Nearly 6,000 Executives Cannot Find the AI Productivity Boom

More than 90 percent. That is the share of senior executives who report no effect from AI on employment at their firm over the past three years.

The figure comes from an NBER working paper built on identical surveys run by the Atlanta Fed, the Bank of England, the Deutsche Bundesbank and Macquarie University between November 2025 and January 2026. Close to 6,000 CEOs, CFOs and senior finance managers answered. Eighty-nine percent of them also reported no effect on labour productivity.

What the AI productivity gains study actually measured

Adoption is not the missing piece. Sixty-nine percent of firms across the four countries report using AI in some form, rising to 78 percent in the United States and falling to 59 percent in Australia. So most of these firms are using the technology, and almost none of them can see it in their own numbers.

That gap is the finding. The surveys asked what firms had observed, not what they expected and not what a vendor had promised them, and the observed effect on both jobs and output sits close to zero.

The forecasts are more interesting than the results

The same executives expect AI to raise labour productivity by 1.4 percent and cut employment by 0.7 percent over the next three years. That nets out to roughly 0.8 percent extra output.

It is a real number and a modest one. It also sits a long way below what markets have been pricing into anything with AI attached to it. These people are not sceptics. They think the effect is coming, they just think it will be small.

Why this matters if you read AI coverage

Most claims about AI productivity come from vendors measuring their own product, or from lab studies on narrow tasks with willing participants. This is firm level, cross country, and answered by the people who would notice a headcount change before anyone else does. Valuations across the sector assume the gains arrive at scale and arrive soon. Three years in, at scale, they have not.

The spending is real and measurable in a way the returns are not yet, which is why the power draw of AI data centres keeps climbing regardless of what the surveys say.

Self reporting is the obvious limitation. A manager who has quietly stopped backfilling roles may never describe that as an AI effect. But if the 1.4 percent forecast turns out to be right, these surveys are the first place it should show up.