Why Automation Stalls:A Dual-Layer Theory of AI Deployment Constraints

Why does AI automation systematically fall short of its technically feasible potential?This paper develops a dual-layer theory of AI deployment constraints to explain thedeployment gap — the persistent discrepancy between what is technically achievableand what is actually deployed at production scale. Drawing on Massenkoff andMcCrory's (2026) finding that AI's observed task coverage remains a fraction of itstheoretical feasibility, and McKinsey's (2025) documentation that only 6% of AIadopting organizations qualify as high performers with meaningful EBIT impact, thispaper argues that the gap can be explained by two structurally distinct layers ofconstraint.At the market layer, we formalize two legal-political distortions absent from standardfactor-cost comparisons: the Zero Rights Defense Discount (ZRDD), which depresseseffective labor cost below the nominal wage through rights-enforcement failures; andthe Accountability Liability Premium (ALP), which inflates effective AI deploymentcost above the nominal price through current legal-subjectivity arrangements as onemajor determinant within the broader accountability-allocation structure. Together, theseproduce a dual wedge rendering nominally viable substitution economically irrationalunder current institutional configurations.At the organizational layer, we develop Deployment Boundary Theory (DBT),extending Cohen and Levinthal's (1990) absorptive capacity framework with three nonsubstitutable constraints — accountability (R), coordination (C), and dynamicmaintenance (D). Critically, the R constraint is exogenously determined by legalinstitutions rather than internal organizational learning, and may rise rather than fall asdeployment scales. We formally define the deployment gap as the set of task–Gorganization pairs that are nominally substitution-advantaged but institutionallyblocked, and derive three testable predictions: the Legal Channel Hypothesis, theCompetitive Externality Hypothesis, and the Non-Monotonic Labor ProtectionHypothesis. Empirical identification strategies exploiting the phased implementation ofthe EU AI Act (2026–2028) are proposed for each.