A year ago this week, Ford’s chief executive Jim Farley stood up at the Aspen Ideas Festival and put a number on the fear. Artificial intelligence, he told the room, would replace “half, literally half” of all white-collar workers in the United States. It was a striking thing for the boss of a car company to say, and it travelled fast. Twelve months on, it is worth asking a plain question: how is that prediction ageing?
Not brilliantly, so far. The 2025 numbers are in, and they make for bleak reading, though the cause is mostly something other than AI. Challenger, Gray & Christmas counted 1,206,374 announced job cuts across US employers last year, up 58 percent on 2024 and the seventh-highest annual total since 1989. AI was named as the reason in just 54,836 of them, a little over four in every hundred. The rest came from government reductions, closures, restructuring, and old-fashioned economic caution. The machines are not, on this evidence, halving anything yet.
We Have Watched a Machine Come for a Job Before
Put the large language model to one side for a moment and look at the automated teller machine. When banks began bolting cash machines to their walls in the 1970s and 1980s, the logic looked airtight: the ATM does what the teller does, so the teller is finished.
It did not turn out that way, and the man who documented why is the economist James Bessen. Writing for the IMF, he showed that as more than 400,000 ATMs were installed across America, the number of bank tellers did not collapse. It rose. The machine cut the tellers a branch needed from twenty to thirteen between 1988 and 2004, which made branches cheaper to run, so banks opened far more of them: urban branch numbers grew by 43 percent. Fewer tellers to a branch, many more branches, and the job survived.
The Teller Survived the ATM. Then the Internet Arrived.
Something subtler happened as well. Once the machine took the cash-counting, the part of the job that remained was the human part: knowing the customer, selling the mortgage, spotting the small-business owner worth keeping. Tellers moved up into what the banks politely called relationship management. The task was automated. The worker was redeployed. That is the pattern Bessen found right across the computer era, from typographers becoming graphic designers to switchboard operators becoming receptionists.
That reprieve, though, depended on the ATM being the only machine in the room, and it never stays that way. Here is the part the comfortable version leaves out: one technology buys time for the next. The cash machine made a branch cheap to run, which is precisely why so many branches stayed open through the 1990s. Then the internet arrived, and online and mobile banking did the thing the ATM never could. They made the branch itself redundant.
Look at what followed. The number of bank branches in the UK fell from 14,689 in 1986 to 5,745 in 2023, according to the House of Commons Library and the ONS. By 2021, Barclays was telling MPs that 74 percent of its customers already banked by phone, online, or mobile. In the United States, teller numbers have slid too, to around 347,000, with the Bureau of Labor Statistics expecting a further 13 percent fall by 2034. The teller was not automated out of a job by the cash machine. The teller’s workplace was shut a wave later, once the customers had moved online, and those were local jobs on local high streets. That is the real shape of the ATM story, and it is worth keeping in mind the next time someone assures you a new machine will leave employment untouched.
Where the Comparison Holds, & Where It Breaks
So what does the cash machine tell the director staring down a generative-AI budget line? Two things, and they point in opposite directions, which is rather the point.
The reassuring half is real. Automating a task is not the same as deleting a role, cheaper processes can grow demand instead of shrinking headcount, and the human residue of a job tends to become more valuable. On that reading, Farley’s “literally half” is the Frey and Osborne paper all over again: that 2013 Oxford study famously judged 47 percent of American jobs at high risk of computerisation, and here we are, more than a decade on, with employment having grown.
The uncomfortable half is the arc we have just traced. The teller kept the job through the first wave and lost the workplace to the second, and the gap between the two was about fifteen years. Generative AI may run the same play faster. It goes straight at the cognitive layer, the drafting, the analysis, the first-pass reasoning that used to be how a junior lawyer or analyst earned their keep, which is the very ground the relationship-era teller retreated to. And where the ATM hit one occupation, AI arrives across dozens at once, so there is no obvious neighbouring trade waiting to absorb the displaced. The first wave of a new technology usually looks benign, and that is the trap, because it is the second wave, the one the first makes room for, that ends up closing the branch.
In Britain the near-term picture is quieter, and pointing the same way. The latest KPMG and REC UK Report on Jobs found permanent placements falling at their fastest rate in ten months in May, with employers leaning on temporary and interim staff where they needed heads at all. That is the same freeze-and-flex behaviour I wrote about in when the workaround becomes the workforce.
The Forecast, With Its Assumptions on the Table
Here is where I will put a number down, with the caveats visible instead of buried. Over the next three years, I do not expect anything close to a halving of white-collar employment. My base case is that AI-attributed cuts keep climbing from that four-percent share but stay a minority of total layoffs through 2028, while the deeper effect shows up in roles that are never advertised: the analyst who is not backfilled, the team of eight that becomes a team of five. The senior layer thins from below.
That call rests on three assumptions. That the models keep improving at roughly today’s clip, with no step change that genuinely automates judgement. That firms keep hitting the same integration and trust problems that have made AI adoption slower inside large organisations than the headlines suggest. And that no external shock, a recession or a regulatory clamp, forces cuts faster than the technology alone would. Break any one of those and the range widens. A true reasoning breakthrough, and the branch-closing second act arrives in five years, not twenty. A sharp downturn, and plenty of firms will use AI as the cover story for cuts they already wanted to make, which is arguably happening now.
For a senior executive, forget guessing the date. The task is to make yourself the part of the work a model cannot carry: the judgement, the accountability, the relationships that keep a business honest under pressure. That is a positioning problem as much as a performance one. It is worth checking whether your executive CV proves judgement instead of listing responsibilities, whether your LinkedIn profile shows a decision-maker or a job description, and, if the worst does arrive, knowing that career transition and redundancy support exist as a floor under you, not a mark against you.
Frequently Asked Questions
Will AI really replace half of white-collar jobs?
No serious labour data supports that yet. In 2025, AI was named in about 55,000 of more than 1.2 million announced US job cuts, under five percent of the total. The larger forces were government reductions, closures, and economic caution. The honest position is uncertainty about the pace, and calm about the near term.
Is the ATM comparison too optimistic?
It is, if you stop the story in about 2005. Teller numbers held up while the ATM was the only machine, then the internet arrived and the branches themselves closed. UK bank branches fell from 14,689 in 1986 to 5,745 in 2023. The lesson is that one technology buys time before the next one removes the job, and the second wave took the local high-street jobs with it.
Which executive roles are most exposed?
Roles built mainly on producing analysis, drafting, and first-pass synthesis are most exposed, because that is what current models do well. Roles built on judgement, accountability, negotiation, and trusted relationships are more defensible, for now.
What should I do about my CV and profile now?
Shift the emphasis from tasks to judgement: decisions made, risks carried, outcomes owned. Those are the things a model cannot claim on your behalf, and they are what a senior hirer is actually buying.
Farley may yet be proved right about the destination. He is almost certainly wrong about the speed, and speed is what matters when there is a career to plan. The cash machine did come for the teller in the end. It just took a generation, and it took the whole branch with it. The people who came through it were the ones who saw the second wave coming and made themselves useful in the ways a machine could not copy. That option is open to you too, and the sooner you take it, the longer your reprieve.
