Between 2018 and 2024, the share of postings requiring three years of experience or less fell sharply across fields:
| Field | 2018 | 2024 |
|---|---|---|
| Software development | 43% | 28% |
| Data analysis | 35% | 22% |
| Consulting | 41% | 26% |
These numbers describe a subtler shift than headcount reduction. The tasks that used to define early-career work are being quietly removed from junior job descriptions. A study of 62 million workers across 285,000 firms found the mechanism is slower hiring, concentrated in roles where AI handles the tasks that juniors once cut their teeth on. Entry-level postings overall have declined roughly 35 percent since January 2023, with AI-exposed roles dropping faster still.
The work that's disappearing is repetitive, low-stakes, unglamorous. Debugging code. Reconciling data. Reviewing documents against expectations. It was also the primary channel through which professionals learned what "normal" looks like. Those two facts coexisted in the same tasks, and automating one meant automating the other.
An economist who spent two years manually cleaning datasets carries a different intuition about data quality than one who never touched the raw material. That intuition doesn't appear on a résumé. The person who has it often can't fully articulate what they know. A recent theoretical paper formalizes this with an uncomfortable finding: entry-level automation can reduce long-term welfare even without reducing employment, because it disrupts the transmission of tacit knowledge between experienced and junior workers. The learning channel that organizations relied on was so embedded in the structure of work that nobody recognized it as infrastructure until it started thinning.
Some companies are pushing against the current. Cognizant hired 25,000 fresh graduates in 2025 and plans to exceed that in 2026, arguing that early-career talent is more critical in an AI-first world. McKinsey has proposed "apprenticing new employees into co-intelligent workflows." But nobody has demonstrated what that apprenticeship looks like when the foundational tasks have already been absorbed by agents. You can't apprentice someone into work that no longer exists in a form they can touch.
Organizations have always grown their own judgment through a slow, inefficient, sometimes invisible process. People learning by doing work that was just hard enough to be instructive and just routine enough to be recoverable when they got it wrong. The current compression happened task by task, job description by job description, as individual teams made individually rational decisions to let agents handle the repetitive stuff.
Maybe judgment can be built through different channels: simulation, structured exposure, new forms of mentorship we haven't invented yet. Maybe the old apprenticeship model was always inefficient and something better is possible. The experiment is already running, in thousands of organizations simultaneously, and nobody designed it. We'll find out what judgment requires by watching where it fails to appear.