A theoretical model identifies a sharp, calculable threshold at which AI replaces human employees: once risk‑adjusted AI costs cross the substitution cutoff, firms can switch abruptly to AI, yielding hybrid workplaces and flatter hierarchies while middle managers face elevated automation risk.

Main Finding

The paper introduces the Human–AI Task Allocation (HAT) model and derives the Human–AI Substitution Principle: a precise, risk‑adjusted condition under which replacing a human employee with an AI agent minimizes organizational cost. Under a maintained Human–AI Cost Asymmetry assumption (AI capability costs grow no faster than human skill costs), the model shows that substitution decisions depend jointly on risk-adjusted costs, organizational depth, deployment scale (amortization of AI training cost), skills, and strategic adaptation. Key organizational consequences include abrupt workforce transitions, persistent hybrid human–AI organizations driven by risk heterogeneity, elevated vulnerability of middle managers, threshold-shaped vulnerability for highly skilled workers, and a tendency toward flatter hierarchies with wider spans of control.

Key Points

• Human–AI Cost Asymmetry (central assumption): AI capability/deployment costs increase no faster than human skill acquisition costs. This structural asymmetry is the foundation for substitution results.
• Risk-adjusted decision rule (Human–AI Substitution Principle): a human is optimally replaced when the AI’s risk‑adjusted effective per-task cost is at most the human’s risk‑adjusted effective cost — accounting for reliability, compliance, reputational risk, and for AI, training‑cost amortization across deployments.
• Deployment-scale effects: AI has large fixed training costs and low marginal costs; per-deployment AI cost falls as deployment count grows (T_k / n_k → 0). This amplifies adoption incentives and can generate cascades when scale is reached.
• Hierarchy and spans of control: embedding task allocation in a Garicano‑style hierarchy endogenizes organizational depth and level sizes. AI adoption tends to reduce optimal depth and widen spans of control (flatter organizations).
• Middle-management vulnerability: under plausible parameter regimes the model predicts elevated automation risk for middle managers — their roles are especially exposed by the combination of task codifiability, cost escalation with depth, and AI cost profile.
• High-skill threshold: the vulnerability of highly skilled workers is non-monotonic; substitution depends on a skill threshold shaped by organizational depth, baseline costs, and relative risk differentials.
• Hybrid organizations and persistence: heterogeneity in operational risk and reliability can sustain mixed human–AI staffing without imposing ad hoc minimum-human quotas.
• Strategic dynamics: in a game-theoretic extension, humans invest in upskilling and AI developers invest in capability/risk mitigation. Asymmetric cost structures shape equilibrium dynamics and long-run substitution outcomes.
• Normative scope: the model is an optimization (organizational) framework — it characterizes the cost‑minimizing allocation under stated assumptions, not the empirical timing of AI progress or macroeconomic effects.

Data & Methods

• Model type: Analytical/theoretical model (no empirical dataset). The HAT framework is a tractable mathematical optimization model of task allocation in hierarchical organizations.
• Organizational structure: Hierarchy with levels, spans of control and level sizes derived following the knowledge-hierarchy literature (Garicano-style). Organizational depth is endogenized.
• Agents and tasks: Humans indexed j with skill u_j and AI agents indexed k with capability u_k compete for task allocation. Tasks are allocated to minimize total risk‑adjusted organizational cost.
• Cost structure:
  • Human cost C_ij(u_j): rising with skill (education, training, experience).
  • AI cost decomposed: fixed training/development cost T_k and marginal operational cost M'_k largely independent of capability; per-deployment training cost T_k / n_k declines with deployment scale n_k.
  • Risk enters as additive/adjusting components (e.g., reliability, compliance, reputational) to produce risk‑adjusted effective costs \tilde C.
• Core assumption: Human–AI Cost Asymmetry — AI capability costs scale no faster than human skill costs. This is a maintained modeling premise; alternative assumptions yield different regimes.
• Analytical results: Theorems and corollaries derive substitution thresholds, comparative statics (how substitution responses vary with risk, scale, depth), conditions for hybrid equilibria, organizational flattening results, and middle/high‑skill vulnerability zones. Proofs are in the appendix.
• Strategic extension: a dynamic/game-theoretic layer where humans choose upskilling effort and AI actors choose capability/risk-mitigation investments; Nash equilibria characterize long-run adaptation patterns.

Implications for AI Economics

• Move from technical-capability focus to organizational optimization: Predictions about automation should incorporate risk, hierarchy, and deployment-scale economics — not just technical feasibility.
• Role of scale economies: Because AI fixed costs can be amortized, industry- or firm-level scale will be a key determinant of adoption timing and potential cascades; small organizations may retain humans longer even for technically automatable tasks.
• Risk & governance matter: Operational, regulatory, and reputational risk (modeled as risk‑adjusted cost) can sustain human roles and shape hybrid architectures. Improvements in AI reliability or regulatory changes that lower AI risk can materially accelerate substitution.
• Organizational structure will change: Expect pressures toward flatter hierarchies and wider spans of control where AI reduces the need for intermediate managerial layers; empirical work should test predicted depth reductions and span widening as AI is deployed.
• Middle-management and high-skill impacts are nuanced: Policymakers and firms should not assume only low-skill jobs are at risk. Middle managers may be disproportionately exposed, and high-skill workers face substitution only above/below threshold conditions tied to depth and costs.
• Workforce planning and training policy: Because substitution depends on relative cost slopes and risk, policies that affect human-skill costs (subsidies for training) or raise AI deployment costs (strict liability, compliance burdens) can change substitution outcomes.
• Empirical tests & calibration: The model yields testable hypotheses — e.g., relationship between firm size (or deployment count) and AI adoption rates, correlation of hierarchy depth with substitution patterns, and the persistence of hybrid staffing in high‑risk tasks. These suggest directions for empirical validation and calibration.
• Limitations and scope cautions: Results are conditional on the Human–AI Cost Asymmetry and other modeling assumptions; political, legal, macroeconomic, and innovation‑direction effects are outside scope and can alter real-world outcomes. Empirical calibration and extensions (e.g., multi-task externalities, market competition among firms, behavioral frictions) are recommended next steps.

Suggested next steps for researchers/practitioners - Empirically estimate risk‑adjusted effective costs and AI deployment amortization in industries to test the Substitution Principle. - Study firm-level changes in hierarchy and span of control as AI systems are rolled out. - Use the model to evaluate policy levers (training subsidies, regulation of AI risk) via counterfactuals that alter relative cost slopes or risk parameters.