What automation did last time.

The confident claim that this technology, finally, is the one that ends work is one of the most reliable products of industrial civilization. It has been manufactured nonstop since at least 1811, in the same packaging every time: a genuinely impressive machine, then a wave of visible task destruction. Next comes a forecast stretching that destruction to whole occupations — and a quiet failure of the forecast over the following decades. The machines were real. The task destruction was real. The mass unemployment never arrived.

This matters now because the AI debate is running on forecasts again, and forecasts should be graded against base rates. The historical record of automation is the base rate. It does not prove that AI will behave like the tractor or the ATM. History rhymes rather than repeats, and the last section of this article takes the differences seriously. But a prediction about AI and work must explain why two centuries of identical predictions failed, and why this one escapes the pattern. Until it does that, it has not earned confidence.

What follows is the record: the recurring anxiety, the task arithmetic that keeps resolving it, the two case studies every forecaster should know by heart, and the one factor — training capacity — that history says decides who actually benefits.

Two hundred years of the end of work

Start with the panic itself, because its persistence is data. Economic historians have surveyed technological anxiety from the Luddite risings onward, and they find the same worries recurring in nearly every generation (Mokyr, Vickers & Ziebarth, 2015). The worries: machines will permanently displace labour at mass scale; the work that remains will be degraded or meaningless; and — in a strand that alternates with the first two — invention itself is running out.

English textile workers smashed frames in 1811. In 1930, at the start of a century of unmatched job creation, Keynes coined “technological unemployment” and predicted a fifteen-hour work week. In the early 1960s, alarm about automation was serious enough that a US presidential commission was convened to study it. It reported, in effect, that the problem had been overstated. Each panic felt uniquely justified at the time, because each was anchored to a technology that really did change the world.

The historians’ point is not that the worry was foolish. Short-run disruption was real, and for some groups it was brutal. Displaced handloom weavers did not glide into the factory jobs that would later employ their grandchildren. In the early industrial decades, living standards for many workers stagnated even as output soared (Mokyr, Vickers & Ziebarth, 2015).

The point is that the forecasts failed in a consistent direction. They overweighted the destruction they could see, and underweighted two things they could not: the re-bundling of existing jobs around new tasks, and the creation of whole categories of work that did not yet have names. A forecaster in 1900 could see the horse being displaced; almost none predicted the car mechanic, the traffic engineer, or the motel. That asymmetry — visible losses, invisible gains — is the recurring error, and it is worth understanding why it keeps happening.

Jobs are bundles, not lumps

The forecasts keep failing the same way because they treat jobs as the unit that automation acts on. The economics says otherwise: the unit is the task. A job is a bundle of tasks, and machines absorb the bundle selectively (Autor, Levy & Murnane, 2003). Some tasks are routine and can be written as rules; others need judgment, dexterity, persuasion, or tacit knowledge nobody can fully write down. The landmark task-level analysis of computerization showed exactly this signature in the data: computers replaced routine mental and manual tasks — the ones that follow explicit rules — while complementing non-routine analytic and interactive work. Occupations did not disappear so much as tilt: the routine share shrank, the judgment share grew.

Once you see jobs as bundles, the survival of employment through two centuries of automation stops being mysterious. When a machine takes over part of a job, three things happen to the rest of it. First, the remaining human tasks often become more valuable, because they are now the scarce input — the analyst freed from arithmetic does more analysis, and the analysis is worth more. Second, automation lowers the cost of the output. When demand for that output is elastic, cheaper output means more output — which can mean more employment even in the automating occupation itself. Third, workers move: the two-century fall of farm employment from most of the workforce to a rounding error was matched by growth in occupations no farmhand foresaw (Autor, 2015).

Autor’s synthesis added a further observation about why the frontier recedes more slowly than demos suggest. Many economically valuable tasks rest on tacit knowledge — we know more than we can tell. Tasks that resist explicit description resist automation long after nearby tasks fall. None of this is a law of nature. It is an accounting identity plus a set of empirical regularities. But the regularities have held across every major automation wave for which we have data, which is why the burden of proof sits with forecasts that ignore them.

The ATM paradox

If the task framework has a mascot, it is the bank teller. The automated teller machine was engineered, explicitly and by name, to automate the core task of a specific occupation. It spread through the United States from the 1970s onward, reaching hundreds of thousands of installed machines — and over the same decades, the number of human bank tellers did not collapse. It rose, modestly but unmistakably, into the 2000s (Bessen, 2015).

The mechanism is the task arithmetic working in the open. ATMs absorbed cash-handling, which meant a branch could run with roughly a third fewer tellers than before. Cheaper branches changed the banks’ calculus. Urban branch counts rose by roughly forty percent as banks competed on convenience and presence — and more branches, each with fewer tellers, netted out to more tellers overall (Bessen, 2015).

Meanwhile the job itself re-bundled. With the cash drawer automated, tellers shifted toward relationship work: resolving problems, advising customers, selling products that machines cannot sell. Detailed occupation-level analysis of computer automation finds this is the general pattern, not a curiosity. Occupations that used computers heavily saw tasks change and skill demands rise. But employment in those occupations tended to grow faster, not shrink, as partial automation raised the value of the remaining human work (Bessen, 2016).

Two honest footnotes belong on the story. The reprieve was not permanent. As mobile banking absorbed still more of the bundle, teller numbers have drifted down since the 2010s — task re-bundling buys decades, not immortality. And the new teller job demanded different skills than the old one; the cash-handler and the relationship banker share an occupation code but not a skill set. Both footnotes point the same direction. The interesting question about automation is almost never “will the occupation exist?” It is “what will it be made of, and who gets retrained into the new bundle?”

The dynamo delay

The teller story explains why occupations survive automation. The second canonical case explains why the payoff takes so long to arrive — and it is the one that should most discipline expectations about AI timelines. In 1990, computers were visibly everywhere while measured productivity growth was flat. Confronting that puzzle, the economic historian Paul David reached for a precedent: the electric dynamo (David, 1990).

Electric power was commercially available from the 1880s. The aggregate productivity surge from electrification did not show up until the 1920s — a lag of roughly forty years. The delay was not about the technology maturing; it was about everything around the technology. Factories of the steam age were built vertically. Their machines were slaved by belts and shafts to a central power source, and their layouts were dictated by the physics of shaft transmission.

The first factory electrifications simply swapped a motor in for the steam engine and kept the layout — “group drive” — and gained almost nothing. The gains came when engineers grasped what unit drive made possible: a small motor on each machine, factories laid out single-storey around the flow of materials rather than the geometry of the drive shaft, overhead cranes, easy rearranging, natural light. Getting there meant scrapping capital, redesigning buildings, rethinking supervision, and retraining a workforce. A generation of complementary investment passed before the headline technology paid (David, 1990).

The lesson David drew for computers applies with at least equal force to AI. The productivity effect of a general-purpose technology is gated not by what it can do in a demo, but by the pace of organizational redesign around it. A model that drafts a document in seconds changes little if the workflow around the document — review, accountability, compliance, the definition of the job itself — is still built for the old process. Group drive is the default first deployment of every general-purpose technology, AI included. Bolt it onto existing workflows, harvest almost nothing, and conclude too early that the technology is hype. The historical base rate says transitions are slower than demos suggest — and that the winners are the firms doing the unglamorous redesign work during the lag.

The two-ledger accounting

The task lens and the delay explain single episodes. To grade forecasts you also need the aggregate bookkeeping, and modern labour economics supplies it. Automation runs two ledgers at once. The displacement effect removes tasks from workers. The reinstatement effect creates new tasks in which labour has a comparative advantage — tasks that often did not exist before the technology (Acemoglu & Restrepo, 2019). Net employment and the labour share of income depend on the race between the two, plus the productivity effect of cheaper output raising demand across the economy.

History is full of the reinstatement ledger being written in real time, almost always unforecast. Mechanized farming displaced field labour, and the vast growth of clerical and office work that industrial coordination demanded followed. The computer destroyed typing pools and ledger clerks, and reinstated software developers, systems administrators, database designers — job titles that simply did not exist a generation earlier.

The framework also carries a warning that separates it from complacency. The two ledgers need not balance, and nothing guarantees that reinstatement keeps pace. The same analysis finds that over roughly the last four decades, displacement sped up while reinstatement slowed relative to the postwar pattern — one candidate explanation for sluggish wage growth and a declining labour share (Acemoglu & Restrepo, 2019). The honest reading of the record is therefore not “it always works out”. It is “it has usually worked out, through a mechanism that is measurable, and the measurements say the balance can tilt”.

The race between education and technology

If reinstatement is not automatic, what decides whether workers end up on the winning ledger? The best long-run answer comes from a century of American wage data. It is uncomfortable for anyone hoping technology alone settles the question. Goldin and Katz framed the twentieth-century wage structure as a race between technology, which raises the demand for skills, and education, which raises the supply (Goldin & Katz, 2008).

For most of that century, America won the race by expanding schooling faster than technology expanded skill demands. The high-school movement put secondary education within reach of ordinary workers decades before other countries followed. The supply of skilled labour grew, and the skill premium — the wage gap between more- and less-educated workers — narrowed even as technology raced ahead. When schooling levels plateaued in the 1970s while skill-biased technology kept accelerating, the premium exploded and inequality widened (Goldin & Katz, 2008). Same technological forces, opposite outcomes for who gained — the difference was the training pipeline.

For the AI transition the transfer is direct. Whether the technology produces broadly shared gains or a winner-take-most premium is not a property of the models. It is a property of how fast the supply of complementary skills grows — which for working adults means employer-run reskilling far more than it means universities. Training capacity is the policy variable hiding inside every automation forecast.

Grading the modern forecasts

Which brings us to the most-cited automation forecast of the era. In a working paper circulated in 2013 and published in 2017, Frey and Osborne put about 47 percent of US employment in occupations at high risk of computerisation. The risk window: plausibly a decade or two (Frey & Osborne, 2017). The number escaped the paper and became folklore — “half of all jobs” — repeated in headlines, keynotes, and policy documents.

More than a decade on, the forecast can be graded, and the grade is instructive rather than merely embarrassing. There was no wave of occupation extinction. Through most of the following decade, rich-country unemployment sat near historic lows, and the occupations flagged at highest risk mostly kept employing millions.

The misses were the classic ones this article has catalogued. The method scored whole occupations rather than tasks, so it read “many tasks in this job are automatable” as “this job is at risk” — precisely the aggregation the teller case refutes. It estimated technical susceptibility, not economic adoption, skipping cost curves, demand elasticity, and the dynamo-style organizational lag — and it had no reinstatement ledger at all. To be fair to the authors, they said much of this themselves. The paper estimated what could be automated, not what would be, on no particular timetable (Frey & Osborne, 2017). The lesson is less about one paper than about a genre: a forecast that scores occupations instead of tasks, ignores adoption economics, and books no task creation is repeating a method that has already been graded against reality once — and failed.

What the evidence doesn’t show

Base rates are for disciplining forecasts, not for dismissing them. Several genuine limits keep the historical record from settling the AI question on its own:

• History rhymes; it does not repeat. Every prior automation wave absorbed relatively narrow families of tasks — mechanical, then routine and codifiable. A technology that reaches into non-routine cognitive work has far thinner precedent, and the tacit-task refuge that protected past occupations (Autor, 2015) may be smaller this time. The base rate is a prior, not a verdict.
• The speed may differ. The dynamo delay reflected physical capital: buildings, shafts, machines (David, 1990). Software diffuses faster than factories are rebuilt. Organizational and institutional lag still applies — but forty years is a historical observation, not a law, and the lag could compress.
• Aggregate comfort was never individual comfort. Even in episodes where total employment held, displaced individuals suffered real and persistent earnings losses, and some never recovered their trajectory. The record’s reassurance is statistical; the costs were borne by particular people, which is exactly why transition support matters.
• Distributional pain was real and sometimes long. Early industrialization saw decades in which output soared while ordinary workers’ living standards stagnated (Mokyr, Vickers & Ziebarth, 2015), and the recent tilt toward displacement over reinstatement (Acemoglu & Restrepo, 2019) shows the ledgers can stay unbalanced for a long time.
• Occupation-level base rates hide within-occupation churn. “Tellers survived” is true at the occupation code and false at the skill profile — the surviving job demanded different capabilities (Bessen, 2016). Individuals experience the churn even when the category persists.
• Institutional context differs across countries. The race between education and technology played out differently wherever training systems differed (Goldin & Katz, 2008), and economies with strong retraining institutions have historically absorbed shocks with less scarring. Base rates drawn mostly from US data travel imperfectly.

Where the evidence stops

1. 1History rhymes; it does not repeat
2. 2The speed may differ
3. 3Aggregate comfort was never individual comfort
4. 4Distributional pain was real and sometimes long
5. 5Occupation-level base rates hide within-occupation churn
6. 6Institutional context differs across countries

© 2026 FUTURE PROOF™

What the record instructs

Taken together, the historical evidence converges on a short list of instructions for firms facing the AI transition — none of them exotic, all of them routinely ignored.

Watch tasks, not headlines. The unit of change is the task bundle (Autor, Levy & Murnane, 2003). A firm that keeps a live map of which tasks in which roles are being absorbed, augmented, or newly created will see its own teller paradoxes and dynamo delays coming. One that tracks occupation-level punditry will not. The practical form of this is a skills-and-tasks inventory that gets re-measured, not a one-off consulting deck.

Budget for redesign, not just licences. The dynamo lesson is that the technology’s price is the small part of its cost (David, 1990). The firms that harvested electrification early were the ones that paid for new factory layouts. The AI equivalent is workflow redesign, role redefinition, and process re-engineering — group drive is a choice, and it is the default choice. Expect the payoff to lag the deployment, and plan the interim measurement accordingly.

Start the reskilling before the displacement. The single strongest historical regularity about who benefits is the supply of complementary skills (Goldin & Katz, 2008). Waiting until a role is visibly obsolete to begin retraining repeats the pattern that made past transitions cruel for single workers even when aggregates held. The re-bundled teller learned relationship banking while still employed as a teller (Bessen, 2015); that sequencing — adjacent skills built from a currently valuable seat — is the humane and the economical order.

Track the reinstatement ledger deliberately. New task categories will appear around the technology (Acemoglu & Restrepo, 2019), mostly unannounced, inside existing job titles first. Firms that name them early — prompt evaluation, model supervision, AI-output review, whatever the local variants turn out to be — can route displaced capacity toward them. The other path is hiring cold from outside while cutting jobs inside — the visible signature of a firm that lost track of its own two ledgers.

None of this requires believing the optimistic reading of history. It requires only taking seriously the mechanism that has governed every prior wave: tasks re-bundle, payoffs lag, skills decide the distribution. The firms that act on the mechanism do better under nearly any scenario for the technology itself — which is what makes it a strategy rather than a forecast.