JPMorgan Chase has more than 450 agentic AI use cases in production. Not pilots. Things that run every day inside a bank of 318,512 people. That headcount is the part worth staring at, because it barely moved. Underneath the flat line, roles did move: operations down roughly 4%, support down 2%, client-facing up 4%. The bank’s attrition runs near 10% a year. Call it 30,000 people who leave on their own, and then get replaced, or don’t, or turn up in a different building. No press release goes out for that. “We reassigned some people” has never once lifted a share price.
Now the number everyone quotes instead. Bloomberg reported on July 1 that the financial-activities and information sectors are shedding an average of 28,000 jobs a month in 2026, against a broader American labour market that added 113,000 a month through May. The two sectors with the deepest AI penetration are contracting while everything else grows. Net across the economy is still positive by 85,000 a month, which is cold comfort if you write software or process loan files for a living.
That payroll data is real. What it does not contain, anywhere, is a cause. It’s not because a sector is shrinking that the machines are the thing shrinking it.
The cheapest sentence in a layoff memo
Deutsche Bank analysts gave this a name I wish I’d thought of: AI redundancy washing. Their warning is that it becomes a defining feature of 2026, and the supporting figure is the one that should make you re-read every workforce announcement from the past year. Close to 6 in 10 companies admit they frame reductions as AI-driven when the actual driver is financial pressure. Sam Altman has said the same thing out loud — companies are blaming AI for layoffs they were going to do regardless. Odd thing for him to volunteer, given which story sells more inference.
Nothing mysterious in the mechanism. Two memos empty the same 400 desks. One says the company over-hired in 2021 and misjudged demand. The other says the company has automated a function and is redeploying capital toward growth. Identical desks, identical severance, and analysts price the second one higher, because the first is a confession of a forecasting error and the second is a strategy. Given a free choice between those, nobody in corporate communications picks the confession.
I should flag my own sourcing here, because I’d want that if I were reading this. The JPMorgan role shifts and the Deutsche Bank figure both reach me through an aggregation of displacement statistics, not through the bank’s own filing or the analyst note itself. I went looking for the primary documents and mostly found compilations quoting other compilations. So treat the direction as solid and the decimals as soft. If someone has the underlying Deutsche Bank note, I want it.
The deployment arithmetic doesn’t support the story
Here’s the part that made me sit up. Roughly 88% of organizations report using AI in at least one function. Only 7% have fully scaled it across the organization, and 56% of CEOs report zero measurable return in the past twelve months. Put those three numbers in a row and ask which cohort is displacing labour at industrial scale.
Not the 56%. A company that cannot detect a dollar of return did not eliminate roles because of an efficiency gain it can’t find in its own books. Maybe the 7%, and 7% of enterprises is a thin base to hang a national labour trend on. The honest reading is that the displacement narrative is running well ahead of the deployment.
Where displacement does show up, it looks like composition rather than volume. The pattern in the reporting on where AI is actually replacing jobs is a freeze at the entry level, plus internal redeployment of everyone already inside. Same total headcount, older workforce, fewer doors in. Which maps neatly onto JPMorgan’s own split: operations and support down a couple of points, client-facing up four. And franchement, that’s the part I’d worry about if I were 23 with a fresh degree in something adjacent to spreadsheets. It’s a slower and far less cinematic story than mass replacement, and it moves the cost onto a different person: not the incumbent, the applicant.
Klarna shows what the credible version looks like. Its agent handled work equivalent to 853 full-time staff by the third quarter of 2025 and delivered about $60 million in cost savings, per a roundup of production agent deployments. A named system, a named workload, a named quarter, a number you could argue with. Compare that to the standard announcement, which names no system, no volume, and no before-and-after, then attributes the cut to AI in the third paragraph. One of those is a measurement. The other is a mood with a headcount attached.
Me, I don’t buy the clean version of either side. Anyone who claims AI has already eaten a third of a million jobs a year can’t show you the substitution. And if you think nothing at all is happening, explain why the contraction sits in exactly the two sectors that adopted fastest, and why entry-level hiring in knowledge work has gone quiet.
Where I could be wrong about this
Here’s the best case against everything above, and it’s the job that never gets posted. AI washing requires an announcement to wash. Attrition with no backfill produces no memo, no severance line, no press cycle. It is exactly what a flat headcount plus 10% annual churn looks like from the inside. JPMorgan holding at 318,512 while shifting roles internally is consistent with my thesis and equally consistent with a bank quietly declining to replace the analysts who left. Same data, opposite conclusions, and I can’t separate the two with public numbers.
Consider Goldman Sachs’s projection: 300 million jobs globally exposed to automation, then 97 million new roles created against 85 million displaced. Net plus 12 million, set against 300 million exposed. That’s 4% of the affected pool, a rounding error dressed as a forecast. When your error bar is wider than your result, you’re describing uncertainty, not the future.
What would change my mind is a specific signature: sector payroll shrinking while that sector’s output and revenue climb. Fewer people, more product. That’s substitution, and it’s distinguishable from a downturn. If the S&P Global employment work or the government series shows information-sector output rising through 2027 on a smaller payroll, I’ll drop the washing thesis without much of a fight.
Until then, one test for the next earnings call you sit through. When a company credits AI for a headcount reduction, listen for whether it names the system and the volume that system handles. Klarna said 853. Most say nothing, and nothing is a number too.
