Execution and Evaluation: A New Occupational Measure and Long-Run Employment Gradients

Measurement

• The paper scores all 19,265 O*NET task statements under a fixed rubric to measure occupation-level execution share and AI capability on the same task universe.Execution describes producing first-pass output, while evaluation supplies verification, certification, responsibility, and judgment; occupations can contain both functions.Source: §1 · p. 2
• Adding execution share to O*NET’s evaluation-activity composite increased R2 by only 0.006 on the common cognitive sample, and execution became insignificant.In the full controlled specification, execution contributed incremental R2 of 0.011 versus 0.058 for the composite and was marginally significant; the paper presents the measure as a reproducible task-level axis, not a large predictive gap.Source: §1 · pp. 2–3

Employment chronology

• Execution-intensive white-collar occupations had lower employment growth in every window since 2012, and equality of execution slopes through 2025 could not be rejected.For 2019–2025, adding two-digit SOC major-group fixed effects reduced the execution coefficient from about −0.70 to −0.14 and removed its significance, indicating a largely between-family association.Source: §1 · p. 3
• The early-2023 AI-capability gradient was negative in every available window and steepened significantly in 2022–2025 versus 2019–2022, but cannot be causally attributed to AI.Capability-exposed occupations already had lower employment growth before generative AI.Source: §1 · p. 3
• Post-2022 employment contraction was concentrated among occupations high in both observed AI use and execution share, but the joint-cell pattern is descriptive, not a multiplicative treatment effect.Usage is measured within the outcome window and cannot be separated from vintage-valid capability as a correlate; neither the continuous interaction nor family-adjusted super-additivity test was significant.Source: §1 · p. 3
• A positive wage coefficient for execution-intensive occupations cautions against reading the employment gradient as a simple inward labor-demand shift, while the CPS entry-age employment test is null.The CPS margin does not capture hiring or labor-force entry.Source: §1 · p. 3

• The roadmap moves from testable predictions and measure validation against O*NET activities to long-run, cross-sectional, post-2022, wage, and seniority facts, then identifies evidence needed for causal and delayed-mechanism tests.Source: §1 · p. 4

2.1 Production and execution-share sorting

• Execution and AI capability are distinct task dimensions: an AI-capable task may still require a human verdict, while execution may exceed current AI capability.The model assigns each role an execution-task fraction and a separate AI-capable share.Source: §2.1 · p. 4
• At a fixed AI-capability share, conditional human employment declines more as execution share rises, in both level and proportional terms.The conditional labor quantity falls for any positive substitution elasticity.Source: §2.1 · pp. 4–5

2.2 From the model to the empirical specifications

• The empirical approximation predicts negative execution and execution-by-capability coefficients, while the capability coefficient alone is unsigned.Output expansion, complementarity, and oversight requirements can offset conditional displacement.Source: §2.2 · p. 6
• Employment-stock associations combine substitution, scale, labor supply, worker composition, and task-mix changes, and do not identify an AI treatment effect.The interaction is harder to identify in a single occupation cross-section; both the continuous interaction and family-adjusted super-additivity test are insignificant.Source: §2.2 · p. 6

2.3 Execution practice and the evaluator pipeline

• Evaluation skill accumulates through human execution practice, while the senior-worker stock depends on entry cohorts from earlier periods.The delayed fall in senior employment and evaluation quality cannot be tested with a sample ending in 2025.Source: §2.3 · pp. 6–7
• If entry-level execution falls, entry employment falls immediately, while the trained senior stock is unaffected for L periods before converging toward a lower steady state.The delayed predictions require post-cohort-lag data linking entry hiring and execution practice to later senior stocks and evaluation performance.Source: §2.3 · p. 7

3.1 Task-level execution and AI-capability shares

• Scoring 19,265 O*NET task statements separates execution from evaluation and assesses AI capability as technical feasibility, not deployment conditions.The execution rubric scores producing or performing tasks above evaluating tasks; capability asks whether a frontier language model with typical tool access can perform a task to a professional standard. Both occupation measures are generated regressors, and reported standard errors do not propagate scoring uncertainty.Source: §3.1 · pp. 7–8
• Using O*NET 26.1 (November 2021, before generative-AI diffusion) as baseline, rebuilding execution share on 30.3 (May 2026) yields occupation-level correlation 0.996; capability versions correlate 0.998.Task-description revisions after 2021 therefore do not account for occupational rankings.Source: §3.1 · p. 8

3.2 Employment, wages, seniority, and samples

• Employment and mean wages come from national cross-industry Occupational Employment data; regressions contain 337 cognitive occupations, while the harmonized long-run panel contains 323 cognitive occupation groups.Source: §3.2 · p. 9

3.3 Separability, routine intensity, and reproducibility

• Within cognitive occupations, execution share correlates −0.35 with 2026-vintage capability and 0.43 with routine intensity, while capability and routine intensity correlate −0.07.Execution retains about 94 percent of its standard deviation after capability is partialled out, and all four execution-capability quadrants are populated. Routine intensity classifies procedure; execution share classifies whether workers produce output or judge it.Source: §3.3 · pp. 9–10

3.4 What the task-level measure adds beyond O*NET evaluation activities

• Against O*NET evaluation, routine, and manual activity composites, execution share has coefficients of −0.058 across 759 occupations and −0.045 within cognitive occupations.The corresponding R2 values are 0.45 and 0.33; the paper characterizes this relation as convergent validity, not evidence of an entirely new latent trait.Source: §3.4 · p. 11
• In the full controlled employment specification (n = 284), execution share is −0.376 (t = −1.95; incremental R2 = 0.011) versus the composite’s +0.104 (t = 4.9; 0.058).Evaluation-intensive occupations grew over 2019–2025, and the execution share’s increment was modest and control-dependent.Source: Table 2 · p. 12, §3.4 · p. 12
• The composite’s coefficients were +0.099 alone and +0.092 conditional on execution share per standard deviation, while execution share falls to −0.020 (t = −1.5) conditional on the composite.The composite’s positive association is not an artifact of suppression by the execution measure.Source: §3.4 · p. 12
• The measure’s distinctive contribution is a common task-level architecture that pairs directly with capability scores, supports joint-cell decomposition and alternative vintages, and transports to databases without O*NET Work Activity ratings.Its main incremental value is task-level alignment with AI capability and portability, not superior predictive fit against the existing O*NET activity composite.Source: §3.4 · p. 12

4.1 The long-run gradients, 2012–2025

• The execution gradient was negative in every window: −0.088 (2012–2016), −0.136 (2016–2019), −0.103 (2019–2022), and −0.136 (2022–2025).Equality was not rejected (p = 0.21), but the minimum detectable slope difference was about 0.089 per year; the two clean pre-pandemic windows establish that the gradient predates generative AI.Source: Table 3 · p. 13, §4.1 · p. 13
• Vintage-valid capability joint slopes were −0.056 (2016–2019), −0.035 (2019–2022), and −0.089 (2022–2025); equality was rejected (p = 0.020).The 2022–2025 slope was 0.054 per year more negative than the pandemic-window slope (p = 0.009), but the design dates the change rather than identifying AI as its cause.Source: §4.1 · p. 14
• The intensive-margin panel omits entrants, but 51 occupations appeared by 2022–2024 against one exit, and entrants resembled the balanced sample on capability and execution shares.Capability shares were 0.49 versus 0.49 and execution shares were 0.61 versus 0.63 for entrants and the balanced sample, respectively.Source: §4.1 · p. 14

4.2 The 2019–2025 cross-section and occupational families

• Controlling for routine intensity and rate exposure, execution and capability coefficients were −0.701 and −0.468, respectively, for 2019–2025 cognitive-occupation employment change.The interest-rate control was insignificant, so that channel did not drive the occupation-level result.Source: §4.2 · p. 14, Table 4 · p. 15
• Adding two-digit SOC major-group fixed effects reduced the execution coefficient from about −0.70 to −0.14 and removed its significance; the association is primarily between occupational families.Capability also became insignificant, so the cross-sectional pattern does not establish differences between execution- and evaluation-heavy occupations within the same family.Source: §4.2 · p. 15
• The controlled execution coefficient remained −0.45 with the rebuilt Acemoglu and Autor (2011) routine measure and −0.738 with vintage-valid Eloundou exposure.The vintage-valid exposure was released in 2023, after the start of the 2019–2022 window in which its slope is reported.Source: §4.2 · p. 15

4.3 Post-2022 concentration in observed usage and execution

• Observed Anthropic usage cannot be treated as predetermined adoption: it comes from one platform during the 2022–2025 outcome window, and employment changes or common demand shocks may affect usage.Source: §4.3 · p. 16
• Usage predicts a 2.6 percent larger decline (t = −3.1), but with earlier Eloundou exposure usage is −1.8 percent (p = 0.089) and capability is −2.6 percent (t = −3.1).The 2026-vintage capability score is insignificant conditional on usage; usage and earlier capability point to nearly the same occupations and cannot be separated as correlates.Source: Table 5 · p. 17, §4.3 · p. 17
• The high-usage, high-execution cell declined by 11.5 percent (t = −5.3), and its coefficient remained significant with controls and family fixed effects (−0.056, p = 0.009).The formal baseline super-additivity contrast was −0.047 (p = 0.14), so the cell pattern does not establish super-additivity.Source: Table 5 · p. 17, §4.3 · pp. 17–18
• Within major groups, usage remained negative (−0.024, t = −2.5), while the execution main effect did not (−0.009, insignificant).The continuous execution-by-usage interaction was also negative and insignificant (−0.011, p = 0.34).Source: §4.3 · pp. 17–18

4.4 Wages and seniority

• The execution coefficient for mean-wage change was 0.102, so execution-intensive occupations had lower employment growth but higher relative mean-wage growth.Mean occupational wages cannot distinguish price changes for fixed workers from changes in who remains employed; employment coefficients describe net occupational change, not an identified inward demand shift.Source: Table 6 · p. 18, §4.4 · p. 18
• The person-level CPS seniority test was null: the entry-age-by-post-2022 interaction was small, insignificant, and slightly positive for execution share, capability, and both jointly.The sample conditions on labor-force participation and does not observe hiring or failure to enter an occupation, so the null rules out large effects only on this narrow margin.Source: §4.4 · pp. 18–19

4.5 Findings and limits

• Table 7 collects the empirical objects above alongside the interpretations they support, distinguishing what the data establish from stated limits.Source: §4.5 · p. 19, Table 7 · p. 19

5 Discussion and Conclusion

• The human-rated O*NET evaluation composite is strongly related to execution and predicts employment growth more strongly, while execution adds modest, control-dependent explanatory power.The measure’s contribution is its task-statement-level definition and direct pairing with capability on the same universe and weights.Source: §5 · p. 19
• The task-level architecture can be reconstructed across vintages, applied to other task corpora, and attached to occupations, firms, postings, or worker histories.It supports future designs using adoption, hiring, worker, or vacancy evidence; the authors also call for human coding to validate the model-based measure.Source: §5 · pp. 19–20
• The execution gradient predates generative AI, shows no detected break, and lies largely between broad occupational families; the post-2022 capability gradient is statistically distinguishable but not causally attributed.The paper says its evidence documents chronology, not an AI-caused execution effect.Source: §5 · p. 20
• The paper does not identify an AI-caused execution effect; stronger evidence would require predetermined or plausibly exogenous adoption and hiring or vacancy flows.The proposed pipeline interpretation remains a hypothesis, and the current sample ends before its predicted later effects can be observed.Source: §5 · p. 20
• The authors’ defensible contribution is the measurement framework and documented chronology, while the longer-run pipeline interpretation remains untested.They propose following entry hiring, accumulated practice, later senior stocks, and downstream evaluation performance.Source: §5 · p. 20

A.2 Proof and interpretation of Proposition 1

• At fixed capability, the proportional decline in human execution demand becomes more negative as execution share rises.The framework signs the execution and interaction coefficients, but leaves the reduced-form capability coefficient to the data.Source: §A.2 · pp. 23–24
• A capability-only exposure measure contains no independent execution share, so it cannot generate an execution-share coefficient conditional on capability without another task attribute.This distinction does not rule out broader task models that include task-specific elasticities, regulation, complementarities, or evaluation tasks.Source: §A.2 · p. 24
• Conditional on output, higher capability lowers human execution demand, but output expansion and evaluation complementarity can offset that effect in observed occupation stocks.Capability operates only through execution and vanishes as e_j approaches zero.Source: §A.2 · p. 24
• Interaction tests have the predicted sign but are imprecise because execution and capability are correlated and their product is highly collinear with the levels.Predetermined adoption variation in hiring or vacancy flows would provide a cleaner test.Source: §A.2 · p. 24
• Because output is held fixed and the oversight residual is assumed common across occupations, employment coefficients describe net occupational changes rather than pure conditional displacement.The text notes that regulation, licensing, certification, and organizational design may make the oversight residual vary.Source: §A.2 · p. 24

A.3 Proof and extended implications of Proposition 2

• Senior mass follows the no-shock path for the first L periods, then converges geometrically to a lower steady state.The post-lag recursion converges at rate 1 − δS to S∞ = µN∞/δS, below µN0/δS.Source: §A.3 · p. 24
• Under Assumption 2, entry-level displacement exceeds senior displacement at a given AI-capable share, with a steeper execution gradient at entry.The corollary states a career-stage gradient; the occupation-level execution score itself cannot test the assumption because it does not vary by seniority.Source: §A.3 · p. 24
• After the cohort lag, lower novice inflow cuts senior headcount and practice-depleted cohorts can lower evaluative skill, reducing evaluation capacity through quantity and quality.Average evaluative skill and downstream outcomes may deteriorate even conditional on the number of senior evaluators.Source: §A.3 · p. 25

A.4 A quantitative structure: uses and limits

• The quantitative structure is a calibrated illustration, not a structural estimate; its entry response is assumed, and it is not a full cohort model.Its per-worker evaluation service omits an evaluation wage, hiring margin, and firm choice over evaluation intensity.Source: §A.4 · pp. 25–26

A.5 Predictions and comparison with alternative mechanisms

• Near-term predictions only partly separate mechanisms: P2 distinguishes career-stage mechanisms from capability-only exposure, while P1 requires execution share to vary beyond knowledge rank or codifiability.A general task model can also generate an execution-share effect; P1 is distinctive relative to commonly used one-dimensional automatability measures.Source: §A.5 · p. 26
• A delayed decline in senior headcount is generic to entry-fed pipelines, but declining evaluative skill per senior and outcomes at fixed evaluator count require the practice-to-skill channel.The channel is conditional: the model does not derive that AI reduces entry-level human execution or that execution practice builds evaluative skill.Source: §A.3 · p. 25, §A.5 · pp. 26–27
• The contribution is a measurable execution–evaluation axis linking removed execution practice to later evaluative skill, not merely the generic possibility of a pipeline.The distinction from alternatives lies in the explicit measure and modeled link from execution practice to later evaluative skill.Source: §A.5 · p. 27
• Removing entry-level execution reduces the practice through which workers acquire evaluative judgment, so P4 concerns human-capital quality rather than only training quantity.Unlike the apprenticeship-viability account, which lowers firms’ private return to training juniors, the execution-evaluation mechanism can apply even if formal training continues while AI takes over judgment-building practice.Source: §A.5 · p. 27
• Ide and Talam`as (2025) produce the same near-term pattern through general equilibrium, while this model differs in delayed predictions for the senior stock and evaluation quality.Firm organization, wages, and compute prices lie outside this paper’s cross-sectional labor-market model.Source: §A.5 · p. 27
• Formal model closures differ in which patterns they predict: capability-only exposure lacks an independent execution-share term, while other accounts predict distinct quantity and quality margins.Knowledge hierarchy predicts neither P3 margin; dynamic tacit knowledge predicts P3a and reaches P3b/P4 only if accumulated skill is evaluative; apprenticeship viability predicts P3a, not evaluation quality at fixed numbers. These comparisons concern stated assumptions, not entire literatures.Source: §A.5 · p. 28, Table 8 · p. 28

B.1 Weighting, batch order, and score uncertainty

• The six-year employment coefficient ranges from −0.70 to −0.74 across four aggregations, with nearly unchanged t-statistics.O*NET task-importance and expected-frequency weights yield occupation scores correlated 0.98 to 0.99 with the baseline among cognitive occupations.Source: §B.1 · p. 28
• Repeated randomized scoring yields a 0.989 correlation and a 2.0-point average difference, so temperature zero is not perfectly deterministic.The 500 audited statements were scored in production-style adjacency, randomized order, and an identical randomized repeat.Source: §B.1 · p. 28
• Adjacency shifts the occupation-level execution share by −0.011, while drift from earlier production scores has a 6.5-point mean absolute difference and +2.1-point mean shift.Grouped and randomized scores correlate 0.95 and differ by 5.8 points on average; adjacency shifts statement scores by −0.9 points.Source: §B.1 · p. 28
• Employment-regression standard errors treat generated scores as fixed, and the audit’s task-level mean absolute error is about 10 points.Repeated-score or errors-in-variables estimates are noted as natural ways to account for scoring uncertainty.Source: §B.1 · pp. 28–29

B.2 Sample flow and unconditional employment relationships

• Table 9 traces the scored O*NET universe to the estimation samples, while Figure 3 plots unconditional six-year employment relationships.The OEWS universe exceeds the matched index count because some occupations lack a scored O*NET counterpart.Source: Table 9 · p. 29, §B.2 · p. 29, §B.1 · p. 29

B.3 Off-diagonal occupations and window-specific routine controls

• Table 10 lists occupations with the largest execution-share versus routine-intensity gaps in both directions.The comparison helps identify occupations where the two measures diverge most sharply.Source: §B.3 · p. 29
• Univariate routine coefficients are −0.016 (p = 0.07) in 2019–2022, −0.031 (p < 0.01) in 2022–2024, and −0.006 in 2024–2025.The comparisons use the balanced cognitive sample with routine intensity (N = 334).Source: §B.3 · p. 29
• Execution coefficients remain negative and significant in all three windows (−0.27, −0.25, and −0.18; all p < 0.05) when routine intensity, execution, and capability enter jointly.Routine intensity is insignificant in the first and last windows and equals −0.022 (p < 0.01) in 2022–2024.Source: §B.3 · p. 30

C.1 Post-2022 thresholds, surface, and family-wise inference

• The high-high coefficient is −0.082 (t = −3.5) at 40th-percentile thresholds and −0.135 (t = −5.2) at 60th-percentile thresholds.The corner contrast grows as comparisons sharpen: −0.115 at the median, −0.143 (t = −3.7) for top-versus-bottom terciles, and −0.187 (t = −3.7) for top-versus-bottom quartiles; these are variations on one descriptive object.Source: §C.1 · p. 30, Figure 4 · p. 31
• The global no-association test across five thresholds yields p = 0.0005, but the high-high cell pattern does not itself test interaction.The observed maximum |t| of 5.33 exceeds every one of 2,000 outcome-permutation draws.Source: §C.1 · p. 30
• The super-additive component is insignificant after family-level dependence and studentization are imposed (exact family-wise p = 0.202).Figure 4 shows a deepest joint-high cell, but its continuous interaction is insignificant (t = −1.3).Source: §C.1 · p. 31, Figure 4 · p. 31
• Office and administrative support accounts for 30 of 85 high-high occupations; dropping SOC 43 leaves a high-high coefficient of −0.075 (t = −3.6).Within the cell, coefficients are −0.192 (t = −5.0) for SOC 43 and −0.075 (t = −3.5) for remaining occupations.Source: §C.1 · p. 31

C.2 Weighting and influence

• Inverse-sampling-variance weighting gives an execution coefficient of −1.03 (t = −4.8) and a 2026 capability coefficient of −0.24 (p = 0.09), similar to employment weighting.The 2019 and 2025 estimates use non-overlapping survey panels, so no covariance term is required for this calculation.Source: §C.2 · p. 31, Table 11 · p. 32
• Leave-one-family-out execution coefficients range from −0.790 to −0.487 and are significant in every replicate, but the within-family fixed-effects estimate remains null.These checks establish influence robustness for the between-family association, not a within-family relationship.Source: §C.2 · pp. 31–32

C.3 The three-window decomposition and sample definition

• Execution slopes are −0.08, −0.13, and −0.12 per year, with no rejected break; retrospective 2026 capability slopes are −0.03, −0.07, and −0.06, also without rejected breaks.For execution, diffusion-versus-pre has p = 0.35 and equality across all three windows has p = 0.61; 2026 capability pairwise and joint p-values range from 0.35 to 0.64.Source: Table 12 · p. 32, §C.3 · p. 32
• Vintage-valid Eloundou slopes are −0.052 (s.e. 0.018), −0.071 (0.023), and −0.123 (0.037) per year across the three windows; the 2024–2025 slope exceeds 2019–2022 by 0.070 (p = 0.041).The slopes are significantly negative in every window; the alternative window partition corroborates the steepening detected in Table 3.Source: §C.3 · pp. 32–33
• Adding 117 healthcare-support and other in-person service occupations changes the sample population and yields a different execution-capability pattern, driven by pandemic-sensitive services.Those occupations had mean 2019–2022 employment change of −5.6 percent, versus −0.5 percent in the ten white-collar majors; this is not a timing result within white-collar work.Source: §C.3 · p. 33

The execution-evaluation rubric

• The measure distinguishes execution—producing an output or carrying out an action—from evaluation—judging an output or situation—and can correlate with routine intensity without collapsing into it.Source: The execution-evaluation rubric · p. 33

The AI-capability rubric

• AI capability is scored separately as technical feasibility of completing a task end to end at professional standard, relative to the model frontier at the scoring date.The rubric excludes legality, licensing, accountability, and cost; execution and capability cross on two distinct axes.Source: The AI-capability rubric · pp. 33–34

Protocol, common to both runs

• Two independently scored audit samples correlate 0.991 with each other and 0.924 with production scores without level bias.Agreement is unweighted on a draw that oversamples hard cases, and the two audit coders' exact model identifiers were not logged.Source: Protocol, common to both runs · p. 34

E CPS Seniority Specifications

• The outcome is employment among labor-force participants in cognitive occupations, so it omits labor-force exit and workers who never enter an occupation.The age comparison of 22–30 versus 35–64 is broader than job seniority or tenure.Source: E CPS Seniority Specifications · p. 35
• The estimated execution-triple standard error implies a minimum detectable effect of roughly 0.008, or 0.8 percentage points, on an outcome with mean 0.976.The design rules out effects of that size or larger on this margin, but is uninformative about smaller effects or entry hiring.Source: E CPS Seniority Specifications · p. 35
• The entry-by-post triple is small, insignificant, and slightly positive, and remains nearly unchanged with occupation-by-entry or entry-by-year fixed effects.The fixed-effects specifications likewise report a small, positive, insignificant triple interaction.Source: E CPS Seniority Specifications · pp. 35–36, Table 15 · p. 36, Table 16 · p. 36
• The specifications use CPS ASEC observations from 2019–2025 and compare entry ages 22–30 with ages 35–64.Estimates use ASEC weights, standardized task measures, and occupation-clustered standard errors.Source: E CPS Seniority Specifications · p. 36

F A Quantitative Scenario Exercise

• The scenario is not an estimate: its decline magnitude is calibrated, and the white-collar sample lacks the required pre-2022 execution trend.The paper identifies timing logic, rather than an auxiliary-model estimate, as the useful output.Source: F A Quantitative Scenario Exercise · p. 37
• The senior evaluation stock stays unchanged until the first affected AI-era cohort reaches seniority, so divergence begins near 2022 + L.The full 2022–2025 practice deficit reaches the senior stock near 2025 + L, or 2031–2040 across the three lags.Source: F A Quantitative Scenario Exercise · p. 37
• With cohort lags L ∈{6, 10, 15}, the full practice deficit becomes visible near 2031, 2035, and 2040, respectively.Changing the lag shifts the onset of decline but not the common long-run asymptote in this scenario.Source: F A Quantitative Scenario Exercise · p. 37, Figure 6 · p. 38
• The proposed sequence is practice falling first, a stable senior evaluation stock until cohorts mature, and later deterioration in evaluation quality.The sequence is conditional on the practice deficit; its dates follow cohort arithmetic and are relatively insensitive to shock magnitude or depreciation.Source: F A Quantitative Scenario Exercise · pp. 37–38