$2.5 Trillion Spent, Almost Nothing to Show For It

6,000 executives were asked what AI actually changed at their company. Nine out of ten said: nothing measurable.

Welcome Back,

A new National Bureau of Economic Research working paper surveyed nearly 6,000 CEOs, CFOs, and senior executives across the United States, United Kingdom, Germany, and Australia, and the headline finding is genuinely striking, close to 90% of firms reported no measurable AI impact on employment or productivity over the past three years. That's not a small, cherry-picked sample either, this spans four major developed economies and thousands of the people actually making budget decisions.

The gap between that finding and the current AI investment climate is hard to overstate. Global AI spending is tracking toward roughly $2.5 trillion in 2026, and big tech capital expenditure alone is climbing toward $800 billion. Companies like Salesforce, Amazon, and Klarna have all cited AI in layoff decisions this year. And yet, according to this survey, nine times out of ten, the companies actually running the technology can't point to a single productivity number that moved because of it.

Today we look at the specific numbers behind that 90% figure, the surprisingly small amount of time executives themselves actually spend using AI, why economists are reaching for a 40-year-old paradox to explain what's happening, and the one factor the data shows genuinely does move the needle.

📉 The Numbers Behind the Headline

The paper, titled "Firm Data on AI," comes from researchers Ivan Yotzov, Jose Maria Barrero, Nicholas Bloom, Philip Bunn, Steven J. Davis, and co-authors, revised in March 2026. More than 90% of surveyed firms reported no effect on employment. Eighty-nine percent said the same about labor productivity, measured specifically as sales per employee, a concrete, comparable metric rather than a vague self-assessment.

What makes this genuinely puzzling rather than simply a story about slow AI adoption is that adoption itself is not the problem, 69% of the firms surveyed are actively using AI. It just isn't showing up where boards usually look first. Usage is concentrated in text generation using large language models, visual content creation, and data processing using machine learning, real, active use cases, not shelfware nobody touches.

That 1987 quote is doing real work in how this story is being covered. Solow made that observation about computers broadly during an earlier wave of technology investment that also failed to show up in aggregate productivity data for years, and economists are explicitly drawing the parallel to AI now, not as a throwaway reference, but as a genuine historical precedent worth taking seriously.

Widespread adoption and measurable impact turned out to be two very different things in this survey.

⏱️ 1.5 Hours a Week

Here's a detail that goes a long way toward explaining the whole survey, about two-thirds of executives said they regularly use AI themselves, but the average use was only 1.5 hours per week. A quarter of the executives surveyed said they don't use it at work at all.

That distinction, individual time savings versus organization-wide productivity, is probably the single most important thing to hold onto from this entire survey. The two things get conflated constantly in AI coverage, and this data is a useful, concrete reminder that they're measuring genuinely different phenomena.

📊 What Executives Actually Expect Next

Despite reporting almost no measurable impact so far, the same executives surveyed still hold genuinely substantial expectations for AI's near future. They forecast AI will increase productivity by 1.4% and output by 0.8% over the next three years, alongside a projected 0.7% cut to employment over that same window.

There's a genuinely interesting mismatch buried in that forecast, too. Employees themselves, separately surveyed, actually reported an expectation of a 0.5% increase in employment over the same period, the opposite direction from what their own executives are projecting. That's a real disconnect between how leadership and workforce see the same near-term future.

Executives and employees hold genuinely different expectations about where AI-driven hiring is headed next.

🏭 The One Study That Explains the Lag

MIT Sloan research offers one of the more plausible explanations for why the gains haven't shown up yet. Research by Kristina McElheran, Mu-Jeung Yang, Zachary Kroff, and Erik Brynjolfsson, focused on US manufacturers, found that companies adopting industrial AI often took a short-term productivity hit before any gains showed up later.

That's a genuinely useful piece of context, it suggests the "no measurable impact" finding might partly reflect a real, temporary implementation dip, companies restructuring workflows around a new technology, absorbing training costs, and working through integration friction, rather than proof that AI simply doesn't work. Whether that dip fully explains a finding this large and this consistent across four countries is a separate, harder question.

Despite the flat productivity numbers, real companies are still citing AI in real layoff decisions this year, Salesforce, Amazon, and Klarna among them. That's worth sitting with directly, layoffs are happening in the name of AI efficiency at the same time as a large-scale survey finds no measurable productivity gain to justify them, at least not yet.

🎯 What Actually Moves the Needle

Buried inside the same dataset is a genuinely actionable finding, one specific investment shows up as making a real, measurable difference. Investment in workforce training amplifies AI productivity benefits by 5.9 percentage points, a substantial effect given how close to zero most of the other productivity numbers in this survey land.

That size gap is worth flagging too, medium and large firms experience substantially stronger AI productivity gains than smaller companies, which raises genuine concerns about widening competitive gaps between companies with the resources to invest properly in training and integration, and smaller firms that adopt the same tools without the same supporting infrastructure around them.

Workforce training investment was the single factor in this survey with a genuinely measurable positive effect.

🧠 AI Spotlight Analysis

There's a real tension in this story worth naming directly, a Financial Times analysis of S&P 500 earnings calls found 374 companies mentioning AI, most describing its implementation as entirely positive. That's the story companies tell publicly. Privately, in a rigorous academic survey, the same broad population of firms reports almost no measurable impact at all. Both things are apparently true simultaneously, which says something genuinely interesting about the gap between earnings-call optimism and what actually shows up in a company's numbers.

This isn't necessarily evidence that AI doesn't work, the historical Solow paradox parallel is a real, defensible explanation, and the MIT manufacturing research showing a temporary productivity dip before gains materialize adds real weight to a "too early to tell" reading. But it is genuinely strong evidence against the more breathless version of the AI transformation narrative, the one where adoption alone is assumed to translate quickly and automatically into measurable business results.

The workforce-training finding is probably the most useful, actionable takeaway buried in all of this. If there's one lever this data actually shows moving the needle, it's not more compute or a bigger model, it's investing in the people who have to actually use the tool well enough for its benefits to show up in the numbers that matter.

🔗 Sources and Further Reading

Thanks for reading AI Spotlight.

Our mission is simple: deliver clear, trustworthy, and actionable AI insights that help professionals stay ahead without the hype.