Mapping algorithmic expropriation: occupational exposure to Agentic AI under the multidimensional ITEA framework v3.0

Abstract The empirical literature on occupational exposure to artificial intelligence (Eloundou et al., 2024; Brynjolfsson, Li and Raymond, 2025; Massenkoff and McCrory, 2026) has converged on metrics calibrated to the generative AI regime, where workers retain workflow control. However, the institutional consensus crystallised in 2025–2026 (Bain Technology Report 2025; McKinsey State of AI 2025; IBM Institute for Business Value, 2025; Microsoft Ignite 2025) identifies Agentic AI as a structurally distinct regime. We argue that this transition is not an incremental extension of capability but a qualitative shift in the mechanism of intellectual capital transfer: under generative AI, worker contribution is diffuse; under Agentic AI, it crystallises into proprietary structural capital. We introduce the term algorithmic expropriation to denote this uncompensated transfer, and we apply the ITEA v2.2 Framework —a multidimensional system of eight indicators applied to 1,016 occupations from O*NET 30.2 (February 2026), with effective analysis on 878 occupations with complete data— to map the exposed population. We operationalise this concept through the Algorithmic Expropriation Index (AEI), built independently on 47,810 O*NET tasks, and convergently validated against the canonical literature benchmark, the AI Occupational Exposure index (AIOE) of Felten, Raj and Seamans (2021). Version 2.2 incorporates the thirteen improvements articulated in the internal evaluation report (PMJ Report Nº 010, April 2026), organised in four blocks: (a) editorial corrections, (b) bibliographic insertions anchored in Acemoglu and Restrepo (2022) and Webb (2020), (c) external convergent validation against AIOE and policy quantification of the top 20, and (d) formal derivation of the Insider–Outsider inversion. The central novelty of v2.2 over v2.1 is item (c): we report r(OAEI v2.1, AIOE) = 0.807 over 769 SOC 6-digit common occupations (p < 10⁻¹⁷⁷), validating the framework's external convergence with the canonical literature metric. This result exceeds the ρ = 0.684 reported by previous works in the author's programme for r(AEI, AIOE), evidencing that OAEI v2.1 captures the dimension of occupational exposure with greater robustness than the auxiliary AEI. The previous v2.0 introduced three central methodological innovations over v1.45: (i) reconceptualisation of the Occupational Resilience Index (IRO) as the Adaptive Resilience Index (IRA), isolating the variance of adaptability orthogonal to formal qualification; (ii) recalibration of the Operational Algorithmic Expropriation Index (OAEI) to a normalised [1, 100] scale via min-max transformation; (iii) explicit articulation of the AEI as convergent validity criterion, on the basis of 47,810 classified tasks. Version 2.1 updated the empirical base to O*NET 30.2 with cross-version validation r = 0.983. Version 2.2 preserves the four accumulated innovations (v2.0, v2.1) and adds external convergent validation against AIOE as fifth innovation. Six findings emerge from the expanded empirical analysis. (1) The convergent validity of ITEA with the AEI remains robust (r = 0.720, n = 923, p < 10⁻¹⁴⁸). (2) External convergent validation against AIOE reaches r = 0.807 over 769 common occupations (p < 10⁻¹⁷⁷), the new central finding of v2.2, superior to intra-corpus benchmarks (ρ = 0.684 in P2 QE for r(AEI, AIOE)). (3) The documented Insider–Outsider inversion is preserved: the maximum exposure quadrant is “high GEE × high IRA” (mean OAEI = 61.80, n = 295). (4) Sectoral concentration is sharp: SOC 15 (Computer and Mathematical, mean AEI = 66.2), SOC 13 (Business and Financial Operations, AEI = 64.9), SOC 19 (Life, Physical, Social Science, AEI = 64.9) and SOC 29 (Healthcare Practitioners, AEI = 64.2). (5) The top 20 OAEI aggregates approximately 2.97 million workers in the US (1.8% of the workforce, per OEWS BLS 2023), prioritising targets for early warning systems and contractual redesign mechanisms. (6) The formal derivation of the Insider–Outsider inversion, articulated in §2.4.bis on the canonical condition (wᴵ − wᴼ) > (F + T)·r of Lindbeck-Snower (1988), shows that under Agentic regime with orchestration cost c_agent → 0 the condition holds unconditionally and qualification ceases to protect. This articulation closes the conceptual residual identified in previous evaluations.