The 80/20 of Human Value & the Career-Ladder Trap

What AI covers, and the residual it doesn’t

The honest starting point is that the technically-automatable share of work is large. McKinsey’s 2023 analysis decomposed some 850 occupations into more than 2,100 detailed work activities and concluded that current generative AI and other technologies “have the potential to automate work activities that absorb 60 to 70 percent of employees’ time today” — up from a prior ~50% estimate, with the jump driven specifically by generative AI’s command of natural language, which underlies activities accounting for 25% of total work time.⁹ A 2025 update restates the figure as roughly 57% of U.S. work hours technically automatable.¹⁵ This is the “80%” in the colloquial 80/20: most of what most jobs consist of, by time, is in principle reachable.

But “reachable in principle” is a statement about tasks, not jobs, and the two McKinsey figures are explicitly technical-potential numbers, “not the inevitable loss of jobs.”¹⁵ What remains after the codifiable majority is stripped out is the high-value residual — the 20% that is judgment, empathy, creativity, accountability for outcomes, and the handling of genuinely novel situations. David Autor names the constraint precisely. Many tasks “people understand tacitly and accomplish effortlessly but for which neither computer programmers nor anyone else can enunciate the explicit ‘rules’ or procedures” — Polanyi’s paradox, after the philosopher who observed in 1966 that “we know more than we can tell.”⁷ The tasks “that have proved most vexing to automate,” Autor writes, “are those demanding flexibility, judgment, and common sense — skills that we understand only tacitly.”⁷

Real usage data confirms the split shows up in behavior, not just in theory. The Anthropic Economic Index, mapping roughly a million Claude conversations to the U.S. Department of Labor’s O*NET task taxonomy, found AI use leaning toward augmentation (57%) over automation (43%), with only ~4% of occupations using AI across at least three-quarters of their tasks and ~36% using it for at least a quarter.¹⁰ The authors’ own reading: “the future might be one where most current jobs evolve rather than disappear.”¹⁰ Autor’s deeper warning is that commentators “tend to overstate the extent of machine substitution for human labor and ignore the strong complementarities” — that focusing only on what is lost “misses a central economic mechanism… raising the value of the tasks that workers uniquely supply.”⁷

Tacit knowledge: why the residual resists codification

The residual is not simply “harder” work; it is structurally different in how it lives in a person. Polanyi’s claim, as carried into labor economics by Autor, is that a great deal of expert performance is tacit — knowable in the doing but not fully articulable in rules. “When we break an egg over the edge of a mixing bowl, identify a distinct species of birds based on a fleeting glimpse, write a persuasive paragraph, or develop a hypothesis to explain a poorly understood phenomenon, we are engaging in tasks that we only tacitly understand how to perform.”⁷ Because the rules cannot be written down, classical rule-based automation could not capture them.

Machine learning partially circumvents this by inferring the input–output mapping from examples rather than from explicit rules — which is why the Brynjolfsson–Li–Raymond support-agent system could, in their words, “distinguish successful behaviors of the top performers, including those they tacitly apply,” and disseminate them down the experience curve; the authors describe these methods as enabling computers “to perform nonroutine tasks that rely on tacit knowledge and experience.”¹ But the circumvention is partial and conditional. It works where abundant labeled examples and clear feedback exist; Autor argues neither machine learning nor environmental control fully resolves Polanyi’s paradox for genuinely judgment-intensive work.⁷ The deeper point for strategy: even when AI learns to imitate tacit behavior from data, the data has to come from humans who possess the tacit skill in the first place. The model is downstream of the expertise it distills — which is precisely why how that expertise gets formed is not a side question.

How expertise is actually formed: learning by doing

The mechanism by which novices become experts is one of the most durable findings in economics, and it is not classroom instruction. Kenneth Arrow’s 1962 paper states it as a generalization “so clear that all schools of thought must accept it”: “Learning is the product of experience. Learning can only take place through the attempt to solve a problem and therefore only takes place during activity.”⁸ Arrow anchored this in hard production data — the airframe “learning curve,” where labor-hours per unit fell as a power function of cumulative output, and the Horndal iron works, where output per man-hour rose about 2% a year for fifteen years with no new investment, a gain “which can only be imputed to learning from experience.”⁸

Arrow added a second, less-quoted regularity that is decisive for the AI question: learning from “repetition of essentially the same problem is subject to sharply diminishing returns… To have steadily increasing performance, then, implies that the stimulus situations must themselves be steadily evolving rather than merely repeating.”⁸ Expertise forms not by doing easy tasks forever but by working a frontier of progressively harder problems. This is the economic substance behind apprenticeship: entry-level work is the supply of “stimulus situations” at the right difficulty. A 2026 Federal Reserve / academic team makes the application explicit, observing that “senior workers began as junior workers who accumulated relevant expertise by executing domain-specific entry-level tasks,” and that those tasks “are not merely low-value work — they are the curriculum through which workers accumulate the human capital that makes them productive later in their careers.”³

Generative AI and skill: the novice-lift finding

The single best-identified result on AI and skill is Brynjolfsson, Li and Raymond’s Generative AI at Work, which studied the staggered rollout of a generative-AI conversational assistant across 5,172 customer-support agents. Access raised productivity — issues resolved per hour — by 15% on average, “with substantial heterogeneity across workers.”¹ The heterogeneity is the story: “less skilled and less experienced workers improve significantly across all productivity measures, including a 30% increase in the number of issues resolved per hour,” while “AI has little effect on the productivity of higher-skilled or more experienced workers,” who even showed “a small decrease in the quality of conversations.”¹ Most strikingly for the apprenticeship question: “treated agents with two months of tenure perform just as well as untreated agents with more than six months of tenure,” and “low-skill agents begin communicating more like high-skill agents.”¹ The tool moves novices down the experience curve faster — and the authors find evidence of durable learning, with gains persisting during software outages when the AI was unavailable.¹

The field-experimental BCG study by Dell’Acqua, Mollick and colleagues — 758 consultants randomized to GPT-4 access or not — reaches a parallel conclusion on a more complex knowledge task, and adds the critical caveat. For 18 realistic tasks inside the AI’s “jagged frontier,” consultants with AI completed 12.2% more tasks, 25.1% faster, with quality more than 40% higher; and the gap compressed — below-average performers gained 43% against their own baseline versus 17% for above-average performers.⁵ But for a task deliberately chosen to sit outside the frontier, AI-equipped consultants were 19 percentage points less likely to reach the correct answer, because AI “produce[s] incorrect, but plausible” outputs that humans over-trust.⁵ The leveling effect is real, and so is its boundary: outside the frontier, the human’s tacit judgment is what averts the confident error — exactly the residual capability that takes experience to build.

The career-ladder trap: substitution vs complementarity over a career

Put the two findings together and the tension is sharp. AI is most useful precisely at the entry level — and entry-level tasks are the curriculum. The 2026 model by Afrouzi, Blanco, Drenik and Hurst formalizes this as a duality operating within a single career. A fall in the price of AI “substitutes for workers on entry-level tasks and complements senior workers in creating the new tasks the occupation performs.”³ The first effect — cheaper automation of entry tasks — can “narrow the pipeline through which workers acquire skill.” The second — AI complementing managers — can let them “expand the task frontier, increasing opportunities for entry-level workers to learn.”³ Which effect dominates is not guaranteed by the technology; it is an equilibrium outcome.

The model’s result is that “economies with high learning capacity admit pairs of stationary equilibria,” and “because automation equilibria exist in pairs, the economy cannot smoothly transition between them. If coordination falls on the zero-learning branch, entry-level work is fully automated” — a human-capital trap.³ The trap exists because the value of an entry-level task to future human-capital formation “[is] not internalized by atomistic workers or firms”; when the resulting skill is general rather than firm-specific, no single firm captures the return on training, so each rationally under-invests — a classic learning externality. The proposed first-best is a policy pair: “a tax on automation profits with a subsidy on frontier-maintenance expenditures.”³ The plain-language version, from the companion Fortune piece: by automating entry roles, firms risk “eviscerating the pipeline of competent senior workers they might need in future, trading short-term cost savings in the present for long-term stability.”¹⁴ The model also cites direct experimental evidence — Shen and Tamkin (2026) found that randomly assigning workers to use AI “reduced their accumulation of skills needed for career progression.”³

Labor-market evidence on the entry rung — read with caution

Is the trap visible in the data yet? The most cited recent evidence is the Stanford Digital Economy Lab’s Canaries in the Coal Mine?, which uses high-frequency administrative payroll data from ADP. Its headline: “early-career workers (ages 22–25) in AI-exposed occupations experienced 16% relative employment declines, controlling for firm-level shocks, while employment for experienced workers remained stable.”⁴ Several features make it more than a raw correlation. The decline is concentrated “in applications of AI that automate work, but not those that most augment it” — the automate/augment split predicted by the theory.⁴ Adjustment runs through employment, not wages, “suggesting possible wage stickiness.”⁴ And the results survive excluding tech firms and remotable occupations, and do not appear in pre-2022 data, including the COVID shock.⁴ Corroborating observational data: SignalFire’s analysis of 650M+ professional profiles finds new graduates were just 7% of big-tech hires in 2024, with new-grad hiring “down 25% from 2023 and over 50% from pre-pandemic levels in 2019,” and “the share of new graduates landing roles at the Magnificent Seven… dropped by more than half since 2022.”¹¹ CNBC reports unemployment for labor-market “new entrants” hit “a nine-year peak,” and that the bachelor’s-vs-high-school unemployment gap for 20–24-year-olds is “the smallest it has been since at least the early 2000s.”¹³

The counter-evidence: precise nulls and epistemic humility

Intellectual honesty requires giving the strongest counterweight full weight, not a token mention. Humlum and Vestergaard link large-scale AI-adoption surveys to administrative records for the entire Danish labor market and find, two years after ChatGPT’s launch, “precise null effects on earnings and recorded hours at both the worker and workplace levels, ruling out effects larger than 2%.“⁶ They examined 11 highly exposed occupations and found “no significant effects in any.”⁶ Most pointed for this paper’s thesis: their data lets them split employment trends by whether firms actually adopted AI, and “while we replicate the pattern of declining early-career employment in exposed occupations in Denmark, our difference-in-differences analysis reveals that the declines are not driven by firms adopting AI chatbots.”⁶ What they do find is reorganization beneath the surface — new tasks in “content generation, AI oversight, and AI integration” — and occupational switching, with adopters who switch seeing earnings grow 12 percentage points more, “though still too few to move average earnings.”⁶

The Yale Budget Lab reaches a similarly cautious verdict using a synthetic difference-in-differences design on U.S. Current Population Survey data: “we generally find no statistically or economically significant effects as of yet” on the employment or wages of AI-exposed occupations.¹² Critically, it flags why it might be missing a real effect rather than disproving one: the CPS is “somewhat underpowered for subgroup analysis, like the 22–27 year old recent college graduates that have been a special focus” — exactly the group the ADP-based Stanford study can resolve.¹² Its appendix shows “roughly half a percentage point increase” in unemployment for exposed occupations in the most recent quarter, “more for the 16–34 year old subsample,” but “both are statistically insignificant as of the first quarter of 2026.”¹² Its honest summary: “Even if no effects are evident in 2022 or 2023, they may yet become evident in 2026 or 2027.”¹²

Read together, these do not cancel out — they locate the disagreement. The studies that find an entry-level effect use the finest data (ADP payroll records that resolve the 22–25 cohort); the studies that find a null use broader aggregates that, by their own admission, may be too coarse to detect a narrow early-career effect, or measure a different country and outcome. The reconciliation the Danish paper offers is the most useful: the early-career decline is real, but it may not be cleanly attributable to firm-level AI adoption — which would mean the labor market is reorganizing for reasons (anticipation, restructuring, the broader downturn) that AI is entangled with but not the sole driver. That is a more defensible reading of the current evidence than either “AI is gutting entry-level jobs” or “nothing is happening.”

Strategic implications

The pieces resolve into a coherent, if uncomfortable, picture. AI is genuinely strong on the codifiable majority of work and genuinely weak on the tacit, high-judgment residual — and it is most immediately useful at the bottom of the experience curve, where it can lift a novice to near-veteran output on in-frontier tasks.¹⁵ The trap is that the entry-level tasks AI most readily absorbs are the same tasks through which humans accumulated the tacit residual in the first place.³⁸ For an individual firm optimizing this quarter, automating the rung is rational; the cost — a thinner senior pipeline years out — is a learning externality the firm does not fully bear, which is exactly why the equilibrium can settle on the bad branch without anyone choosing it.³

Three implications follow that the evidence actually supports. First, treat junior roles as apprenticeship architecture, not just cost centers: if AI does the codifiable drafting, the human-capital value of an entry role now lies in supervising, correcting, and judging AI output — which means redesigning the role around Arrow’s “steadily evolving stimulus situations,” not eliminating it.⁸ Second, the augment/automate choice is the lever that decides which equilibrium you are in: the Stanford data shows declines concentrated where AI automates and growth where it augments, so deliberately deploying AI to expand what juniors can attempt — rather than to remove the attempt — is the firm-level version of “frontier maintenance.”⁴³ Third, hold the causal claims loosely: the strongest displacement evidence is correlational and contested by precise nulls, so betting the workforce strategy on either extreme is unjustified.⁶¹² The defensible posture is to assume the residual human value is real and durable, that it is built by doing, and that an organization which automates away its own apprenticeship is optimizing a number it can see against one it cannot — until the senior bench it needed fails to exist.

References