There's a moment that happens in conference rooms and on Slack threads thousands of times a day, and nobody designs for it. A junior analyst watches a senior colleague pause over a number. The number looks fine. It passed every check. But something about it, in context, against a pattern the senior person has absorbed over years and couldn't fully articulate if asked, makes them stop. The junior analyst registers this. Files it somewhere pre-verbal. Months later, they'll pause over a different number for reasons they can't quite explain either.

Cognitive scientists who study expertise call this cognitive apprenticeship: learning that happens through proximity to skilled practitioners doing real work, absorbing knowledge that masters of a domain often cannot make explicit even to themselves. The Dreyfus model of skill acquisition, developed for the Air Force in 1980, traces the path from novice to expert as a gradual shift from rigid rule-following to a largely intuitive mode of operation built on exposure to thousands of concrete cases. None of this was something organizations planned. It was a byproduct of having people in the room. Accidental infrastructure, performing a function no one had named.

Entry-level positions in the U.S. have declined roughly 35% since January 2023, according to labor research firm Revelio Labs. Among the fifteen largest tech companies, recent graduates made up just 7% of new hires in 2024, down from 11% two years prior. Meanwhile, 35% of roles still labeled "entry-level" now require years of prior experience. The entry point is narrowing from both directions, and the immediate logic is clean: AI handles the tasks that junior employees used to do, so you need fewer of them.

That logic measures the task. It has nothing to say about what the task was incidentally producing: the slow, situated accumulation of judgment that turns a rule-follower into someone who knows when to override the rule. The pausing. The pattern recognition that only forms through years of being wrong in recoverable ways, in proximity to people who've been wrong before you.

We know what happens when you strip those conditions out. When U.S. surgical residency programs capped trainee hours in 2003, 78% of program directors reported diminished competency development. Residents couldn't observe the natural course of disease across shifts. They extended their training through fellowships, at significant cost. The system found a detour.

The UK's dismantling of engineering apprenticeships in the 1980s produced a skills crisis that arrived ten to fifteen years later, precisely because 42-month apprenticeships require foresight that quarterly budgets don't reward.

IBM noticed. The company is tripling entry-level hires in 2026. Its CHRO, Nickle LaMoreaux, was direct about the reasoning:

The specifics are worth lingering on. Their entry-level developers, who might have spent 34 hours a week writing code in 2024, now spend significant time talking to clients, consulting across teams, navigating the kind of ambiguity no coding assistant resolves. They preserved the friction. Changed the tasks around it. They saw the accidental infrastructure and chose to maintain it on purpose.

And then it folds in on itself. The people best positioned to notice what's disappearing are the ones whose judgment was formed by the very thing being optimized away. Recognition requires exactly the kind of pattern-matching that develops slowly, through exposure, in the presence of people who've seen enough to pause over a number that looks fine.