The Bottom Rung Is Being Sawed Off, and Nobody Has a Plan to Rebuild It

Let me start with a number that should unsettle anyone who thinks about where tomorrow's senior professionals come from. A Harvard study published in mid-2025, tracking 62 million workers across 285,000 U.S. firms¹, found that junior employment at companies adopting generative AI declined by roughly 7.7% relative to non-adopters within six quarters of adoption. Senior employment at those same firms? Unchanged. The researchers, Seyed Hosseini and Guy Lichtinger, coined a term for this: "seniority-biased technological change."

That phrase is worth sitting with. It doesn't mean AI is eliminating jobs broadly. It means AI is selectively removing the bottom rungs of career ladders while leaving the upper rungs intact. And the mechanism isn't dramatic. Companies aren't firing juniors. They're just quietly not posting the positions. As the Harvard researchers noted², the decline is "driven primarily by reductions in hiring, rather than staff being sacked." No pink slips, no headlines. The pipeline just narrows, invisibly.

This finding isn't isolated. A Stanford Digital Economy Lab working paper³ by Erik Brynjolfsson, Bharat Chandar, and Ruyu Chen — using ADP payroll data covering 50 million workers — found a 13% relative decline in employment of early-career workers (ages 22–25) in AI-exposed occupations between late 2022 and mid-2025. The decline persisted after controlling for macroeconomic effects and interest rate changes. And the Burning Glass Institute's "No Country for Young Grads" report⁴ documented sharp drops in the share of entry-level postings (requiring three years or less of experience) in AI-exposed fields: software development fell from 43% to 28%, consulting from 41% to 26%, data analysis from 35% to 22% — all between 2018 and 2024⁵. Total postings in those fields stayed flat or rose. Senior hiring held steady. Companies are hiring the same number of people. They're just skipping the juniors.

So the displacement is real. But the deeper question is what it means. Conventional wisdom treats this as a jobs problem — an unemployment concern to be addressed with retraining programs and career services. I think that framing misses the structural risk entirely. The real danger isn't that today's graduates can't find work. It's that when they're 35, they won't have the experiential foundation to do the work that only humans can do: exercise judgment under uncertainty, manage crises, lead organizations through novel situations.

Here's the mechanism. Cognitive scientist Gary Klein spent decades studying how experts actually make decisions in high-stakes environments — firefighters, military commanders, ICU nurses. His Recognition-Primed Decision model shows that expert judgment isn't built from knowing more rules. It's built from having seen more situations. Senior professionals carry a mental library of cases, anomalies, and near-misses accumulated through direct experience, and they pattern-match against that library in real time. There are no shortcuts to building that library. You build it by doing the work.

The work that AI is now doing instead.

I want to be fair to the optimistic case, because it has a genuine kernel of truth. The strongest counter-argument goes like this: the entry-level work being automated — document review, first-draft financial models, boilerplate code — was often rote, low-feedback drudgery. It was naive repetition, not the deliberate practice that Anders Ericsson showed actually builds expertise. If AI handles the drudgery and junior workers instead spend their time evaluating AI outputs, catching errors, and making higher-order judgments, they might actually develop faster. Every previous technology scare — spreadsheets, Westlaw, search engines — triggered identical "missing rung" warnings, and every time, new developmental pathways emerged.

That argument is historically well-grounded. I take it seriously. And I think it's wrong this time, for a specific reason that distinguishes the current moment from all prior analogies.

When spreadsheets automated manual calculation, the analyst's job remained. It was transformed — you built models differently — but the analyst still defended assumptions to a senior, still received feedback on their reasoning, still participated in the evaluative loop that builds expertise. When Westlaw automated case retrieval, the associate still drafted the brief and got it red-lined by a partner. The sub-task changed; the developmental context survived.

What's happening now is different in kind. Tools like Harvey AI and GitHub Copilot aren't marketed as sub-task helpers. They're marketed as producing what a junior would produce — the entire first draft, the complete document review, the preliminary analysis. When AI produces the output and a senior reviews the AI's work rather than the junior's work, the junior is removed from the feedback loop at precisely the point where domain expertise would form. They become an observer of evaluation rather than a participant whose reasoning is evaluated. That's a fundamentally different developmental relationship.

And the evidence that firms are compensating for this by redesigning training? It's thin. LeadDev's AI Impact Report 2025⁶ found that 38% of engineering leaders agreed that "AI tools have reduced the amount of direct mentoring junior engineers receive from senior engineers." Meanwhile, 54% expect junior hiring to decline⁷ as a long-term result of AI coding tools. The mentorship is contracting at the same time the hiring is contracting. That's not a pipeline being rebuilt in a different shape. That's a pipeline being drained from both ends.

I initially found the medical residency analogy compelling — the ACGME duty hour reforms that capped resident work hours and led to concerns about reduced operative experience. But the research on actual patient outcomes is more nuanced than the alarm suggested. A study published in PMC⁸ found that internists exposed to the 2003 duty hour reforms did not show statistically different rates of patient mortality or length of stay compared to pre-reform cohorts. A JAMA study of the 2011 reforms⁹ found no association with changes in general surgery patient outcomes or resident examination performance. The perception of competence gaps was strong among program directors, but the objective outcome data was largely flat. This is actually an important nuance: expert perception of degradation can diverge from measured reality. It means we should be cautious about partner complaints and managing-partner surveys as evidence of actual skill decline.

But here's what makes me reject the fully optimistic reading: the medical residency case reduced hours while keeping residents in the same environment doing substantially similar work. The current AI transition is reducing the work itself while (in many cases) also reducing the headcount of juniors present in the environment. A 2025 IDC/Deel survey¹⁰ found that 66% of global enterprises plan to cut entry-level hiring due to AI. Employment for developers aged 22–25 dropped nearly 20%¹¹ from its peak in late 2022, while developers over 26 remained stable or grew. When the junior professional isn't even in the building, the ambient learning that optimists count on — corridor conversations, informal mentorship, observing how seniors react to the work — doesn't happen.

The asymmetry of this situation is what keeps me up at night. If I'm wrong and AI-supervised evaluation turns out to build domain expertise just as well as independent production, we lose nothing by having built robust developmental programs alongside AI tools. If the optimists are wrong and we discover in 2033 that a generation of professionals lacks the judgment to lead, rebuilding the apprenticeship infrastructure takes another decade. The cost of being wrong is radically different on each side, and the institutions that should be hedging against the downside mostly aren't.

AWS CEO Matt Garman called replacing juniors with AI¹² "one of the dumbest things I've ever heard." Harvard Business School's Amy Edmondson called cutting entry-level jobs¹³ "short-sighted," noting these positions are "crucial for developing future leaders." Stack Overflow's CEO told the BBC that neglecting junior talent pipelines will create problems companies can't solve later. The people saying this most loudly are the ones closest to the talent pipelines in question.

What should readers watch for? Three specific indicators will tell us whether this is a cyclical adjustment or a structural crisis. First, track entry-level hiring rates in AI-exposed sectors through 2027–2028: if they remain 15%+ below 2019 baselines while overall hiring in those sectors recovers, the structural thesis holds. Second, watch for peer-reviewed longitudinal studies tracking the competence development of 2023–2026 cohort professionals at the 5-year mark — these should start appearing around 2028–2030 and will be definitive. Third, monitor whether professional associations (the ABA, AICPA, CFA Institute) roll out scaled training redesigns or merely publish white papers. Scaled programs by 2028 would suggest institutional adaptation; white papers without implementation would confirm the pipeline is being severed by inertia. I think the third indicator is the most telling, because it captures the gap between knowing the problem and doing anything about it. Right now, the knowing is widespread. The doing is not.