A task-by-task model of five professions leaves 57 to 73 of every 100 entry-level hours

The arithmetic in Kolb's own model is worth doing before the staffing conversation. His research on how AI recomposes entry-level roles put five professions through a task-by-task model and found that between 57 and 73 hours of every 100 hours of pre-AI work survived ^[6]. That leaves 27 to 43 hours per 100 gone ^[7]. Those hours had names in the old apprenticeship: pulling comparables, cleaning data, building slides, checking citations, reading cases, updating tickets, reconciling accounts ^[15].

They were doing two things at once. Kolb wrote that the first memo, first model, first document review and first reconciliation "did two jobs at once: they shipped, and they taught" ^[3]. His warning is the load-bearing sentence for anyone holding a hiring plan: "Remove the routine work too casually, and the firm saves hours today while draining tomorrow's pool of people able to verify, decide and take responsibility," he wrote ^[4]. He names what gets short: "The new scarcity is verification capital" ^[5].

The payroll figure needs reading precisely. Kolb cites Stanford researchers using ADP records who find employment for 22- to 25-year-olds in the most AI-exposed occupations sits 19% below where it would have been had it kept pace with their less-exposed peers, while employment for experienced workers held steady ^[8]. That is a gap against a counterfactual, not a measured fall in the number of young people employed. It measures hiring, and nothing in it measures whether any firm's stock of people who can check an answer has thinned. The article names no author or venue for the Stanford analysis and gives no method or profession list for the five-profession model ^[18].

The redesign Kolb proposes moves juniors off first drafts and onto comparing AI output against source material, identifying failure modes, explaining their verification logic and presenting tradeoffs to senior reviewers ^[12]. Read that from the senior's side. Verification that a junior does badly at first has to be checked by someone whose hours are the most expensive in the building, and the source puts no price on that. The board-deck version of this is that AI removes 27 to 43 of every 100 entry-level task hours and the headcount plan follows ^[7]. It is incomplete because the ladder that keeps judgment being taught, from assisted production up to accountable recommendation, runs through manager and partner time ^[13].

One more distinction, because the two clocks here are not the same. A smaller entry cohort is bookable this quarter; the capability gap has no date attached, and Kolb only wrote that AI can remove enough exposure to create one ^[17]. What a firm can check now is narrower than the essay's frame and more useful: whether juniors still meet the failure modes that used to teach them, why a comparable felt wrong and why a clean-looking citation failed ^[16]. Kolb's cheapest version of that check is a short confidence note in which a junior explains why an output should be trusted, revised or rejected ^[14].