Generative AI’s metered delivery can hollow out local demand and trigger self-fulfilling automation cascades: because retained payrolls finance customers, widespread automation shrinks markets and can make remaining human-intensive services unprofitable, yet coordinated expectations or temporary policy bridges can sustain a high-employment human-augmented equilibrium.

Main Finding

Generative-AI changes who bears the fixed cost of automation (upstream model training vs downstream payroll). Because automation reduces a firm’s break-even scale while human-augmented services earn greater incremental surplus from a larger market, downstream firms face a coordination game: payroll-funded demand can make human augmentation privately optimal in a high-employment equilibrium but make automation privately optimal in a low-employment equilibrium. Flexible wages do not generally eliminate this multiplicity — when the wage‑income (demand) effect dominates the fall in a firm’s wage bill, production modes are strategic complements and both a high‑demand human‑augmented and a low‑demand automated equilibrium can persist. Expectations and path-dependence can then generate self-fulfilling cascades of automation or recoveries of augmentation. Welfare analysis shows the high‑employment (human‑augmented) endpoint is locally first-best in the coordination region; optimal policy combines a Pigouvian adoption wedge (to internalize demand spillovers) with a temporary “bridge” (subsidy/coordination device) when the low state is self-sustaining.

Key Points

• Mechanism
  • Downstream firms choose between human‑augmented mode (H) — requiring a minimum team of size ℓ and thus an ongoing payroll cost — and automated mode (A) — which displaces a share d of that team and pays for capability per use.
  • Household spending in the local nontradable market depends on local wage income: M = M̄ + µ w(L)L. Payroll thus generates demand that benefits all local suppliers.
  • A firm’s private gain from augmentation: G(x) = κ M(x) − d ℓ w(L(x)), where κ ≡ b_H − b_A > 0 (human mode’s incremental market return).
  • If G′(x) > 0 (condition (12) in paper), augmentation becomes more attractive the more other firms augment — strategic complementarity via payroll‑driven demand.
• Reverse Big Push
  • Opposite to the classic big‑push: here the scalable upstream technology is available, and it is the presence of local payrolls that expands downstream markets. Widespread automation can shrink demand enough to make remaining teams unprofitable — a self‑reinforcing “reverse big push.”
• Defensive vs offensive automation
  • Defensive automation: attractive because it lowers break‑even scale (d large enough), but human mode still yields larger incremental surplus in large markets (κ>0). This is the central case leading to multiplicity.
  • Offensive automation: automated mode has larger incremental surplus; then automation can dominate across states and no multiplicity arises.
• Role of wages
  • Allowing wages to fall with employment (w = w(L)) is stabilizing (reduces the cost of maintaining teams) but also reduces local demand (through µ wL). Multiplicity survives whenever the demand (wage‑income) effect on G exceeds the direct cost reduction from wage decline.
  • With linear labor supply w(L)=w0+γL the condition and thresholds can be expressed in closed form.
• Dynamics & expectations
  • Myopic Poisson revision: inherited share x converges to x=0 or x=1 depending on whether x is below or above static unstable interior x*.
  • Forward‑looking firms with Poisson revision and switching costs: two polar perfect‑foresight equilibria (all future revisers automate or all augment) can both be supported from the same initial x in a nonempty interval [x_E_H, x_E_A] — hence self‑fulfilling cascades or recoveries are possible.
  • Random public shocks: under Frankel & Pauzner regularity, there is a unique state‑contingent path; in the vanishing‑friction limit, selection follows risk dominance (human augmentation picked if x < 1/2; automation if x > 1/2).
• Welfare & policy
  • In the coordination region, the human‑augmented endpoint is the local first best but decentralized firms ignore positive demand spillovers of payrolls.
  • A state‑contingent Pigouvian wedge can internalize the marginal externality; because welfare is nonconcave in adoption share, a permanent marginal subsidy may be insufficient — temporary coordination subsidies (bridges) may be necessary to push the economy past critical thresholds.
• Quantitative illustration
  • The paper calibrates a transparent parameterization anchored to U.S. professional‑services revenue‑to‑payroll ratios, local employment multipliers, markups, and task‑exposure / realization rates.
  • Central calibration: low/medium‑autonomy uses → unique augmentation; high‑autonomy uses → lie inside coordination region (near risk‑dominance boundary); very high‑autonomy uses → automation dominant.
  • Wage adjustment compresses (but does not always eliminate) the coordination interval.

Data & Methods

• Analytical model:
  • One‑unit mass of differentiated nontradable service sectors; each sector has one incumbent firm.
  • Workers heterogeneous in participation cost c(z); local employment L determines wage w(L)=c(L) (general upward‑sloping supply). Households spend fraction µ of wage income on local services.
  • Two production modes per firm: H (payroll fixed cost ℓ, revenue coefficient b_H) vs A (retains (1−d)ℓ workers, revenue coefficient b_A). Profit flows linearized as π_j(M,w)=b_j M − wage bill.
  • Key object: G(x) = difference in profits between H and A evaluated at aggregate employment implied by x (share of H firms).
  • Condition (12): κ µ [w(L)+L w′(L)] > d ℓ w′(L) ensures G′(x)>0 (strategic complementarity).
• Static results:
  • Classification of equilibria under flexible wages (Proposition 1): unique augmentation region, unique automation region, and coordination (multiplicity) region with an interior unstable x*.
  • Closed forms for special case w(L)=w0+γL and for perfectly elastic labor supply (γ=0) providing cutoffs d1, d2.
• Dynamic results:
  • Poisson arrival of revision opportunities (rate ρ); firms may be myopic (maximize current profit) or forward‑looking (discount r, switching costs FH, FA).
  • Derivation of integral conditions (Φ_A, Φ_H) that determine whether polar automation/augmentation paths are perfect‑foresight equilibria (Proposition 3).
  • Closed forms for expectation thresholds in linearized benchmark; expressions for overlap width and dependence on switching costs, discounting, and adjustment speed.
• Stochastic selection:
  • Public profitability fundamental z follows Brownian motion; with affine G(x,z)=Γ(x−x*)+z, impose Frankel & Pauzner regularity → unique state‑contingent monotone equilibrium. In fast‑revision limit, selection goes by risk‑dominance (Burdzy et al.).
• Quantitative illustration:
  • Parameters set using observed revenue/payroll ratios in professional services, local employment multipliers, markups differences, and task automation exposure mapped to realized displacement; used to locate sectors relative to coordination region and risk boundary.

Implications for AI Economics

• New coordination failure channel: Unlike standard labor‑demand externalities that make automation unambiguously deficient, generative‑AI’s separation of upstream fixed cost and downstream payroll creates a demand‑based coordination problem. Policies focused only on worker retraining or innovation subsidies may miss demand spillovers from payrolls that sustain higher‑value, human‑augmented services.
• Local labor markets matter: Effects of automation depend on local wage responses and the share µ of wage income spent locally. Cities/regions with stronger local multipliers are more likely to sustain augmentation equilibria; weaker local demand amplifies automated cascades.
• Timing and expectations are critical: Identical fundamentals and inherited employment can lead to diverging paths depending on firms’ expectations — hence observable early adopters and credible coordination policies can deterministically affect whether a broader cascade occurs.
• Policy prescriptions
  • Pigouvian adoption wedge: a state‑contingent margin (subsidy on augmentation or tax/fee on automation) that internalizes payroll demand externalities.
  • Temporary coordination bridges: short‑lived payroll subsidies, hiring credits, or guaranteed contracts can be socially valuable to push employment shares past tipping thresholds when the low state is locally self‑sustaining.
  • Targeting: interventions should focus on service lines in the coordination region (high automation potential but positive κ), not on lines where augmentation is uniquely optimal or where automation is clearly dominant.
• Empirical priorities
  • Measure local revenue‑to‑payroll ratios and local spending shares µ by sector and region to identify sectors at risk of self‑fulfilling automation cascades.
  • Estimate κ (human mode incremental return) and d (displacement share) across tasks/occupations to map uses into augmentation/coordination/automation regions.
  • Track distribution of early adopters and expectation formation to anticipate cascades.
• Broader takeaways
  • Upstream availability of high‑capability models does not imply uniform downstream automation — who bears ongoing payroll vs metered usage is crucial.
  • Standard labor‑market stabilizers (wage flexibility) can compress but not necessarily eliminate coordination failures; policy should consider demand externalities explicitly.
  • The model links micro choices about workplace organization (team vs metered use) to macro/local equilibrium selection, providing a framework to study heterogeneous outcomes across sectors and places.
• Limitations / open questions
  • Model abstracts from broader general equilibrium interactions (e.g., reallocation of displaced worker income across regions, owners spending locally, capital accumulation).
  • Assumes linearized profit forms and symmetric sectors; richer heterogeneity and nonlinear demand could change quantitative thresholds.
  • Upstream R&D and pricing of model tokens are taken as given; endogenous upstream pricing or constrained supply could alter adoption incentives.
  • Empirical validation needed: mapping task exposure and realized displacement to the model’s parameters (κ, d, µ, ℓ) is nontrivial but crucial for policy design.

If you’d like, I can: - Produce a one‑page graphical schematic of the mechanism and thresholds; - Translate the calibration steps into a checklist for empirical work (data sources and estimators for κ, d, µ, ℓ, local multipliers); - Simulate comparative statics for a range of parameter values (e.g., how changing µ or γ shifts the coordination region). Which would be most useful?