Between the 1980s and 2010, U.S. banks installed roughly 400,000 ATMs across the country, and bank teller employment did not collapse. It grew, from about 500,000 to nearly 600,000. The machines were doing the tellers' job, and the head count went up anyway.
That story is one of the cleaner versions of a pattern that runs through the economic record whenever a big new technology shows up, and it is the right starting point for thinking about AI and jobs.
Two Roads After a Leap
Every big technology leap presents companies with the same fork: use the new capability to do existing work with fewer people, or use it to do work that was not economically possible before. The first path is cost savings. The second is growth.
The cost-savings path has a ceiling, because you can only cut payroll down to zero in a given role and the saving stops the moment the last person is gone. The growth path has no obvious ceiling, because the new floor on what is affordable changes which markets exist in the first place. Banks could not have opened tens of thousands of new branches at the old per-branch cost. Once ATMs lowered that cost, the unit economics of a branch flipped, and the question stopped being how to cut staff and started being how many more locations to open.
More Tellers, Not Fewer
Economist James Bessen tracked the ATM case carefully. Between 1988 and 2004, the number of tellers required to run an average urban branch fell from about 20 to 13, which is the kind of efficiency gain that usually shows up as a headline about layoffs. Instead, banks used the savings to open more locations, urban branches grew by about 43%, and total teller employment kept climbing. The job changed shape along the way, with less cash handling and more product sales and relationship work, but the category got bigger, not smaller.
The cost cut was real enough. The growth that stacked on top of it was the larger story.
What Excel Actually Did
VisiCalc shipped in 1979, and Excel arrived for the Mac in 1985. The standard story is that spreadsheets gutted bookkeeping, and they did. Mid-skill bookkeeping clerks lost ground for decades. The less-told half is that accountant employment grew over the same stretch, because companies could now run forecasts, scenario models, and analyses that were too expensive to commission by hand. As NPR's Planet Money documented in its history of the electronic spreadsheet, accountants ended up doing more work, not less, because the lower unit cost of arithmetic expanded what was worth doing in the first place.
The category changed shape rather than shrinking.
The Long Lag
The 1980s and 1990s computerized white-collar work end to end and produced what economists called the productivity paradox, where capital investment in IT shot up while measured productivity growth slowed. Robert Solow's 1987 line was that you could see the computer age everywhere except in the productivity statistics, and the resolution took about a decade and a half to arrive. By the late 1990s the gains showed up, software developer employment grew into a multi-million-job category that did not exist at scale before, and entire industries restructured around computer-enabled growth rather than computer-enabled head count cuts.
Further back, the same pattern appears with electricity, the steam engine, and the assembly line. Each one produced a wave of cost cuts on existing work and a much larger wave of new work the technology made affordable.
Layoffs in the Lede
The current AI moment looks different from the inside, because the headlines are almost entirely about cuts: Klarna's AI assistant doing the work of 700 customer service agents, IBM planning to automate 30% of back-office roles, Ford's CEO saying half of white-collar America is on the line. The public discourse has settled almost entirely on the cost-savings axis, and the reason is that cost cuts are easy to predict. A CFO can sign off on cutting head count and write the savings into next quarter's plan, while the growth bet requires imagining a market that does not exist yet, building a product that depends on cheap intelligence, and finding customers who did not have a budget for that thing yesterday.
The growth bet is where the money has been every previous time, and the early signs of it are visible already. A 2025 paper on AI rebound effects noted that healthcare imaging volumes have continued to rise rather than fall in settings that adopted AI assistance, because the lower per-study cost made more studies clinically and economically viable. The same dynamic is starting to show up in legal review and analyst work, though the longer-term data on those is still being written.
William Stanley Jevons described the underlying mechanism in 1865, watching coal use rise as engines became more efficient. The prediction that falls out of his observation is that cheaper intelligence will be used more rather than less, and that the use cases nobody bothered to pursue at the old price will end up forming the largest part of the new market.
The Bet
Whether AI takes our jobs is partly a forecast and partly a choice. If companies on the cost-savings path outcompete companies on the growth path, payrolls shrink and the displaced workers do not all find new categories to land in. If the growth path wins, which is what the historical record consistently says, the jobs change shape faster than they disappear and new categories show up that we cannot fully name yet.
The two paths are not symmetric, since one has a floor and runs out while the other has not yet hit a ceiling that anyone can point to.
The branch managers in 1985 did not know that putting cash dispensers in their lobbies would mean more tellers, just doing different work. They were not making a sociological prediction, they were making small business decisions inside a competitive market, and the aggregate outcome was growth. The question for 2026 is whether the companies making decisions about AI right now are picking the same path, or whether something about this technology raises the cost-savings ceiling enough to break the historical pattern. The honest answer is that we will not know for another five to ten years, and the base rate is that the growth bet wins, the jobs change shape, and the people who learned to use the new tools end up doing work the old market could not afford.