A century of automation history comes down to one ratio: how fast the work disappears against how fast people were already leaving.
The short version: Economists are digging through the 1920s and 1930s for clues about AI, and the sharpest lesson in that history is not that technology kills jobs or creates them. It is about speed measured against turnover. AT&T’s telephone operators quit at up to 40% a year, and a dial cutover still erased 50% to 80% of young operator jobs in a city within minutes. Teamsters shrank by about 1.8% a year as motor trucks arrived, and the ones who stayed got raises. Every occupation has a no-layoff speed limit, which is simply the rate at which people leave it on their own. On federal projections, customer service reps leave their occupation at about 13% a year. Software developers leave at about 4.8%. That ratio, more than any benchmark score, decides who gets hurt by AI and what it costs the employer to get there.
Phone companies liked to switch a city over to dial service around midnight on a Saturday, when call volume was at its lowest. An operator could leave her switchboard on Friday and come back Monday to find there was no switchboard.
That scene anchors a New York Times piece by Ben Casselman this week on economists who are reading AI through earlier disruptions, from threshing machines to talking pictures. The takeaways are the familiar trio: the wider economy matters, speed hurts, and jobs change more often than they vanish.
All true. But the research underneath those stories holds a tighter rule, and it comes with a number a CFO can actually plan around. The operators were not hurt because automation was fast. They were hurt because it was faster than they were already leaving.
What Happened
Casselman opens in December 1931, at the annual meeting of the American Economic Association, where economists worried that dial phones, talkies and gasoline trucks were destroying work faster than any earlier wave. Unemployment was near 20% and headed for a record 25%. One economist, Ewan Clague, who later ran the Bureau of Labor Statistics, floated taxing machinery to slow its spread.
The modern echo is loud. In July, more than 200 economists and AI researchers, including 16 Nobel laureates, signed a four-sentence letter organized by Stanford’s Digital Economy Lab warning that AI could reshape the economy more than the Industrial Revolution did, over a far shorter stretch of time.
The historians Casselman interviewed offer three lessons:
- Context matters. Displaced phone operators found work quickly in the 1920s boom and struggled badly once the Depression hit.
- Speed matters. Overnight cutovers left operators no time to adjust, while the slow decline of horse-drawn hauling let young men steer elsewhere.
- Jobs change. England had about 220,000 boot makers in 1851 and about 207,000 in 1911, but the work moved from artisan shops to factories as cheaper boots expanded demand.
Meanwhile, the live data is muddier than the headlines. Challenger, Gray & Christmas counted 120,136 announced US job cuts blamed on AI from January through September 2026, about 21% of all cuts and the leading reason for the year. In September alone, AI slipped to the fifth most cited reason, with 3,961 cuts.
The Backstory
Put the two most famous cases side by side and the “speed” lesson turns into something more precise.
The operators. By the early 1920s AT&T was the largest employer in the country, and operators made up about half its workforce. James Feigenbaum and Daniel Gross, whose 2024 study in the Quarterly Journal of Economics is the research behind the Times anecdote, found AT&T’s own internal memos describing operator turnover of up to 40% a year. This was a job young women held for a while and then left.
AT&T knew it, and it tried to use it. Preparing a city for dial took two to three years. The cutover itself took minutes. In between, phone companies hired temporary operators so that when the wires were cut, the people let go were largely temps. That is attrition, engineered on purpose.
It worked for the youngest workers. It failed for the career ones. After a city’s cutover, the number of young women working as operators fell by 50% to 80% almost immediately. Incumbent operators were about 8 percentage points less likely to still be operators a decade later, roughly a third off the normal rate. The damage concentrated among operators 26 and older, who were about 7 points less likely to be working at all. Those who kept working saw their occupation’s typical earnings fall about 5% while similar women elsewhere gained 8%.
The pattern inside the data is the point. The older operators were the low-turnover segment. They were not going to quit on their own, so the cut landed on them.
The teamsters. Kevin Lang and co-authors at Boston University tracked the drivers of horse-drawn wagons as the motor truck closed in. The census counted 421,983 teamsters in 1910 and 350,657 in 1920, a 16.9% drop across a decade, or about 1.8% a year. Young men stopped signing up, the occupation got older, and teamster wages in five cities rose between 1917 and 1919 relative to comparable jobs. Lang calls the extra pay “obsolescence rents.”
Same era, same kind of shock. One transition cut faster than people left. The other cut slower. That difference, not the technology, decided who paid.
The Plan
Here is the part that should change how you read the AI layoff headlines: employers in 2026 appear to be running the teamster playbook, whether or not they have heard of it.
Stanford’s August revision of its “Canaries in the Coal Mine” research, built on ADP payroll data, finds that employment of 22- to 25-year-olds in the most AI-exposed occupations now sits about 19% below where it would be had it tracked less-exposed peers. In raw levels, that group fell about 11% between November 2022 and June 2026 while the least-exposed group grew about 10%. The authors say the adjustment is running primarily through reduced hiring of young workers rather than increased separations, and experienced workers show no comparable gap. We traced the same squeeze through the consulting pyramid in August.
That is the front door closing, not the back door opening. And layoff trackers only count the back door. Challenger’s AI tally peaked at 38,579 cuts in May and fell to 3,462 in August and 3,961 in September.
The front door is also far bigger than people assume. Every occupation loses workers each year to retirements and career switches. The Bureau of Labor Statistics publishes employment and annual openings for each one through 2034; strip out projected growth and what remains is the flow of people leaving. Run that math and customer service representatives alone exit their occupation at about 357,000 a year. Challenger’s nine-month count of AI-blamed layoffs across every industry in America is 120,136.
Divide each occupation’s exits by its size and you get its no-layoff speed limit.
- Customer service representatives: about 13.0% a year, on 2.8 million jobs
- Paralegals and legal assistants: about 10.4%
- Heavy and tractor-trailer truck drivers: about 10.0%
- Software developers: about 4.8%, roughly 88,000 exits a year on 1.7 million jobs
A call center can retire about one seat in eight every year without firing anyone. A software team can retire about one in twenty. On this measure, software developers are not the 19-year-old switchboard operator who was leaving anyway. They are the career operator past 25 who was not.
One honest caveat belongs right here: BLS still projects software developer employment to grow 16% by 2034. The speed limit is not a forecast of decline. It is the price list for a cut if one comes.
The Business Model Angle
Your exit rate is your automation budget. Work removed below an occupation’s natural exit rate costs roughly nothing in severance, notice periods or morale. Work removed above it gets paid for in layoffs. AI savings targets set by capability (“the model can do 30% of this team’s tasks”) instead of by turnover (“this team loses 5% a year”) will overrun on cost or overrun on time, and low-turnover, high-salary functions are where the overrun is largest.
Productivity only rescues jobs if someone cuts the price. Feigenbaum and Gross found cutovers were often paired with telephone rate increases, with AT&T citing the new capital to justify them. The savings went to the company, so no wave of cheaper calls created new phone jobs. The boot makers ran the opposite play: cheaper boots meant more pairs sold and a workforce that barely shrank in sixty years. For AI, the question to ask is whether a law firm, an agency or a software vendor passes its productivity gain through to price. Markets where it does have a path to more volume and more work. Markets where it does not are booking margin.
Contracts bring back the midnight cutover. A telephone exchange was indivisible, so it flipped all at once. AI is bought by the seat, which lets most employers take the slow teamster route. The exception is outsourcing. When a client moves a back-office function from a human vendor to an AI one, the vendor does not get to shrink by attrition. It loses a contract and a floor of agents in one decision. That is why the outsourcing firms that run other companies’ operations, a line of business Accenture runs at scale, and the call-center economies the Times flags abroad, carry the most operator-shaped risk in the AI economy.
The Risk
The speed-limit framing is useful, and it is easy to push too far. Four objections a well-informed skeptic would raise:
Turnover shrinks when you need it most. People quit less in a weak job market, which lowers every exit rate in the chart. The 1930s are the warning: the same dial telephone that displaced workers gently during the 1920s expansion hit hard once the Depression arrived. If AI cuts land in a slowdown, the attrition budget gets smaller exactly when finance teams lean on it.
Attrition is not painless. It is relocated. Feigenbaum and Gross estimate that, given operator turnover, cutovers may have shut off entry-level opportunities for as much as 10% to 15% of peak cohorts of young women. Later cohorts did find work, often in secretarial jobs and sometimes in lower-paying restaurant work. A closed front door hurts people who never show up in a layoff count, and Stanford’s 19% gap says that is already happening.
The technology may outrun every front door. If the work a model can absorb arrives as fast as a software update, the cut a firm wants could equal several years of exits. At that point, the economics favor the cutover, and history’s gentle cases stop applying. That is the scenario the July letter describes.
The numbers are averages and projections. The BLS figures are built on historical behavior and describe whole occupations, while turnover at a single firm can run far higher or lower. AT&T’s 40% is company turnover, which is not the same measure as an occupation-wide exit rate, so the panel in the chart is a benchmark, not a like-for-like bar.
The optimist’s case deserves its place too. Stanford finds no widespread, economy-wide displacement so far. Future cohorts of young women in cutover cities were employed at the same rates as before. And by one estimate cited in the Times, about 60% of US jobs in 2018 were in occupations that did not exist in 1940.
Quick Questions
Is AI causing mass layoffs right now? Not on current evidence. AI-blamed cuts are a minority of announced layoffs and fell sharply after May. The bigger effect is showing up as fewer hires for young workers in exposed occupations, which layoff trackers do not capture.
Why would a low turnover rate make a job more vulnerable? Because it shrinks the cheap path. A team that loses 13% of its people every year can absorb a large automation gain without a single firing. A team that loses under 5% cannot, so any sizable cut there either takes years or arrives as layoffs.
What did the 1931 economists want to do about it? Clague suggested taxing machinery to slow its adoption, an idea that has resurfaced almost word for word. We ran the arithmetic on Bill Gates’s token tax in August.
What should a manager do with this? Map each function by its natural exit rate before setting AI savings targets. Where turnover is high, pace adoption to it and redeploy rather than replace. Where turnover is low, budget honestly for severance and retraining, because the slow route will take years.
The Business Model Analyst Take
The history lesson everyone is drawing from this research is “go slow.” The more useful one is “go slower than people leave.” Speed only means something against a benchmark, and the benchmark already exists for every job in America: its exit rate.
That reframes who should worry. The roles with the fastest churn, like call centers, can absorb automation through hiring freezes, and the pain lands on applicants rather than employees. The roles with the slowest churn, like software development, are where the switchboard math returns, because career workers there are not going to leave on their own. And in both cases, the person who pays is increasingly someone who never got the job in the first place.
Kevin Lang, who studied the teamsters, has a line for colleagues who insist the world has survived every past scare: “I’ve been sick many times, but I’ve never died.” The speed limit will not decide whether AI is the sickness or something worse. It will decide whether the cost lands on paychecks or on applications that never get answered.