AI coding assistants produce consistent individual-level productivity gains — 20–55% fastertask completion in controlled experiments — yet these gains systematically fail to translateinto organizational-level throughput improvement. This survey synthesizes approximately 40independent empirical studies covering over 500,000 software developers (2023–2026),including randomized controlled trials, large-scale telemetry analyses, organizational surveys,and code quality assessments. The evidence converges on a paradox: organizations adopting AIcoding tools observe increased individual output (more tasks completed, more pull requestsmerged, more code generated) alongside flat or declining delivery performance (stagnantdeployment frequency, longer review cycles, degraded code quality, rising complexity). Weidentify six mechanisms through which individual gains dissipate at organizational scale —code quality degradation, review bottleneck inversion, comprehension debt, specificationbottleneck exposure, team cognition disruption, and automation complacency — anddemonstrate that these mechanisms interact multiplicatively, not additively, producingcompounded dissipation that exceeds any single mechanism’s contribution.Drawing on four independent theoretical traditions — Brooks’s essence–accident distinction,Bainbridge’s ironies of automation, sociotechnical systems theory, and context-contingencytheory — we show that the organizational throughput barrier is overdetermined: each traditionindependently predicts its existence through a different causal mechanism, and all fourmechanisms operate simultaneously in AI-augmented software development. We formalize thisconvergence through a mathematical framework structurally analogous to special relativity,where individual coding velocity v faces an organizational throughput ceiling c determined byessential complexity, and coordination costs grow via a Lorentz-like factor = 1/√(1 − v²/c²). γThis framework subsumes Amdahl’s Law as its low-velocity limit, provides micro-levelinterpretation through Little’s Law and M/M/c queueing theory, and generates fourindependent empirical predictions — all of which find support in the reviewed evidence. Wepresent the first survey that treats the individual-to-organizational productivity gap as thecentral phenomenon requiring explanation and provide a predictive mathematical frameworkfor estimating diminishing returns. The paper concludes with twelve specific open researchproblems.
Sophia Franny Philos (Wed,) studied this question.