Automation can lift GDP and consumption if gains are large and ownership is widespread; but when ownership is concentrated and vulnerable, privately chosen automation overexpands and shifts income to capital and high-skill workers, harming low-wealth households.

Main Finding

Automation generates a demand externality: private firms internalize productivity gains and labor-cost savings but do not internalize how automation redistributes income away from low-wealth, high-MPC households and thereby reduces aggregate consumption, prices, savings, capital, and the stationary wealth distribution. Depending on technology, ownership, and exposure parameters, privately chosen automation can be either (a) excessive and welfare-reducing via "demand-base erosion" or (b) socially beneficial and growth-enhancing via "productivity-led capital growth." The key policy levers are taxes/rebates and the domestic pass-through of automation rents.

Key Points

• Central mechanism (marginal object): automation ⇒ redistribution of income ⇒ change in high‑MPC demand ⇒ a missing marginal term in firms’ automation FOC. The decisive objects are dC/da (how aggregate consumption changes with automation) and dBU/da (how exposed low-skill wage bill changes).
• Modeling approach:
  • Static benchmark clarifies the required policy wedge (tax/subsidy) to implement a government target for automation.
  • Stationary heterogeneous-agent general equilibrium (continuous-time Krusell–Smith / Bewley–Aiyagari style) with incomplete markets and skill heterogeneity determines the total-equilibrium derivative Ca(a), including distributional feedbacks.
  • Firms choose automation a to maximize Π(a) = Z(a)K^αL^{1−α} − wH(a) − ΦA(a) − P(a), private marginal benefit M(a) contains productivity and labor-cost terms but not the demand externality.
• Two regimes arise:
  • Demand-base erosion: automation raises productivity and high-skilled incomes but weakens the consumption base (if exposed workers are low-wealth/high-MPC and ownership of automation rents is concentrated), so private automation can be excessive and lower aggregate consumption, capital, and welfare.
  • Productivity-led capital growth: with large productivity gains, strong high-skill complementarity, low obsolescence, and broad domestic ownership of automation rents, automation raises output, capital, and consumption.
• Fiscal closure and policy incidence matter:
  • A tax (e.g., τa) changes the firm’s FOC by adding a wedge but also shrinks the tax base (since a changes), and revenue is realized only on remaining automation rents.
  • Revenues may be rebated via a kernel bs(k) and some fraction ωT may be lost to administrative frictions; the distributional form of rebates (lump-sum vs progressive) crucially affects outcomes.
• Ownership concentration (θE, fraction of automation rents passed to domestic households) is a critical parameter: low domestic pass-through reduces domestic gains from automation and exacerbates demand erosion.
• Model clarifies why representative-agent models miss important effects: it is not just average wages but who receives labor income and who owns capital/rents that determines aggregate demand and the private/social optimum.

Data & Methods

• Theory:
  • Static model: closed-form expression for private FOC and the policy wedge required to implement a government target for automation.
  • Dynamic model: continuous-time heterogeneous-agent model with HJB (household optimization) and Kolmogorov forward equation (stationary wealth distribution). Final-good firm uses Cobb–Douglas technology; automation affects Hicks‑neutral productivity Z(a), paid-task functions h_s(a), and production-task functions ℓ_s(a). Obsolescence δA(a) captures legacy capital reorganization costs.
  • Equilibrium solved on a finite grid via HJB–KFE finite-difference methods (building on Achdou et al., 2022).
• Calibration / empirical anchors:
  • Skill efficiency ratio eH/eU ≈ 1.67, guided by 2024 BLS median weekly earnings and Pew evidence that AI-exposed jobs pay more.
  • Productivity shifter Z(a) calibrated so baseline ψZ = 0.18 (≈20% full-adoption productivity gain); a productivity-led scenario uses larger ψZ (e.g., 0.40).
  • Task exposure parameters chosen so automation displaces low-skilled paid tasks and complements high-skilled tasks (e.g., hU(a)=exp(−3.20a), hH(a)=exp(0.35a)).
  • Ownership pass-through θE varied (baseline 0.45; low 0.15; high 0.75), informed by Distributional Financial Accounts and TIC (foreign ownership).
  • Other parameters: α=0.36, δ0=0.06, δA=0.25, ρ=0.15, γ=2, ϕ=0.01, κ=0.52, fiscal loss ωT=0.15.
• Computation reports both decentralized automation choices (a_KS) and government targets a_G, and evaluates total derivatives Ca(a) and BU,a(a) by re-solving the stationary HJB–KFE for each automation level.

Implications for AI Economics

• Policy and welfare:
  • Automation policy cannot be evaluated purely on output/productivity gains; the distributional incidence and demand-side feedbacks matter at the margin. Policymakers must consider dC/da and dBU/da when assessing automation incentives.
  • Taxes on automation (or alternative redistributive tools) can internalize the demand externality, but design matters: tax revenue is endogenous to adoption, rebates must be targeted (high-MPC, exposed households) to restore demand, and administrative leakage reduces effectiveness.
  • Broader domestic ownership of automation rents (e.g., equity sharing, employee ownership, sovereign or public ownership of returns) reduces the risk of demand-base erosion and shifts outcomes toward productivity-led gains.
• Empirical observables to distinguish regimes:
  • If automation leads to rising aggregate capital, consumption, and broad-based income gains → productivity-led scenario.
  • If automation raises high-skill earnings but reduces aggregate consumption, real capital, or investment → demand-base erosion (look for rising inequality, concentrated ownership of corporate rents, and contraction in consumption among low-wealth exposed groups).
  • Key micro moments: MPCs of exposed workers, fraction of automation rents owned domestically (and its concentration), changes in saving rates and stationary wealth distribution, obsolescence/reorganization costs.
• Modeling and research directions:
  • Empirical work should measure who loses labor income from automation and who gains asset income from automation rents (ownership incidence), and estimate the marginal consumption response (MPC) of exposed households.
  • Extensions: nonstationary aggregate shocks (AI adoption waves), endogenous task creation, multi-sector generalizations, explicit labor supply or bargaining, and richer models of corporate retention vs distribution of rents.
• Takeaway for applied macro and policy: automation can be socially harmful at the margin even while privately profitable when the demand externality is large and ownership is concentrated. Countermeasures include well-targeted redistributive policies, mechanisms to broaden ownership of automation returns, and accounting for obsolescence costs in welfare calculations.