Tech cut jobs and blamed AI two years ago. Banks are doing the same now.

Bloomberg reported this month that top bankers at JPMorgan, Citi, and Goldman have said AI will eliminate some jobs at their firms. The finance sector was widely predicted to be next in line after tech. It now appears to be (3).

Two years is long enough to see how the first wave went. Some of the tech companies that led it are now quietly hiring people back, often at a higher price tag than before. Orgvue's 2025 workforce research found that 39% of business leaders had made AI-related staff redundancies, and 55% of that group now say the decision was wrong. Gartner has projected that by 2027, half of organizations that expected to significantly reduce their customer-service workforce because of AI will abandon those plans (1, 2).

Named cases have moved from fringe to routine. IBM's AskHR system handles 94% of HR queries, and the company has separately announced plans to triple US entry-level hiring in roles widely forecast as AI-replaceable. Ford brought back 350 veteran engineers after AI-based quality systems fell short, and topped J.D. Power's 2026 Initial Quality Survey among mainstream brands. Klarna acknowledged going too far on AI-driven customer-service replacements and began hiring human agents back (4, 5, 6).

An analyst line worth reading twice.

Inside the Bloomberg piece on the bank cuts, Pooja Sriram, senior US economist at Barclays, said something worth reading twice: "Some of this could genuinely be productivity replacing workers. But the narrative that keeps coming up is really a cost-cutting exercise by a lot of firms, given the amount of investments they have committed towards AI" (3).

That is a specific claim, made carefully. Some of the cuts are productivity replacing workers. The pattern in the headlines is not.

AI can replace work. That is not what is being tested here.

There is no serious argument that AI cannot replace positions and tasks. In the categories where it works, it works. That case is settled enough to stop rehearsing.

What is worth looking at is the gap between what measured productivity would justify and what is actually being cut. If the layoffs were being sequenced by productivity data, we would expect them to move role by role, sector by sector, slowing when the data did not confirm the model was ready. That is not the pattern we have. The workforce cuts have moved faster and deeper than measured productivity would support, and they have moved fastest in the sectors under the heaviest capital-markets pressure to show a return on AI investment.

Enterprise AI spending in 2025 ran into the hundreds of billions of dollars (7). Numbers of that size require a visible return. Real productivity data takes years to develop and years more to attribute cleanly to the tool. A headcount reduction shows up in the next quarter.

That is the mismatch. Labor cuts move on quarterly time. Productivity moves on different time intervals and is much harder to measure. When the two get out of sync, executives reach for the tool that can show impact by the next earnings call. A headcount reduction can. A productivity gain that has not yet arrived cannot.

The productivity math has a missing term.

The math being run on AI layoffs, at least the version that shows up in the board deck, is straightforward. Cost of the eliminated roles, subtracted. Cost of the AI system, added. Output of the AI system, added. Net productivity gain, calculated.

That math has a known error. It counts what the tool does. It counts what the eliminated people were doing. It does not count what happens to the productivity of the people who stay.

The research on the remainders is decades old and consistent. Gallup's engagement research has found that layoffs damage engagement, retention, and trust across the surviving workforce. In the HR literature, the pattern has a name: layoff survivor syndrome, first documented by Joel Brockner and colleagues in the 1980s and replicated across dozens of studies since. Trevor and Nyberg's 2008 study in the Academy of Management Journal quantified one edge of it: even a modest downsizing, a 0.5% workforce reduction, was associated with a 2.6 percentage-point increase in voluntary turnover in the following year. The people the company most wants to keep tend to leave first, because the market rewards them for doing so (8, 9, 10).

A cut takes more than the people cut.

When people leave, productivity does not just drop by their share of the output. It drops for the people who stay too.

The load doubles, and it doubles in a specific way. Research on generative AI adoption shows that the tools shift work from doing to overseeing. Employees spend meaningful cognitive effort validating, correcting, and integrating AI output. A 2026 Connext Global oversight report found that only 17% of US adults believe workplace AI is reliable without human oversight; the rest say it needs light review or dedicated supervision. When a company cuts headcount and hands the work to AI, the people who stay carry two loads: the departed colleague's work, and the oversight the AI itself now requires (11, 12).

The trust thins. The workhorses who were carrying the critical work carry more of it, and less of it gets noticed.

That is the term the productivity model leaves out. Layoffs take more than the headcount on your balance sheet. They take some of the productivity of everyone left, and they take it hardest from the ones the company can least afford to lose.

Two years into the AI layoff wave, that missing term is starting to show up on the other side of the ledger. The rehires are the visible correction. The engagement scores, the quiet exits of senior talent, the productivity of the workhorses: those show up as the productivity gain that was supposed to arrive not arriving.

The dominant story of AI layoffs is a story about the model. The story worth watching is the one about the math.

Sources: