Ask a chief executive what artificial intelligence is going to do to headcount at their company over the next three years. Then ask the people who work there the same question. You get answers pointing in opposite directions. Somebody ran that comparison properly this year, at scale, and published what came back.

Nearly six thousand CFOs, CEOs and senior executives across the United States, the United Kingdom, Germany and Australia were surveyed in January, and the National Bureau of Economic Research published the results in February.1 Those executives expect AI to cut employment at their own firms by about seven tenths of one percent over three years. In American firms the figure is 1.2 percent.1 The employees at those same companies expect AI to raise employment by roughly half a percent.1 The authors do the arithmetic on the executive side: across the four countries it comes to something like 1.75 million fewer jobs by 2028 at firms that already exist.1

Two groups inside the same buildings, looking at the same three years, forecasting in opposite directions. Only one of them writes the hiring plan.

The forecast is running ahead of the experience

What makes the executives' number strange is not that it disagrees with the employees'. It is where it came from. The same survey asked those executives what AI had actually done at their own companies over the previous three years. More than ninety percent reported no effect on employment. Eighty-nine percent reported none on labor productivity.1

Their own contact with the tools is thin as well. Around seventy percent of the firms use AI in some form. More than two thirds of the executives say they use it regularly, and regularly turns out to mean about an hour and a half a week. A quarter of them do not use it at all.1

Set those beside each other. A person who has spent ninety minutes a week with a tool, at a company that has logged nothing from it across three years, is confident enough to model a headcount reduction on the strength of it. That confidence came from somewhere and it did not come from the tool in their hands. It came from the demo that ran without a hitch, the analyst slide with the steep line, a competitor's announcement, the uneasy sense that everyone else has worked out something you have not. The forecast is borrowed. It was assembled out of other people's claims about a future none of them have reached either.

A forecast is only as good as the forecaster's contact with the thing being forecast. Ninety minutes a week is not contact. It is a rumor you have decided to believe.

Solow's rhyme

None of which means the gains are imaginary. In July 1987 Robert Solow reviewed a book about manufacturing and left behind a line that long outlived the review, which was that you could see the computer age everywhere but in the productivity statistics.2 He was right for years, and then he stopped being right, because the payoff did eventually arrive. What is easy to forget is why it took so long. The gains showed up once companies rebuilt how the work was done around the machine, not in the earlier stretch when they bought the machine and waited. The lag was organizational.

Read the present gap that way and the executives are not wrong that something is coming. They are wrong about what their own three years of nothing is evidence of. It is not evidence that the technology fails. It is evidence that nobody has yet done the work that makes it pay, which is a different problem, with a different fix, on a different clock.

What a forecast turns into

A prediction made in a conference room does not stay in the conference room. Seven tenths of one percent reads as academic right up to the moment it is spread across a real company and becomes a specific set of choices. A requisition that never opens. A departure nobody backfills. A graduate intake trimmed because trimming looks prudent. The people making those calls are steering by a number they have never personally tested, and the people who absorb the calls usually heard a warmer story on the way in, the one about AI freeing them up for higher-value work. The employees in this survey still expect employment to rise. When the openings stop appearing, they find out which forecast leadership actually believed.

I wrote last month about AI tools dropped into unchanged work with nobody answerable for the result, and I am not going to run that argument again here. This is the same failure one storey up. What went unowned this time is the forecast itself. Somebody produced the number, the number went into a plan, the plan shaped a budget, and at no point was it made to survive contact with the work it claims to describe.

A cheap test

The usual diagnostic asks whether people are using the tools. Adoption dashboards, license counts, weekly actives. Those measure who opened something. The more useful question is the one people would rather not ask out loud. Before a firm's AI forecast hardens into next year's hiring plan, find out who in the room has rebuilt one of their own workflows with the tool from start to finish, and then lived with the result long enough to know what it actually changed.

The leaders who have done that forecast differently. They are slower to promise, more specific about which tasks moved and which did not, more willing to say the timeline runs in quarters of redesign rather than in the speed of a rollout. Their numbers come in smaller and they hold up better. The ones still forecasting from the demo are the likeliest to be surprised, and the likeliest to have somebody else carry the cost of the surprise.

So here is what I would put to any executive team about to sign off on an AI-driven workforce plan. If the person who built the productivity number has spent ninety minutes a week with the tool and can name no effect it has had on the business so far, what is that number a measurement of? Everyone else in your building is currently expecting the opposite of what you have forecast. One of you is going to be wrong, and it will not be the same people who pay for it.