Industrialisation already ran this experiment, and we are misreading the result

Companies are removing people this year to capture a productivity gain that, on their own account, has not arrived.

Read that twice. Both halves are documented.

The Federal Reserve Bank of St. Louis ran about 490,000 earnings call transcripts from 5,198 publicly traded US firms to see how executives talk about productivity. Roughly 95 percent of the productivity sentences that mention AI point at future gains. Not gains booked. Gains expected. Over the same window, utilization-adjusted total factor productivity grew 0.07 percent in the four quarters to the first quarter of 2026.

Now the other half. TechCrunch keeps a running list of large technology companies that announced significant job cuts this year with AI as a stated factor. Those cuts are not expected. They happened, to real people, in named quarters.

The benefit is a forecast. The cost is a fact. Both turn up in the same announcement.

The reflex I cannot switch off is to ask whether the thing being claimed is the thing being measured. Here it is not. A company can tell investors AI will make it more productive next year, tell four hundred people AI has made them surplus this year, and nobody lines the two sentences up.

This is not dishonesty. It is a pattern with a long history, and that history is more useful than either story we normally tell about it.

What the dynamo actually cost

We tell ourselves two stories about technological unemployment. One says we have been here before and it worked out. The other says we have been here before and it did not. Both skip the part that matters, the interval.

Paul David went and looked. In 1990 he published The Dynamo and the Computer, at a moment when economists were staring at computers everywhere and productivity nowhere. Go back, he said, to the last general purpose engine that did this to us.

In 1899, electric lighting reached 3 percent of American homes and electric motors drove under 5 percent of factory mechanical power. It took two more decades for either to reach half. Factory electrification, David writes, "did not reach full fruition in its technical development nor have an impact on productivity growth in manufacturing before the early 1920s." Four decades after the first central power station opened for business. He notes, drily, that an observer in 1900 could fairly have said the dynamos were everywhere but in the productivity statistics.

So why the wait? Not because the motor was slow to arrive. Because of what factory owners did with it once it had.

The first wave used group drive. You bought a motor, bolted it to the overhead shafting, and it turned the same belts the steam engine used to turn. It worked. It saved a little fuel. It moved the productivity needle almost not at all, because the belts stayed, the shafting stayed, and the heavy bracing holding it up stayed. David's phrase is that the old equipment "remained in place as available capacity", pushing the capital to output ratio up and total factor productivity down.

The gain came from unit drive, where every machine got its own motor. And the point of unit drive was never the motor. Once no building needed bracing for overhead shafting, you could build light instead of heavy. One floor instead of four. A layout arranged around the flow of materials instead of the geometry of a shaft.

The productivity did not come from installing the motor. It came from redesigning the factory around it. That took a generation of architects and engineers learning how. David reckons roughly half the five percentage point jump in American manufacturing productivity between 1919 and 1929 traces to the growth in secondary electric motor capacity in that decade.

What the weavers paid while everyone waited

The interval is not free. Somebody pays for it, and rarely the person who collects.

Robert Allen's Engels' Pause gives the British numbers in two lines. Between 1780 and 1840, output per worker rose 46 percent and the real wage index rose 12 percent. Between 1840 and 1900, output per worker rose 90 percent and the real wage 123 percent. The aggregate gain was real. It was enormous. It reached working people sixty years late.

This is contested, and it should be. Allen notes Gregory Clark's case that workers did better than the standard wage series suggest, and calls it "exciting revisionism" that does not convince him. Historians are still arguing. What nobody argues about is what happened at the sharp end.

A single power loom could out-produce ten to twenty handweavers at home, and the machines were big enough to need a factory building, which took the cottage trade off the table. Handloom weavers in English cotton averaged 240 pence a week in 1806. By 1820 they were under 100. In two Lancashire towns family earnings halved in five years from 1814, a figure Acemoglu and Johnson recount from their 2024 review.

The figure I keep coming back to is not a weaver. It is David Ricardo, the most respected economist alive, telling the Commons in 1819 that "machinery did not lessen the demand for labour." Then he watched the power looms. In the 1821 edition of his Principles he inserted a new chapter saying the opposite: "The same cause which may increase the net revenue of the country, may at the same time render the population redundant, and deteriorate the condition of the labourer."

He was nearly fifty, wealthy, at the top of his profession. He changed his mind in public anyway. Think about how rare that still is.

In fairness to the people making the cuts

Several of the executives doing this are more careful than the coverage suggests. Atlassian's chief executive put it squarely: "Our approach is not 'AI replaces people.' But it would be disingenuous to pretend AI doesn't change the mix of skills we need or the number of roles required in certain areas. It does." Microsoft went further and said the eliminated roles were "not being replaced by AI" at all.

That is a defensible position, and it is not the position the market hears.

There is a simpler reading of the numbers too. On figures TechCrunch reports, Amazon, Oracle, Meta and Microsoft account for almost 50,000 of the nearly 140,000 tech jobs cut since January, while pouring hundreds of billions into AI data centre buildouts. Payroll becomes plant. That is capital reallocation, a different animal from a machine doing the job more cheaply than a person.

The retrofit that is being sold as the rebuild

Cutting headcount while leaving the operating model exactly where it was is group drive. A motor bolted to the old shafting, presented as a new factory.

Take Cloudflare, which cut about 20 percent of its workforce, some 1,100 people, in the quarter its revenue hit 639.8 million dollars, up 34 percent and the best in its history. Its chief executive wrote that "the vast majority of those we laid off last week were measurers", naming middle management, finance, legal, internal auditing and revenue recognition. Read it generously and that is a real claim about organisational design. Read it harder and the question is whether the coordination work disappeared or simply stopped having an owner. Only one of those shows up as productivity next year.

Take a public administration that procures the tooling, trains its people, then takes twelve percent out of its establishment because the business case said the tooling would absorb the load. The approval chain has not moved. The file still crosses the same four desks. The volume coming in has not fallen. What arrives a year later is the same queue with fewer people beside it.

Or the professional services firm that cuts its junior intake by a third because the tooling drafts, narrowing at the base the pyramid that makes next decade's partners.

The question in all three is not whether the cut was justified. It is whether anything about how the work is arranged actually changed. If nothing did, the organisation has bought the entire cost of the transition and left the gain exactly where its own earnings call put it. In the future.

What the aggregate cannot see

Stanford's Digital Economy Lab tracks this monthly on payroll data. Take its August 2026 revision in the order it makes its points. First: "We do not see widespread, economy-wide job displacement associated with AI." Second: employment among 22 to 25 year olds in highly AI-exposed occupations sits about 19 percent below where it would be had it kept pace with the same age group in less exposed work. A year ago that gap was 15 percent. Experienced workers show nothing comparable. The adjustment runs through hiring that never happens rather than people being let go. These are descriptive patterns, the authors say plainly, and not causal estimates.

Hold that next to the weavers. The aggregate was fine. The aggregate was better than fine. One identifiable group was destroyed, and the recovery arrived in time for their children.

The people who are also the product

There is a second asymmetry underneath the first, and it turns my stomach a little.

The same Stanford revision finds the declines concentrated in occupations built on codified knowledge, the formal, documented, written down kind you can teach from a textbook. Employment held up, or rose, in work that runs on tacit knowledge, the sort you only get from practice, mentorship and being in the room.

That distinction is not neutral. Codified knowledge is knowledge somebody wrote down. The manuals, the templates, the resolved tickets, the annotated cases, the millions of ordinary competent documents produced by ordinary competent people doing their jobs. That corpus is what the models are good at, because that corpus is what they were built from.

So the same population turns up twice, in two different columns. Once as the source of the training material, collected and monetised with little conversation about it. Once as the cohort not being hired. That is an inference and not a finding. But the weaver kept his skill when the loom took his trade. Here the skill is copied into the machine first, and the job goes afterwards.

That is what industrialisation actually teaches, and it is neither comforting nor apocalyptic. The aggregate is a poor guide to the experience. The gain arrives late, and only where the work is rebuilt around the technology. The interval is paid for by specific people, rarely the people deciding.

What to ask before you cut

No framework here, and I would distrust anyone selling one. There is a posture, and it comes down to refusing to let the two tenses share a sentence.

If the gain is real, you can say where it lands. Which process. Which cycle time. Which unit cost. Measured how, by when. A gain nobody can locate in a process is not a gain, it is a hope with a headcount attached.

If you cannot name it yet, that is a perfectly good answer, and it is the argument for waiting before you remove the people. The record says the gain follows the redesign. It has never once run ahead of it.

And if the redesign has not happened, what you have is a bet rather than a harvest. Bets are legitimate. But a bet described upward as an efficiency and downward as an inevitability is neither of those.

Simon Johnson says it more bluntly than I would dare: "just because you have new miracle machines does not mean most people will benefit."

Ricardo needed two years and a new edition to get there. We have his working, we have the dynamo, and we have four decades of factory owners discovering that the motor was never the point. The question is not whether AI raises productivity. It probably will. The question is who pays for the gap, and whether anyone has told them.

Think about it.