Why the evidence is small and the fear is large

The public conversation about AI and jobs runs on two kinds of numbers, and most people, including most executives, do not distinguish them. One kind measures what could happen. The other measures what has. The gap between them is where a Chief Workforce Officer has to live.

The first kind is exposure. In 2023 researchers at OpenAI and the University of Pennsylvania estimated that around 80 percent of the American workforce had at least a tenth of their tasks exposed to large language models, and about a fifth had half or more. The International Monetary Fund in 2024 put roughly 40 percent of global employment in exposed occupations, around 60 percent in advanced economies. The International Labour Organization's 2025 update finds one worker in four worldwide in an occupation with some exposure. These are serious estimates by serious people, and they measure the same thing: whether the tasks in an occupation are the kind a model can do. They do not measure whether anyone has deployed a model to do them, let alone whether a person lost work as a result.

The second kind is realised effect, and it is much smaller. Humlum and Vestergaard, with Danish register data covering tens of thousands of workers, find no significant effect of generative AI on earnings or hours through 2024, with time savings around three percent on average. The field experiments find real productivity gains, 14 percent in customer support, 26 percent for developers, 40 percent faster writing in Noy and Zhang's study, but those are gains in bounded tasks with human review built in, not job losses. The one finding that looks like displacement, the Stanford decline among young workers in exposed occupations, is confined to one layer and has not yet been replicated.

Daron Acemoglu, in a 2024 analysis later published in Economic Policy, worked through what the exposure numbers imply for aggregate productivity if one is careful about which exposed tasks can actually be done more cheaply, and arrived at a gain of under one percentage point of total factor productivity over a decade. That is not nothing, but it is not a revolution, and it is a useful corrective to the belief that exposure translates into transformation.

Why does the distinction matter for the organisation? Because an executive who reads exposure as destiny will make one kind of error, cutting the entry layer in anticipation of an automation that may arrive slowly or in a different form. An executive who reads the small realised effects as reassurance will make the opposite error, missing the concentrated change at the layer where it is actually happening. Both errors come from reading the wrong number, and neither is corrected by the vendors, whose own research, including a widely quoted survey finding that 86 percent of chief human resources officers see integrating agents as central to their role, is sentiment data from an interested party.

There is also a cautionary tale about the evidence itself. In 2025 the Massachusetts Institute of Technology withdrew a widely circulated paper claiming large productivity gains from AI in materials science, after concluding that its data could not be trusted. The paper had been cited by senior figures before it was withdrawn. The lesson is not that AI research is unreliable. It is that someone in the organisation has to read it with care, distinguish exposure from adoption from substitution, note what is replicated and what is not, and resist both the vendor's slide and the headline's fear.

That reading is a workforce responsibility, not a technology one, because what is at stake is how many people to hire, train and keep. It belongs to the Chief Workforce Officer, who should be the most careful reader of this literature in the building, and who should be the one telling the board which numbers measure the future and which measure the present.