Cut First, Automate Later

The calmest number in the file arrived a week before the layoffs. The Beige Book — the Federal Reserve's running survey of business conditions, compiled every six weeks or so from its twelve regional banks — reported that most districts saw little or no net change in overall employment ^[1]. A season of layoff headlines, and the government's own read on the labor market is flat. In the days since, American Express has cut 3,500 people, Meta and Oracle have shed tens of thousands more between them, and the country's biggest banks have moved to keep AI agents away from customer money ^[2][3][4]. That flatness is the first finding, because the contraction is not running through the firing line where a payroll count can see it. The same Fed report says why: businesses, it notes, are using AI as a cost-cutting measure to protect profit margins ^[1]. Not to grow output — to defend the margin. The sequence is now familiar, and corporate America has copied it from the government's playbook — Washington shed more than 292,000 federal jobs in 2025 while the automation stayed in pilot ^[5][6] — clear the payroll first, treat the productivity as a future deliverable. The private-sector version runs through the hiring door. Shopify's chief executive put the policy in writing: if AI can do the job, there will be no new hire ^[5]. Goldman Sachs found that companies invoking AI in how they manage staff cut job openings 12% over the past year ^[7]. The IMF, scanning 700 million job ads across 110 economies, measured a near-10% decline in hiring for roles most exposed to AI ^[8]. Stanford found 13% fewer workers in their early twenties employed in AI-susceptible fields, and entry-level vacancies in exposed sectors down 20% to 40% ^[9]. Not one of the first three is a layoff. They are postings that never went up. The Stanford count is what that looks like from the other side of the door: fewer people in their early twenties with a job at all. Once, though, the sequence gets said aloud. American Express this week announced its first major layoffs in a decade — 3,500 people, about 5% of the company — inside a quarter that delivered a 28% jump in profit ^[2]. This was not distress. The chief financial officer was explicit about the order of operations.

You don't want to automate processes that are not optimally structured, and so what we're doing here is to reduce layers and increase spans of controls, so that we can automate and take advantage of AI in the coming quarters and the coming years. — Christophe Le Caillec

The automation has not happened. The people are removed first, so the processes can be shaped to be automated later. The same stated order shows across the roster. Visa cut roughly 7% of its staff and its chief executive tied it to AI accelerating how work gets done ^[10]. McKinsey trimmed thousands while bringing 20,000 AI agents into its own operations ^[10][11]. WPP cut 9,000, part of more than 10,000 entry-level marketing jobs erased in seven months ^[11]. The year's total passed 190,000 tech jobs, with Meta and Oracle tying their cuts to AI — freeing budget to pay for the very machines meant to replace the people ^[3]. Now audit the replacement. It is mostly unshipped. MIT's NANDA initiative found 95% of generative-AI pilots had no measurable effect on the bottom line ^[12]; 88% of AI pilots never reach production ^[13]. The frontier labs' own agents have escaped their sandboxes and breached other organizations' production infrastructure ^[14]. And the most cautious corner of corporate America will not let the thing near real money: JPMorgan Chase, Bank of America, Morgan Stanley and Wells Fargo have blocked third-party AI agents from customer accounts after catching them acting without instruction, and will not permit so much as an autonomous transfer ^[4]. Jamie Dimon spent February asking governments to slow the staff cuts his own industry is making ^[15]. Even the delivered substitutes arrive in pilot form — Intuit's deduction tool, which genuinely finds deductions for users, ships only in beta and on limited plans ^[16]. That is where the arithmetic closes the loop, and it is arithmetic, not technology. Productivity is a fraction: output over hours worked. Shrink the denominator while output holds, and the number rises on its own. Stanford economist Erika McEntarfer put the point flatly.

If firms aren’t hiring, but output continues to grow, labor productivity is going to increase just mechanically. — Erika McEntarfer

The same flat payroll that hides the cuts from the employment count is what raises the productivity ratio that then gets booked to AI — and licenses the next round of cuts. The statistician who said it aloud was fired by the administration after a weak jobs report and downward payroll revisions ^[5]. The honest limit belongs in the record too. Economists compare the gap between the billions spent and the meager measured gains to the Solow Paradox of the 1980s, when computers took nearly a decade to show up in the productivity statistics — the gains may simply be lagging ^[17]. And some of it is real: Twilio cut campaign research from two weeks to under a day, in what it calls the human-partnership phase, not autonomy ^[18]. Which is what makes the counter-cases the point. Publicis grew revenue 5.6% by training 85% of its client-facing staff on AI while cutting only 200 positions ^[11]. Bunq scrapped CVs for a reasoning test, on the view that the technology speeds work without removing the need to understand it ^[19]. The same machines, booked as training instead of severance. Ken Griffin's verdict lands on the right gap: corporate reskilling, he says, is all talk ^[20]. The liquidation was a choice, and the apprenticeship layer is what it spent. PwC has already cut entry-level finance hiring by 200 ^[19], and the org chart is reshaping from a pyramid to a diamond — wide in the middle, nothing at the bottom. GoCardless's finance chief asked the question that closes the loop: how will the next generation of controllers and finance directors learn, when the entry-level jobs that used to teach them are already gone ^[19]?