AI deployment alone does not guarantee productivity gains; workforce skills, incentives and organizational flexibility determine whether technical advances translate into value. Managers and policymakers must address learning, capability and incentive structures—not just procure models—to realize AI's economic potential.

Main Finding

Adopting AI in organizations does not automatically yield productivity gains. Realized benefits depend critically on human and environmental moderators—organizational composition, individual baseline capability, practitioner learning dynamics, incentives for fair use, and flexibility of objectives—that determine whether nominal AI capabilities translate into effective productivity improvements. The paper extends an existing partial-equilibrium task-based model (Gries & Naudé, 2022) to make these factors explicit and endogenous.

Key Points

• Core claim: Technical capability (AIT) alone is insufficient; organizational and human factors can substantially attenuate or negate productivity gains from AI.
• Five moderating factors identified:
• Human resource composition — how AI specialists, policy managers, supervisors, and end users are positioned and interact.
• Baseline capability of individuals (κi) — users’ skill to perform tasks unaided and to critically evaluate AI outputs.
• Learning curve of practitioners (λi(τ)) — time and heterogeneity in acquiring effective AI usage; learning-rate heterogeneity can create Matthew effects.
• Incentives for fair use (ωI) — differential access or rewards can encourage misuse, bias amplification, or strategic nonuse.
• Flexibility of objectives and key results (F) — whether organizational goals can be redefined to capture AI-amenable tasks.
• Empirical context: The position draws on field experiments and firm-level evidence showing heterogeneous effects (e.g., large gains for novices in Brynjolfsson et al. 2025; performance degradation when tasks lie outside AI’s capability frontier in Dell’Acqua et al. 2023; concentration of adoption in large productive firms in Calvino & Fontanelli 2023).
• Conceptual contributions:
  • Introduces an organizational effectiveness modifier Ω ∈ [0,1] = ωC · ωI (compositional alignment × incentive alignment) that scales effective AI expertise/utilization (affecting effective bIT).
  • Models capability-adjusted productivity: effective labor productivity ˜γL(z, κi) = γL(z)·g(κi); AI-assisted productivity depends on a capability–task interaction ϕ(z, κi) with different signs inside/outside AI’s reliable task region.
  • Adds temporal learning dynamics: λi(τ) = 1 − e^{−ρi τ}, so effective AI productivity evolves over time: ˜γIT(z, κi, τ) = γIT(z)·ϕ(z, κi)·λi(τ).
  • Endogenous task boundary: organizational flexibility F shifts the effective automation threshold ˜NIT = (1 − F)·(N − 1) + F·NIT.
• Position: Existing economic models that treat organizational/human factors as exogenous miss important channels shaping realized productivity and may overstate AI’s aggregate contribution.

Data & Methods

• Type of paper: Position/synthesis with a formal conceptual extension (no new primary empirical dataset).
• Methods:
  • Literature synthesis of empirical work on AI and productivity, HCI/trust, and education (citing Brynjolfsson et al. 2025; Dell’Acqua et al. 2023; Calvino & Fontanelli 2023; Deng et al. 2025; others).
  • Theoretical extension of Gries & Naudé (2022) partial-equilibrium task model: introduces modifiers and functions (Ω, g(κi), ϕ(z,κi), λi(τ), F) to make organizational and human factors endogenous determinants of effective task-level productivity.
  • Uses stylized functional forms (e.g., exponential learning curve) and threshold definitions (NR for reliable tasks, NIT technical threshold) to illustrate how moderators operate.
• Empirical evidence referenced to motivate components:
  • Field experiments showing heterogeneity in effects by worker skill and task fit.
  • Cross-firm and cross-country adoption patterns showing concentration and role of complementary assets.
• Limitations (acknowledged implicitly): conceptual/qualitative rather than an econometric estimation; specific functional forms and parameter values are illustrative, not estimated from new data.

Implications for AI Economics

• Measurement and macro inference:
  • Productivity accounting should incorporate organizational and human heterogeneity; treating AI ability as a uniform TFP-like factor will misestimate its aggregate impact.
  • The “productivity paradox” can be partly explained by low Ω, mismatched objectives (low F), heterogeneous κi and ρi, and weak incentives for fair/adaptive use.
• Modeling:
  • Economic models of AI should make organizational variables (composition, incentives, flexibility) endogenous or at least explicitly parameterized and estimated, not treated as fixed technology parameters.
  • Task-boundary definitions must allow endogenous movement as organizations re-specify goals to leverage AI.
• Policy and firm practice:
  • Firms: To realize AI gains, invest in complementary assets beyond algorithms—training to raise κi and ρi (reduce learning costs), redesign objectives to capture AI-amenable tasks (increase F), align incentives and access (raise ωI), and reorganize structures to improve compositional alignment (ωC).
  • Education/workforce policy: Focus on baseline capabilities and critical evaluation skills so workers can detect AI errors and integrate AI outputs appropriately.
  • Regulation and governance: Monitoring and incentive systems (to enforce fair use) will matter for realized productivity, especially in high-risk domains (education, hiring, safety-critical work).
• Distributional consequences:
  • Heterogeneous baseline skills and learning rates can amplify inequality (winners capture early productivity gains), and incentive misalignment can erode skill accumulation incentives.
  • Concentration of AI benefits in large, well-resourced firms can persist unless smaller firms gain complementary assets or organizational reforms.
• Research agenda (practical next steps):
  • Empirically estimate Ω, κ-distributions, ρi learning-rate distributions, and F in specific industries via firm-level surveys, field experiments, and randomized controlled trials.
  • Incorporate estimated organizational modifiers into macro/TFP models to refine forecasts of AI’s aggregate contribution.
  • Design and test interventions (training, incentive alignment, organizational redesign) to quantify causal impacts on realized AI-driven productivity.
• Takeaway for AI economists: Accurately predicting and fostering AI-driven productivity requires attention to human and organizational mechanisms; policy and firm-level action on these channels is as important as algorithmic improvements.