AI is compressing the traditional apprenticeship layer in knowledge industries by automating codifiable entry-level tasks and reducing junior hiring throughput. Since late 2022, employment for workers aged 22-25 in AI-exposed occupations has declined 13% relative to older cohorts, primarily through reduced hiring rather than layoffs. The implication: experience formation can no longer rely on volume-based repetition and must be structurally redesigned.

The most comprehensive cross-sector evidence comes from the Stanford Digital Economy Lab’s analysis of millions of ADP payroll records. Since late 2022, employment for workers aged 22–25 in the most AI-exposed occupations has declined 13% relative to older cohorts in the same roles. In absolute terms, entry-level employment in those occupations fell approximately 6% through mid-2025, while employment for workers aged 26 and older grew between 6% and 9%. The employment effects are concentrated in occupations where AI automates task layers rather than augmenting them. Annual salary trends show little divergence by age cohort, indicating adjustment through hiring rather than wage cuts.

The Federal Reserve Bank of Dallas reached a similar conclusion using Current Population Survey data. The employment share of workers aged 20–24 in the most AI-exposed occupations declined from 16.4% to 15.5% between late 2022 and late 2025. The decline was not driven by increased separations into unemployment. Instead, job-finding rates for young labor market entrants fell by more than three percentage points from their 2023 peak. The Dallas Fed analysis states that the pattern does not resemble prior cyclical downturns.

Sector-level data corroborates this aggregate signal.

In consulting, workforce analytics from Revelio Labs show entry-level consultant inflow down 54% year-over-year at mid-2024. Overall consulting job postings declined 26% year-over-year, with the steepest contraction at the base of the staffing pyramid. Senior consultant hiring rose 55% relative to 2020 levels. McKinsey’s global headcount declined from approximately 45,000 to approximately 40,000 between 2023 and 2025, while partner counts remained roughly stable, tightening the junior-to-partner ratio.

In accounting, Big Four firms reduced UK graduate intake between 6% and 29% year-over-year from 2023 to 2024. Accountancy graduate job advertisements declined 44% compared to 2023 levels, exceeding the broader graduate job market decline. These reductions occurred alongside stable or rising partner payouts, indicating cost discipline rather than revenue contraction.

In software engineering, new graduates accounted for approximately 7% of Big Tech hires in 2024, more than 50% below pre-pandemic levels. Indeed postings for entry-level software roles declined roughly 60% over a two-year period. The Stanford analysis shows entry-level software engineers aged 22–25 experiencing employment declines of roughly 20% from their 2022 peak, while older cohorts expanded.

In banking, JPMorgan reported Q3 2024 profits up 12% to $14.4 billion while headcount increased only 1%. Within that flat growth, operations and support roles declined 4% while client-facing roles rose 4%. The firm disclosed measurable AI-driven efficiency gains in software engineering and fraud operations and publicly instructed managers to resist incremental hiring.

In Indian IT services, firms that historically expanded headcount in line with revenue added only 3,910 employees collectively over a recent year, compared with historical quarterly additions often exceeding 10,000. Revenue per employee increased across major providers, indicating a shift from manpower expansion to productivity expansion.

Across these sectors, wage data does not show broad entry-level compression. The labor market signal appears in hiring throughput and staffing composition. Recent college graduate unemployment has risen relative to the overall unemployment rate, while firms report productivity gains, margin stabilization, and flat or declining junior intake.

The consistent pattern across data sources and industries is reduced entry-level hiring intensity in AI-exposed roles. AI-exposed industries are not eliminating entire professions; they are reducing the number of new entrants required to produce a given level of output.

The traditional apprenticeship model embedded training within billable or operational output. Junior employees performed codifiable, repeatable tasks that generated revenue or throughput. That output subsidized supervision by experienced professionals. Over time, repeated exposure converted codified knowledge into tacit judgment.

Generative AI performs an increasing share of this codified task layer. Drafting first-pass legal documents, generating software code, building financial models, and synthesizing research—tasks historically assigned to junior professionals—now fall within the automation frontier.

The Autor–Levy–Murnane task framework distinguishes between routine, rule-based activities and non-routine activities requiring contextual reasoning. Generative AI expands the set of tasks that can be automated beyond traditional routine work, including portions of cognitive knowledge work previously used as training substrate.

The Federal Reserve Bank of Dallas provides an empirical link between task exposure and wage outcomes. Using Bureau of Labor Statistics wage data across more than 200 occupations, Dallas Fed economist Tyler Atkinson calculated an “experience premium”—the percentage difference between experienced and entry-level wages—as a proxy for tacit knowledge intensity. Occupations with low experience premiums, where codified knowledge dominates, show weaker post-2022 wage growth associated with AI exposure. Occupations at the 90th percentile of experience premium show modest wage gains associated with AI exposure. The same technology substitutes for entry-level codified work while complementing experienced professionals.

The structural implication is that codified tasks represent the most advanced work entry-level employees can perform, while those same tasks represent the least distinctive portion of experienced professionals’ roles. When AI automates codified tasks, it reduces the economic rationale for maintaining large junior cohorts.

The Garicano–Rayo apprenticeship framework formalizes this dynamic. Apprenticeship viability depends on the ratio between the productivity of an AI-augmented expert and the standalone output of AI. If AI raises the floor—output without training—nearly as much as it raises the ceiling—output with expert training—then apprenticeship duration and profitability compress. Below a modeled threshold, apprenticeship viability collapses.

Experimental evidence supports the mechanism at the task level. In a large-scale field experiment of customer support agents, generative AI increased average productivity by 14%, with the largest gains accruing to novice workers. New hires reached veteran-level performance in two months rather than eight. In the BCG “jagged frontier” experiment, consultants using GPT-4 completed more tasks faster and at higher quality within the model’s capability frontier, with the largest improvements among lower-performing consultants. Outside the frontier, AI-assisted performance deteriorated on tasks requiring contextual judgment.

Earnings call disclosures align with these findings. Executives in professional services and financial institutions describe AI automating drafting, research, reconciliation, code review, and documentation tasks historically assigned to junior staff. Reported productivity gains of 20–50% are associated with constrained incremental hiring. Revenue per employee increases while entry-level intake contracts.

The structural shift can be summarized as follows: the pyramid model depended on junior repetition financing experience formation. AI absorbs repetition. Experienced judgment is retained and amplified, while the base that historically produced it narrows.

Apprenticeship weakens not because firms devalue experience, but because the codified work that financed experience is increasingly automated.