Competition compels firms to replace workers with cheaper AI agents, raising output but permanently reducing labour demand and shifting income away from wages; productivity gains can therefore coincide with higher unemployment and weaker aggregate demand.

Main Finding

AI agents, when modelled as an independent, low‑marginal‑cost factor that directly competes with human labour at the task level, can generate an endogenous, competition‑driven structural break in production: firms must automate any task for which agents become the lower‑cost provider. This drives irreversible, broad task substitution, rising productivity alongside falling labour demand in affected task clusters, and a persistent shift of income from labour to capital/agent rents with attendant aggregate‑demand risks.

Key Points

• Core mechanism: tasks are assigned to the cheapest input. Falling costs and rising productivity of AI agents shift an "automation threshold" so more tasks are fully substituted by agents.
• Competition is decisive: under contestable/competitive markets non‑adopters lose share or exit, making automation an equilibrium outcome rather than a discretionary firm choice.
• Difference from prior literature:
  • Extends task‑based/routine automation models by explicitly treating AI agents as a separate factor with very low marginal costs.
  • Contrasts with skill‑biased and exogenous‑progress models that typically preserve labour as a necessary factor; here tasks can become permanently dispensable for human labour.
• Macro consequences:
  • Productivity gains need not coincide with stable employment; output can rise while aggregate labour demand falls.
  • Functional income distribution shifts toward capital and agent/owner rents, concentrating gains and creating potential shortfalls in aggregate demand.
• Frictions matter for timing but not the long‑run direction: adjustment costs, regulation, market power, or switching costs can delay diffusion, but sufficiently large relative cost advantages and competitive pressure produce widespread automation.
• Task creation is not assumed to automatically offset substitution; whether new task creation compensates remains an empirical question and cannot be taken as a structural stabiliser in the AI‑agent context.

Data & Methods

• Paper type: theoretical review and model development (no primary empirical dataset).
• Approach:
  • Literature synthesis across growth theory, endogenous innovation, skill‑biased/ routine/task‑based models, and recent robotics/AI automation studies.
  • Development of a task‑based macroeconomic modelling framework that:
    • Treats production as a bundle of heterogeneous tasks.
    • Introduces AI agents as an independent factor that competes with labour on a per‑task unit‑cost basis.
    • Defines an automation threshold determined by relative unit costs; decreasing agent costs shift this threshold over time.
    • Analyses macro effects (employment, output, income distribution) and explores dynamics in a two‑sector extension.
  • Emphasis on endogenous adoption driven by competitive selection (references: Grossman & Rossi‑Hansberg; Zeira; Acemoglu & Restrepo; Aghion & Howitt).
• Assumptions & limitations:
  • Strong role for competition (contestability) in driving diffusion—results are framed as tendencies under competitive pressure, not universal across all market structures.
  • Adjustment costs, complementarities, and managerial/organizational frictions are acknowledged but modelled primarily as timing/transition factors rather than permanent inhibitors.
  • Task creation and reallocation are treated cautiously; the framework does not assume automatic, sufficient creation of new labour tasks to offset substitution.
• Empirical implications suggested (for future testing): measure agent unit costs, task automability across occupations, firm/sector competitiveness, timing of adoption, and distributional impacts on wages and rents.

Implications for AI Economics

• Modelling implications:
  • AI agents should be modelled as a distinct production factor with near‑zero marginal cost potentials and the capacity to fully substitute tasks—not merely as productivity multipliers for labour.
  • Competitive market structure must be explicitly incorporated to capture endogenous adoption and selection effects.
  • Task‑level (not occupation‑level) analysis is essential for accurate forecasting of labour displacement and productivity effects.
• Policy and welfare implications:
  • Aggregate demand risks call for redistribution or demand‑support policies (taxation of capital/agent rents, universal basic income, wage subsidies) to offset concentrated income gains.
  • Labour market policies should focus on broader social insurance and demand support as well as retraining, but retraining alone may not suffice if large swathes of tasks become permanently automatable.
  • Market design and regulation can moderate speed and extent of automation (e.g., rules affecting contestability, antitrust, procurement, sectoral protections), influencing transition dynamics.
• Empirical research agenda:
  • Quantify the automation threshold empirically: task‑level cost curves for AI agents vs. human labour across sectors.
  • Measure how market structure (contestability, concentration) modifies adoption rates and wage/employment outcomes.
  • Track the evolution of income shares (labour vs capital vs agent rents) and test links to aggregate demand and macro stability.
  • Investigate the extent and permanence of task creation induced by AI, and whether such creation can offset substitution at aggregate scale.
• Caution for forecasting and policy design:
  • Forecasts that assume labour will be preserved through upskilling or new task creation may be overly optimistic unless they account for competition‑driven, irreversible task substitution.
  • Interventions aimed solely at accelerating productivity without addressing distributional consequences risk exacerbating demand shortfalls and inequality.

Overall, the paper argues that AI agents introduce a qualitatively different automation mechanism—competition‑driven complete task substitution—that requires rethinking theoretical models, empirical measurement and policy responses in AI economics.