent of their comments, against 75 percent for everything else, an unusually uniform bullishness for a technology whose effects the data still cannot find. Someone is going to be proven right, but the honest position today is that we are cheering a revolution we have not yet learned how to count.