Greater AI use among Chinese listed firms boosted firm and regional productivity from 2011–2023, and part of the gain runs through green financial innovation; the result links corporate digitalization to greener, higher‑quality economic growth.

Jian, W., Lu, H., Yang, Z., & Zhong, Z. (2026). The sustainability payoff of AI: revisiting TFP in corporate and societal performance. International Review of Financial Analysis 110:104891. DOI: 10.1016/j.irfa.2025.104891 (LSE repository: https://researchonline.lse.ac.uk/id/eprint/130473/)

Main Finding

AI adoption intensity raises productivity at both the firm and regional levels in China (2011–2023). Green financial innovation is a significant positive mediator in the transmission from AI intensity to firm TFP and partially mediates the AI → social TFP pathway. The results are robust to event‑study tests, synthetic controls, and an IV specification.

Key Points

• Positive TFP effects:
  • Firm TFP: AI intensity coefficient ~0.10–0.17 (statistically significant across specifications).
  • Social/regional TFP: AI intensity coefficient ~0.055–0.063 (significant).
• Identification and robustness:
  • Staggered-adoption, intensity-based difference‑in‑differences (DID) with event‑study dynamics.
  • Parallel‑trends tests show no pre‑treatment differential trends.
  • Synthetic Control methods show treated firms diverging upward post-adoption (noted from ~2018).
  • Instrumental variable (degree of local AI development) yields positive IV estimates (TFP_it ≈ 0.1560; RegionTFP_rt ≈ 0.1249), supporting causal interpretation.
• Mechanism:
  • Green financial innovation (authors report measures including log count of green patents / green R&D efficiency) increases after AI deployment and significantly mediates the AI → TFP effect at firm and region levels (partial mediation for social TFP).
• Heterogeneity and controls:
  • Models include firm (or region) and year fixed effects, control for size, age, capital intensity, R&D intensity, financing constraints, profitability, export exposure, and financialization. Standard errors clustered at firm level.
• Data and sample:
  • Chinese A‑share listed firms and matched regional aggregates, 2011–2023.
  • Observations N ≈ 68,903; AI intensity standardized to [0,1], constructed from share of AI positions, AI‑related CAPEX/GPU spending, AI patents, and management textual indicators; adoption defined by threshold (score ≥ 0.5) and intensity multiplied by post‑adoption indicator so exposure = 0 pre‑adoption.

Data & Methods

• Data sources: Firm financial statements, Wind, CSMAR, company reports; regional data from China Statistical Yearbook and provincial/municipal yearbooks; green patent data from company reports / patent records.
• Outcome measures:
  • Firm TFP: Levinsohn–Petrin (LP) residuals (industry‑specific), log transformed.
  • Social TFP: region × industry residual/growth‑accounting based index using constant‑price value added, capital stock, and hours worked.
• Treatment variable:
  • AI_Intensityit = (standardized composite AI score in [0,1]) × Postit (Postit = 1 from first adoption year onward). Ensures zero pre‑adoption exposure.
• Econometric strategy:
  • Intensity‑based staggered DID and event‑study specification.
  • Mediation: two‑step mediation models (AI → GreenInnov; AI & GreenInnov → TFP).
  • Robustness: synthetic control, IV (local AI development), parallel‑trend checks, clustered SEs.
• Key sample statistics: mean AI_Intensity ≈ 0.3643 (SD 0.4812); mean firm TFP ≈ 8.18 (SD 1.03); mean GreenInnov ≈ 0.515 (SD 0.219).

Implications for AI Economics

• Empirical evidence that AI, as a general‑purpose technology, produces measurable productivity gains that propagate from micro (firms) to macro (regional) levels — strengthening arguments for AI investment as a driver of aggregate TFP.
• Demonstrates a concrete transmission channel: AI → improved green financial innovation → higher TFP. This links digitalization and sustainability literatures and suggests green finance institutions amplify AI’s productivity benefits.
• Methodological contributions:
  • Intensity‑scaled, zero‑pre‑exposure DID design offers a way to handle staggered adoption with heterogeneous treatment intensity.
  • Firm–region dual framework helps trace micro→macro aggregation and spillovers.
• Policy relevance:
  • Support policies that combine AI adoption incentives with green finance development (e.g., AI tools for environmental risk assessment, credit allocation for green projects).
  • Regional policies to promote diffusion (data sharing, cross‑firm collaboration) may magnify societal productivity returns.
• Caveats / future research suggested:
  • External validity: sample limited to Chinese listed firms and regions — replication in other countries and non‑listed sectors is needed.
  • Measurement nuances: the paper uses multiple operationalizations of GreenInnov (green patent counts vs. R&D efficiency) and a thresholded adoption indicator (≥0.5) — sensitivity to these design choices merits scrutiny.
  • Remaining endogeneity concerns: IV helps but may not fully resolve all channels (e.g., unobserved local policies). Deeper exploration of channels (labor reallocation, firm routines, product quality) and distributional effects is warranted.
  • Long‑run dynamics: more work on long‑run employment, wage, and sectoral reallocation consequences of AI‑driven greening would complement productivity evidence.

If you want, I can (a) extract the exact regression tables and mediation effect sizes from the published supplement, (b) produce a short slide‑style summary, or (c) propose replication tests / extensions (alternative AI intensity measures, worker outcomes, cross‑country comparison).