AI is replacing the work that used to define the first decade of a career. At the same moment, organisations are quietly thinning out the people whose work AI cannot do.
That pairing sits somewhere between obvious and uncomfortable, depending on which part of the workforce you sit in. The data behind it is no longer disputed. The talent decision being made in response to it is almost universally backwards.
What AI is actually replacing
The popular framing of AI in the workplace is that it threatens knowledge workers broadly. The 2025 data tells a sharper story.
Stanford’s Digital Economy Lab released a study in November 2025 with the deliberately unsettling title Canaries in the Coal Mine. It tracked employment in occupations highly exposed to AI from late 2022 onwards. Workers aged 22 to 25 in those roles saw employment fall by roughly 13%. Workers in their forties, fifties and sixties in the same occupations continued to grow.
The mechanism is not redundancy. It is non-replacement. Entry-level vacancies are quietly not being backfilled. The career ladder is losing its bottom rungs.
The Stanford authors are unusually direct about why. AI is replacing codified knowledge, the part of expertise that can be written down, while complementing the experiential wisdom that only comes from years on the job. Other 2025 work in customer support and software development tells the same story. AI lifts the bottom of the distribution faster than the top. Two-month-experience workers using AI now match six-month-experience workers without it. The work AI does best is the kind of standardised, learn-from-a-book task that used to define the first few rungs of a career.
The thing AI cannot replicate
There is a second half to this story that gets less coverage.
Boston Consulting Group ran a study with Harvard Business School using 758 of its own consultants and GPT-4. On standard tasks, AI users completed 12% more work, 25% faster, with 40% better quality. The finding that rarely makes the press summary: when the same study tested tasks designed to fall outside the model’s actual capability, consultants using AI were 19 percentage points less likely to produce correct solutions. AI made experts wrong more often when the problem required judgement AI lacked.
The capability to know when to trust an AI answer and when to override it is itself a function of experience. It is built from a personal library of cases, situations and outcomes that no model has been trained on.
Decades of research in naturalistic decision-making, the field Gary Klein founded by watching firefighters and military commanders make calls under uncertainty, describes the same mechanism. Experts under pressure do not deliberate between options. They pattern-match against situations they have seen before. The library is built by exposure, not by reading frameworks.
This is what is meant by judgement. It is the residual human advantage in the AI era, and it has a clear demographic profile.
The talent decision being made backwards
Put the two findings beside each other.
AI is removing the codified, junior-level work fastest. The cohort whose work AI is actually complementing is the experienced one. The economic logic of an organisation in 2026 should be to lean into that experienced layer, because it is the part of the workforce AI cannot reproduce and which increasingly determines the quality of any AI-augmented output.
What organisations are doing instead is the exact opposite. The over-50 cohort is being quietly thinned through restructures, voluntary exit programmes, redundancy schemes, and the slow erosion of roles experienced workers tend to hold. It is rarely a stated policy. It is almost everywhere a pattern.
The talent decision is being made backwards. The cohort being pushed out is the one most worth keeping. The cohort being squeezed at the bottom is the one whose work AI is already doing. The organisation ends up with no future and no memory.
The cost of forgetting
There is an institutional dimension to this that gets ignored because it does not show up in the next quarterly report.
Roughly 42% of an organisation’s working knowledge sits in the heads of individual employees and nowhere else. Industry estimates put the cost of knowledge loss from rapid organisational change at tens of billions of dollars a year across Fortune 500 firms. The direction is consistent even where the precise figure varies. Restructures remove people, and the people take the unwritten knowledge with them. A newly arrived CEO who clears out the over-50 cohort does not just lose those individuals. They lose the only group who remembers why the last three transformations failed and what is different about this one.
That is not a fairness argument. It is a structural one. The organisation is paying a real cost. It will appear on the books eighteen months later, in the form of mistakes the experienced layer would have caught.
What we are not calling it
Age discrimination is unlawful in most major jurisdictions. It is also one of the most reliably under-reported categories of workplace harm, because it is rarely framed as discrimination by the people doing it.
ProPublica’s multi-year investigation into IBM found the company eliminated more than 20,000 workers aged 40 and over from 2013 onwards. The US Equal Employment Opportunity Commission concluded in 2020 that the layoffs had a clear adverse impact on older workers. More than 85% of those targeted for layoff in that period were older workers, even when rated as high performers. In 2023, former IBM HR professionals filed suit alleging termination linked to age and explicit plans to replace them with AI.
Almost no one running these processes describes them as age discrimination. They are called “right-sizing,” “talent refresh,” “succession planning,” “rebalancing the pyramid.” The language and the outcome have been routinely diverging for at least a decade. What is new in 2025 is that AI has made the underlying decision economically illiterate as well as legally exposed.
Where this leaves you
If you run an organisation that has quietly pushed out the experienced cohort, you have spent real money to remove the layer of your workforce AI cannot replicate, while leaving in place the layer whose work AI is doing without you noticing.
The question is not whether you can afford to keep experienced staff. It is whether you can afford to lose them at exactly the moment they became the most valuable people on your payroll.