AI that lowers coordination costs lets managers supervise more workers, boosting output and compressing hierarchies while widening manager–worker pay gaps; whether gains are shared or concentrated into superstars hinges on who captures the coordination rents.

Main Finding

The paper introduces "agent capital" (KA) — AI that lowers within-firm coordination costs — as a distinct input that reshapes organizational structure. Reducing coordination friction unambiguously raises output and flattens hierarchies, but the distributional consequences depend critically on how KA interacts with managerial skill. If KA is broadly accessible, it functions like general infrastructure with broad-based gains; if KA disproportionately augments top managers, it generates superstar concentration. The model yields five formal propositions and a sharp "regime fork" (general-infrastructure vs elite-complementarity), supported by simulations with heterogeneous workers.

Key Points

• Agent capital (KA): modeled as an input that reduces per-worker coordination cost c(KA) and thereby expands each manager’s span of control Si = 1/ci.
• Coordination-compression functional form: c_i(KA) = c0 / (1 + γ · KA · s_i^β), where
  • c0 = baseline coordination friction,
  • γ = effectiveness of KA at compressing coordination,
  • s_i = manager skill,
  • β = elite-complementarity exponent (controls whether KA is broadly shared or skill-amplifying).
• Endogenous task creation: task frontier T(KA) = T0(1 + δ · KA); δ governs how coordination compression enables new tasks.
• Rent-sharing and wages: team output Yi = A · L_eff,i^α with Cobb–Douglas sharing — manager gets (1−α)Yi, workers split αYi proportional to skills qj.
• Five formal propositions:
• Output Effect: ∂Y/∂KA > 0 (coordination compression raises output holding team allocations fixed).
• Span Expansion: Si strictly increases with KA (hierarchies flatten).
• Manager Demand: Required number of managers falls with KA (fewer layers/managers needed).
• Wage Dispersion: Managerial wage dispersion (Gini) rises with KA for β > 0; the effect strengthens with β.
• Task Frontier: KA can endogenously expand feasible tasks (via δ > 0), absorbing displaced labor and affecting employment.
• Regime fork:
  • General-infrastructure equilibrium (low β): KA is broadly accessible; productivity and employment rise broadly, inequality within firms can fall as employment expands, though manager-worker wage gap widens.
  • Elite-complementarity equilibrium (high β): KA disproportionately benefits top managers; managerial wage concentration (superstars) emerges, producing pronounced inequality.
• Simulation findings:
  • Simulations with heterogeneous workers confirm sharp divergence across parameter regimes.
  • Within simulated firms, overall inequality falls in all regimes as employment expands, but the manager–worker wage gap widens universally.
  • Distributional outcomes hinge on who controls organizational elasticity (the β channel) and on task-creation elasticity δ.
• Empirical context and connections: builds on task-based AI literature (Acemoglu & Restrepo; Agrawal et al.) and organizational models (Garicano; Garicano & Rossi-Hansberg); matches empirical evidence on post-AI hierarchy flattening (Ewens & Giroud; Babina et al.).

Data & Methods

• Analytical model:
  • Team-level production with coordination friction: effective labor L_eff,i = Q_i / (1 + c_i n_i), where Q_i is team quality (sum of worker skills) and n_i is team size.
  • Capacity closure: team sizes bind to technological capacity ni = Si = 1/ci (assumes abundant labor so teams fill capacity).
  • Cobb–Douglas sharing rule: manager residual share (1−α)Yi; workers split αYi proportionally to qj.
  • Elite complementarity parameter β indexes how KA maps into manager-specific effective KA (K_eff,i = KA · s_i^β).
  • Task creation parameter δ governs expansion of the task frontier T(KA).
  • Comparative statics yield five propositions with proofs.
• Simulations:
  • Implemented with heterogeneous workers and a fixed manager pool (example: n = 20 managers).
  • Parameter sweeps across regimes (varying β, γ, δ, etc.) to illustrate regime fork outcomes and distributional dynamics.
  • Simulation details: proportional allocation rules used in many runs; the paper notes potential local non-monotonicities under proportional allocation when low-skill workers admitted dilute effective labor.
• Modeling choices / assumptions to note:
  • Coordination capacity closure (ni = Si) rather than interior optimization.
  • Cobb–Douglas sharing/rent-splitting convention (manager residual).
  • Linear task-creation specification T(KA) = T0(1 + δKA) (claimed not to be essential qualitatively).
  • KA treated as separate from conventional capital and labor (non-rival, scalable).

Implications for AI Economics

• Organizational channel matters: AI’s macro and distributional effects cannot be inferred solely from task-level automation/productivity models; organizational restructuring (spans, layers) is a distinct and powerful channel.
• Ownership and access determine distributional outcomes: whether KA is broadly available (low β) or concentrated (high β) determines whether AI produces broad-based gains or superstar-level concentration. Policy and firm-level governance over AI deployment therefore critically shape inequality.
• Measurement recommendations:
  • Treat coordination-compressing AI as "agent capital" distinct from traditional capital in growth accounting and productivity analysis.
  • Track organizational indicators (span of control, management layers, managerial headcount) alongside task- and occupation-level measures to capture AI’s full effects.
• Testable predictions for empirical work:
  • Firms with higher KA (AI adoption intensity) should show flatter hierarchies, larger spans, and fewer managers per worker.
  • Managerial wage dispersion should rise with measures indicating skill-amplifying AI deployment (signs of high β), while broad-access AI investments should produce smaller increases in managerial dispersion and broader employment gains.
  • Task creation (new occupations, product lines) should correlate with coordination-compressing AI intensity, particularly in activities previously limited by coordination costs.
• Policy levers:
  • Diffuse access to organizational AI (subsidies, open platforms, workforce training) may favor the general-infrastructure regime and reduce superstar concentration.
  • If KA is concentrated, redistribution and targeted labor-market policies may be required to mitigate inequality driven by managerial superstar rents.
• Firm strategy:
  • Firms should treat organizational redesign (flattening, reallocation) as a complement to AI deployment to capture productivity gains.
  • Expect managerial roles to shift (some managerial displacement or role redefinition), with persistent widening of manager–worker wage gaps even when overall employment rises.
• Caveats & directions for future research:
  • Results depend on modeling choices (capacity closure, rent-sharing rule, linear task creation). Alternative wage-setting or endogenous team-size optimization may alter magnitudes or some directions.
  • Empirical validation is needed: measure β-like heterogeneity (who benefits from AI) and estimate γ and δ in field data; use firm-level org charts, payroll, LinkedIn/BLS occupation flows to test predictions.
  • Extensions could endogenize who controls KA, consider multi-firm general equilibrium, or model endogenous wage bargaining.

Summary takeaway: Introducing coordination-compressing AI as a distinct "agent capital" reveals a robust organization-driven mechanism by which AI raises output and flattens hierarchies, but whether those gains are widely shared or concentrated among managerial elites depends critically on whether AI acts as general infrastructure or as an elite-amplifying complement.