Past Technologies Took Decades to Shrink a Job. AI Is Starting to in Three Years.
Robots and ATMs took decades to show up in employment counts. AI's first drop in employment, among young workers in exposed jobs, appeared in three years.
By Geordie Everitt
Every earlier technology that took work from people took decades to show up in the job counts. AI agents have started to show up in three years.
A chatbot at home changes how one person looks things up. An agent in a business does the work someone was paid for: it files the claim, reconciles the invoice, answers the customer. That is where jobs go, so I measured how long earlier technologies took to take them.
The factory
The first industrial robot went into a General Motors plant in 1961. US manufacturing employment peaked in June 1979 at 19.6 million, in the Bureau of Labor Statistics series on the Federal Reserve's FRED database. That is 18 years from the first robot to the peak. By June 2010 employment was 11.5 million, down 41 percent, and it stands at 12.7 million now.
A machine was on the floor for two decades before the job count turned.
The teller
ATMs spread through the 1970s and 80s, and for decades the tellers stayed. The economist James Bessen found that teller employment rose as ATMs spread: each branch needed fewer tellers, branches got cheaper to run, and banks opened many more of them. The Bureau of Labor Statistics says teller growth has followed branch expansion. It counts 339,200 tellers in 2025 and projects a 13 percent decline by 2035, tied to fewer branches.
The ATM took the cash-handling task early, and the job count held for decades because cheaper branches created demand for tellers somewhere else. The task goes first and the job much later, and in between new demand either offsets the loss or doesn't.
What AI shows
Three Stanford economists, Erik Brynjolfsson, Bharat Chandar and Ruyu Chen, used payroll records from ADP to compare workers by age across occupations, ranked by how much of the work AI can do. Workers aged 22 to 25 in the most AI-exposed occupations lost 16 percent of employment relative to those in the least exposed. Older workers in the same jobs held steady or grew. The drop concentrates in jobs where AI automates tasks, and it starts in late 2022.
The consulting firm Challenger, Gray & Christmas counts announced layoffs. It attributed 54,836 US job cuts to AI in 2025, and 101,743 of 443,604 announced cuts in the first half of 2026, which is 23 percent.
Robots took 18 years to turn the job count. AI's first measured drop came in three.
Why it comes faster
A robot or an ATM takes one task out of a job. An agent takes the tasks across a job, the same ones entry-level workers were hired to do and learn from.
The ATM case shows what an offset looks like: cheaper branches made more branches. Whether cheaper cognitive work makes more of itself the same way is the open question, and the payroll data for workers aged 22 to 25 show no offset yet.
Maybe the machine can help
The gap between losing the task and losing the job was the time society had to build something else: new branches for tellers to staff, new work in the same building. That gap is three years now, and the institutions that fill it, schools, tax codes and labor law, still move at the old pace.
The change is cheap intellectual labor, which I wrote about in Money or Happy. A bright person with a model can now read two centuries of economics, run the arithmetic on a proposed policy and draft the bill in an afternoon. The problem runs at the speed of the technology that created it, and that technology is the one tool that works at that speed.