Adoption of advanced, agent-like AI is linked to measurable productivity and growth gains but disproportionately benefits richer countries; developed economies capture markedly larger growth effects than emerging markets, risking reinforcement of existing global inequalities.

Main Finding

Agentic AI — advanced, autonomous AI systems — has significant potential to raise firm productivity and aggregate growth, but current empirical evidence indicates its economic gains are unevenly distributed. The study finds statistically significant positive effects on firm-level productivity and on growth primarily through a productivity channel, with markedly larger gains in developed economies than in emerging ones. Overall, Agentic AI appears more likely to reinforce existing core–periphery inequalities than to reconfigure the global economic order.

Key Points

• Definition: Agentic AI refers to AI systems capable of autonomous decision-making and task execution (beyond narrow, assistive tools).
• AI Adoption Index: constructed from AI investment, business adoption, and innovation output indicators to proxy diffusion of advanced/agentic capabilities.
• Main estimates:
  • Firm-level productivity: β = 0.18 (p < 0.01).
  • Macroeconomic channeling through productivity: mediation β = 0.35 (p < 0.01).
  • Direct macro effect (not via productivity): β = 0.09 (weaker).
• Heterogeneity:
  • Developed economies: growth impact ≈ 0.33.
  • Emerging economies: growth impact ≈ 0.15.
  • Quantile regressions indicate larger benefits accrue to higher-performing firms/countries (i.e., top quantiles capture disproportionate gains).
• Interpretation: Agentic AI raises productivity and output, but structural conditions (infrastructure, skills, institutions, capital) mediate how much benefit is realized, amplifying existing inequalities.
• Policy recommendations emphasized by the study: invest in digital infrastructure, human capital, and policies that promote inclusive diffusion of advanced AI capabilities.

Data & Methods

• Data sources: World Bank (World Development Indicators; Enterprise Surveys) and OECD AI indicators, covering 2015–2024.
• AI Adoption Index: composite index combining measures of AI-related investment, firm-level adoption rates, and innovation outputs (e.g., patents, R&D indicators) to capture advanced AI diffusion including agentic features.
• Empirical strategy:
  • Panel-data fixed effects regressions to estimate firm-level productivity and country-level growth relationships, controlling for time-invariant unobserved heterogeneity and common time trends.
  • Mediation analysis to decompose total macro effects into productivity-mediated and direct channels (reporting stronger productivity-mediated effects).
  • Quantile regressions to assess distributional heterogeneity across firms/countries.
• Controls and design features: models include standard firm and country controls (e.g., firm size, sector, capital intensity; GDP per capita, institutional controls), year and country/firm fixed effects, and robustness checks (alternative index specifications, lag structures, and subsample analyses).
• Limitations noted by authors: reliance on composite proxy for agentic-capable AI (measurement error risk), observational design (potential endogeneity and omitted variable bias), and varying data coverage across countries and firms.

Implications for AI Economics

• Growth modeling: empirical support that advanced/agentic AI primarily operates via productivity improvements — growth models should explicitly represent productivity-channel mechanisms and complementarities with human capital and infrastructure.
• Distribution and inequality: Agentic AI is likely to exacerbate cross-country and within-country inequalities unless diffusion barriers are addressed; researchers should incorporate heterogeneous adoption and institutional capacity into macro- and trade-oriented models.
• Measurement needs: better, more granular metrics distinguishing agentic capabilities from general AI adoption (e.g., autonomy levels, task scope, deployment contexts) are necessary for improved causal inference.
• Policy design: to realize more equitable benefits, policies should target:
  • Digital infrastructure and connectivity (to lower adoption costs).
  • Education and reskilling (to raise absorptive capacity).
  • Inclusive diffusion strategies (support for SMEs, technology transfer, affordable access).
  • Competition and regulatory frameworks that prevent winner-take-most dynamics.
• Future research directions: stronger causal identification (natural experiments, instruments), firm- and sector-level case studies of agentic deployment, dynamic general-equilibrium analyses of long-run distributional effects, and targeted studies in low- and middle-income countries to understand binding constraints to realizing Agentic AI gains.