Listed Chinese firms that adopt generative AI strengthen their supply-chain resilience, mainly via innovation, greater organizational investment and downstream customer optimization; however, benefits are partly undermined by rising supplier concentration that creates a trade-off between flexibility and coordination risk.

Main Finding

Generative AI (GAI) adoption materially improves firm-level supply chain resilience (SCR) among Chinese A‑share listed firms (2017–2024). The effect operates mainly through (1) enhanced technological innovation, (2) intensified organizational investments, and (3) reduced downstream customer concentration. However, GAI also reduces upstream supplier concentration, which—because supplier concentration can aid post‑shock recovery—creates a suppression (trade‑off) that partially offsets resilience gains.

Key Points

• Sample & scope: Unbalanced panel of 4,453 Chinese A‑share firms, 27,382 firm‑year observations (2017–2024); finance and real estate excluded.
• Core result: GAI adoption (text‑based measure) has a positive, statistically significant effect on a composite SCR index.
• Positive transmission pathways:
  • Technological innovation (R&D intensity) increases with GAI and partially mediates the GAI → SCR relationship.
  • Comprehensive organizational investment (selling + admin + financial expenses scaled by assets) rises with GAI and partially mediates the effect.
  • Downstream customer concentration declines with GAI; lower customer concentration improves SCR and partially mediates the effect.
• Suppression pathway:
  • GAI adoption is associated with reduced upstream supplier concentration (top‑5 supplier share). Because supplier concentration is positively associated with SCR (via relational assets that speed recovery), this reduction suppresses part of GAI’s net positive effect.
• Interpretation: GAI enhances information processing, scenario simulation, and innovation capacity, but creates a technology–organization fit challenge: flexibility gains (diversification) can erode coordination benefits (relational stability), producing trade‑offs for resilience.
• Robustness: Winsorization, industry/year fixed effects, clustered SEs, and instrumental variable strategies are used to address endogeneity.

Data & Methods

• Data sources: Annual reports (text) from Sina Finance; financials from CSMAR; patents from CNRDS.
• GAI measure: Natural log of (frequency of predefined GAI‑related keywords in annual reports + 1). Keywords include terms such as “artificial intelligence,” “deep learning,” “natural language processing,” “semantic search,” etc.
• Dependent variable (SCR): Composite index constructed via the entropy weight method over five dimensions:
• Resistance capability (e.g., ln(AR / revenue), negative direction),
• Recovery capability (residual of EBIT per employee regressed on firm covariates),
• Operational capability (AP turnover, AR turnover),
• Supply–demand matching (ln |current net inventory − lagged net inventory|, negative),
• Renewal capability (ln( # granted invention patents +1)).
• Mechanism variables:
  • Technological innovation (TI): R&D expenditure / total assets.
  • Upstream supplier concentration (SC): Purchases to top‑5 suppliers / total purchases.
  • Downstream customer concentration (CC): Sales to top‑5 customers / total sales.
  • Comprehensive organizational investment (COI): (Selling + administrative + financial expenses) / total assets.
• Controls: Firm size, age, leverage, revenue growth, quick ratio, operating cash flow, board size, independent director ratio, plus year and industry fixed effects.
• Econometric strategy:
  • Baseline: Two‑way fixed effects regression (firm year panel) with firm‑clustered robust SEs.
  • Endogeneity: Bartik (shift‑share) instrument and lagged province‑by‑industry mean instrumental variables employed to strengthen causal claims.
  • Mediation analysis: Baron & Kenny stepwise regressions to identify partial mediation and suppression effects.
• Sample processing: Exclusions for ST/*ST/suspended firms, single‑firm industries, short panels, missing values; continuous variables winsorized at 1%/99%.

Implications for AI Economics

• Causal evidence of macro‑technology effects: The paper supplies large‑scale, quasi‑causal evidence linking GAI diffusion to firm resilience outcomes—important for economic models of technology adoption that must account for externalities across firm networks.
• Mechanisms matter: Results show that productivity/resilience gains from AI are mediated by (a) R&D/innovation spillovers and (b) organizational investment (coordination, governance, human capital). Economic evaluations of AI should incorporate the costs and timeline of organizational adaptation (the “productivity J‑curve”).
• Network externalities and trade‑offs: GAI can reconfigure buyer–supplier network structures in ways that simultaneously reduce vulnerability (by diversification) and remove relational assets that support recovery. Models of AI’s industry‑level welfare effects should account for these network trade‑offs and heterogeneous effects across upstream vs downstream links.
• Policy and managerial relevance: Policy incentives for AI adoption should be paired with support for organizational change, data governance, and mechanisms that preserve productive supplier relationships. From an economic policy perspective, subsidies or training programs that accelerate the organizational investments accompanying AI may increase social returns.
• Measurement approach: The text‑based measure (keyword frequency) is scalable for large samples but likely noisy (captures disclosure intensity/PR as well as substantive adoption). Future economic work should triangulate such measures with usage logs, implementation depth, or procurement data.
• Directions for further research: Heterogeneous effects across industries, firm sizes, governance regimes, and cross‑country contexts; dynamic (longer‑term) productivity and employment impacts as organizational investments mature; welfare implications of shifts in network structure.

Limitations (brief): keyword disclosure may imperfectly proxy true GAI deployment or intensity; supplier/customer concentration and SCR are measured with financial proxies that may omit nonfinancial relational factors; results are based on Chinese listed firms and may not generalize globally.