Late last year, research teams at three central banks and a university put a blunt question to nearly 6,000 senior executives across the United States, the United Kingdom, Germany and Australia: over the past three years, what has artificial intelligence actually done to your business? More than 90 percent answered that it had done nothing at all to employment. Eighty-nine percent said the same about labor productivity.
That is the headline result of Firm Data on AI, NBER Working Paper 34836, circulated in February 2026 and revised in March by a 13-author team including Stanford economist Nicholas Bloom, Steven J. Davis, Ivan Yotzov and Jose Maria Barrero. Six months on, it remains the most-cited number in the argument over whether the AI buildout is producing anything a firm can actually measure.
The instrument matters here. Identical questions were fielded between November 2025 and January 2026 by the Federal Reserve Bank of Atlanta, the Bank of England, the Deutsche Bundesbank and Macquarie University, drawing responses from CEOs, CFOs and senior finance managers. The survey did not ask whether AI is important. It asked executives to quantify its effect on their own firm, in percentage terms, over a defined window. Adoption was not the constraint: 69 percent of firms reported using AI, ranging from 78 percent in the US to 59 percent in Australia, with LLM text generation the most common application at 41 percent.
Among the minority who did report an effect, productivity tilted faintly positive, an average boost of 0.29 percent. According to Charter, which covered Bloom presenting the data in San Francisco in February, just five percent of executives said AI had reduced headcount, mostly by less than five percent, while four percent said it had increased headcount. Asked about the wave of layoffs announced under an AI banner, Bloom pointed instead at interest rates, tariffs and earnings pressure, and was direct: As of yet, there has not been a big effect.
The same executives expect a very different next three years: productivity up 1.4 percent, output up 0.8 percent, employment down 0.7 percent, roughly two-thirds of that through slower hiring rather than cuts. About 3,000 US employees surveyed in parallel expect the opposite, a 0.5 percent employment gain, and forecast less than half the productivity improvement their bosses do. There is a clear misalignment here across society, Bloom said.
Which sectors did report effects. A companion paper, NBER 34984 by Salome Baslandze, Zachary Edwards, John Graham, Ty McClure, Brent Meyer, Michael Sparks, Sonya Waddell and Daniel Weitz, surveyed roughly 750 executives, 603 through The CFO Survey run by the Atlanta Fed, Richmond Fed and Duke. Gains were real but narrow, concentrated in high-skilled services and especially finance. It also produced the cleanest statement of the puzzle: firms reported output per worker rising 1.8 percent from AI in 2025, but the gains implied by their own revenue and headcount figures were much smaller in every major industry. The authors invoke Robert Solow directly, noting current revenue-based gains remain well below the late-1990s IT surge.
Census data points the same way. The Business Trends and Outlook Survey AI supplement published in April found roughly 96 percent of AI-using firms reported no change in total employment over the prior six months. Adoption is real but shallow: about 20 percent of businesses reported AI use in early 2026 under a broadened definition, versus 30 percent among firms with 250 or more employees and 17 percent among firms under 20. Nearly four in ten information and professional services firms use AI; fewer than one in ten in agriculture, transportation, food services or construction do.
The gap between nothing and seven hundred billion
Four hyperscalers are on track to spend close to 700 billion dollars on AI infrastructure in 2026, roughly 70 percent above 2025. Reconciling that with nine-in-ten reporting no effect requires taking the survey design seriously.
Three things are happening at once. The capex is concentrated in a handful of firms building capacity, while the surveys sample the broad economy, where a typical adopter uses AI for marketing copy and customer service. The questions are also binned: a firm with a genuine but modest 1 or 2 percent efficiency gain plausibly lands in the no impact bucket, which makes 90 percent no effect compatible with a small, diffuse, positive average rather than zero. And the respondents are thin users, averaging about 1.5 hours of AI use per week, with 28 percent reporting none at all.
The strongest counterargument is that firm-level self-reports are simply the wrong instrument. Block announced roughly 4,000 job cuts this year that Jack Dorsey tied explicitly to AI efficiency; Amazon has told staff to expect a smaller corporate workforce. Mark Ma of the University of Pittsburgh, writing in Fortune on 22 August, argues the causation runs backwards from what managers assume: AI-linked layoffs damage employee sentiment, and employee sentiment, not management optimism, predicts whether gains reach the financials. Apollo chief economist Torsten Slok takes the other side, insisting there is zero evidence of job losses because of AI and that the buildout is stoking both employment and inflation. Both can hold. The aggregate is quiet while the distribution is violent.
The honest reading is that AI in 2026 looks like a general-purpose technology in its awkward middle, adopted faster than it is absorbed. Bloom offered his own caveat: these surveys cover existing firms, and the reallocation may run through new entrants no incumbent panel will capture.
Watch three things over the next two quarters. Whether the CFO Survey shows realized 2026 productivity converging on the 1.8 percent executives reported for 2025 or falling short again. Whether the routine clerical share of employment declines on the 2.19 percent path firms projected for 2028. And whether the BTOS employment question, stuck at 96 percent no change, finally moves. If none of the three budge by year-end, the gap between what executives are spending and what they are measuring becomes the story.
“As of yet, there has not been a big effect.”— Nicholas Bloom, Stanford economist and co-author, NBER Working Paper 34836