Generative AI boosts task-level output in writing, support, consulting and software development, but these gains rarely translate into automatic firm-wide productivity increases; realizing firm-level value requires workflow redesign, skilled workers, data and cloud infrastructure, and robust digital ecosystems — especially in China.

Main Finding

Generative AI can raise productivity at the task level (notably writing, customer support, consulting, and software development), but these gains do not automatically translate into broad or uniform firm‑level productivity improvements. The economic value of GenAI depends on a task-to-firm conversion process shaped by model capability, task fit, human–AI calibration, organizational complementary assets, and regional digital infrastructure. For Chinese firms, measurable gains are most likely when GenAI adoption is embedded in cloud infrastructure, data readiness, skilled labor, workflow redesign, and strong digital ecosystems.

Key Points

• Task-to-firm conversion framework:
  • GenAI is not a standalone productivity shock; its firm-level impact arises from how task-level improvements aggregate given organizational and regional context.
  • Five interacting factors determine realized value: model capability, task fit, human–AI calibration, organizational complementary assets, and regional digital infrastructure.
• Empirical pattern:
  • Consistent task-level productivity gains documented for specific activities: drafting/editing text, routine customer support, consulting-analytics workflows, and parts of software development (e.g., code completion, debugging).
  • Evidence includes randomized experiments and field studies showing time savings and quality improvements on well-defined tasks.
• Limits and risks:
  • Overreliance on GenAI can introduce errors, hallucinations, and misalignment with organizational goals, reducing net gains or creating new costs (supervision, verification).
  • Performance is uneven across tasks and domains; gains are larger when tasks are structured, repetitive, or have clear evaluative feedback.
  • Organizational frictions—poor workflows, lack of training, legacy IT—impede diffusion and scaling of gains.
• China‑specific observations:
  • Productivity benefits are conditional on cloud adoption, data quality/availability, digital skills, and ecosystem support (startups, platforms, talent pipelines).
  • Regional variation and firm heterogeneity (state-owned vs. private, export orientation, firm size) shape adoption speed and benefit capture.

Data & Methods

• Approach:
  • Literature synthesis and conceptual framework building rather than new firm-level empirical analysis.
  • Integrates evidence from multiple strands: AI and ML capability studies, task‑level productivity experiments (RCTs and field experiments), general‑purpose technology (GPT) theory, digital economics, and China‑focused digital transformation literature.
• Types of evidence reviewed:
  • Controlled task experiments measuring time-on-task, output quality, and downstream decisions.
  • Field deployments and case studies in businesses using GenAI tools.
  • Theoretical and empirical work on technology diffusion, complementarities, and organizational capital.
  • Sectoral and regional studies, including China-focused analyses of cloud adoption, data governance, and digital ecosystems.
• Limitations noted:
  • Rapid evolution of models limits the longevity of specific performance claims.
  • Publication and selection biases toward successful interventions and well-resourced firms.
  • Generalizability concerns: many experiments occur in high-income settings or with digitally mature firms; firm-level aggregation remains challenging.

Implications for AI Economics

• Measurement and modeling:
  • Researchers should move beyond treating GenAI as a single exogenous shock; models must incorporate task heterogeneity, complementarity, and organizational frictions.
  • Microdata linking task-level outcomes to firm processes (workflows, IT, management practices) are crucial to estimate true firm‑level productivity effects.
• Policy and investment:
  • Public policy that improves regional digital infrastructure (cloud, broadband), data interoperability, and workforce digital skills will raise the likelihood that GenAI yields broad productivity gains.
  • Policies should also address verification standards, liability for AI-generated errors, and incentives for workflow redesign and training.
• Firm strategy:
  • Firms should prioritize complementary investments (cloud migration, data pipelines, management practices, retraining) and redesign workflows to capture GenAI benefits.
  • Effective human–AI calibration (task allocation, oversight, quality checks) is essential to avoid hazardous overreliance.
• Labor and distributional effects:
  • GenAI amplifies complementarities: it substitutes for routine parts of tasks but complements skilled, evaluative, and integrative labor; distributional outcomes depend on re‑skilling and organizational change.
• Research agenda:
  • More randomized or quasi‑experimental firm‑level studies that track adoption, complementary investments, and long-run output are needed.
  • Comparative work across countries and regions (including more China micro‑evidence) will clarify how infrastructure, institutions, and ecosystems mediate impacts.
  • Models of GPT diffusion should incorporate endogenous organizational adoption and regional spillovers to better predict aggregate productivity dynamics.