China's AI pilot regions modestly strengthen corporate supply chains: listed firms in pilot areas show a consistent 0.0177-point uplift in measured resilience (95% CI 0.0074–0.0281), driven by greater absorptive capacity, resource integration and innovation. Gains concentrate in the east, tech-heavy sectors and state-owned firms, and propagate to neighboring areas, suggesting policy-driven AI adoption bolsters resilience but favors already-advantaged firms.

Main Finding

• China’s national AI pilot policies (new-generation AI innovation pilot zones) causally increase enterprise supply chain resilience. Estimated average treatment effect: +0.0177 units (95% CI: [0.0074, 0.0281]) for firms located in pilot regions.
• The effect operates mainly through three firm-level dynamic capabilities: absorptive capacity, resource-integration capability, and innovation capacity. Effects are stronger where regional/firm digital foundations and capital investment are higher.
• The policy effect is heterogeneous: larger in eastern regions, central (core) cities, technology-intensive industries, and state-owned enterprises. Policies also mitigate damage from external shocks (especially in later stages) and produce positive spatial spillovers to non-pilot regions.

Key Points

• Policy evaluated: China’s “New Generation AI Innovation” national pilot zones (first designated in 2019; 18 pilot cities by April 2024).
• Average treatment effect: +0.0177 on the authors’ composite supply-chain-resilience index (95% CI as above).
• Mechanisms (mediators):
  • Absorptive capacity proxied by share of employees with bachelor’s degrees or higher.
  • Resource-integration capability proxied by log(365 / inventory turnover rate).
  • Innovation capacity proxied by number of granted patents.
• Heterogeneity: more pronounced impacts for firms in eastern regions, central cities, tech-intensive industries, and SOEs.
• Shock-mitigation: AI policies reduce firms’ vulnerability to external shocks; the dampening effect grows stronger in later stages of shocks.
• Spatial effects: pilot-region improvements spill over to neighboring/non-pilot regions.
• Comparative methodological claim: Double Machine Learning (DML) with random forest learners provides a more robust identification than conventional DID/fixed-effects when many covariates and nonlinearities exist.

Data & Methods

• Sample: 4,144 A-share listed Chinese firms matched to prefecture-level city data; panel 2016–2023.
• Outcome: A firm-level supply-chain-resilience index constructed from multiple dimensions (resistance, adaptation, recovery) — the paper builds a composite measure (drawing on common practices such as entropy-weighted indices across lifecycle stages).
• Identification strategy:
  • Quasi-experimental policy variation: timing/placement of national AI pilot zones.
  • Estimation: Double Machine Learning (DML) framework using random forest regressors as the machine-learning “nuisance” learners. DML is used to flexibly control high-dimensional covariates and reduce model-specification bias relative to standard DID/fixed-effects.
• Mediator analysis: tested the three dynamic-capability channels (absorptive, resource-integration, innovation) as mediators of the policy → resilience effect.
• Additional analyses: heterogeneity checks (region, city type, industry tech intensity, ownership), external-shock interaction tests (timing/stage effects), and spatial spillover analysis.
• Robustness: the paper emphasizes DML’s advantages in handling many covariates and potential omitted-variable bias; it reports confidence intervals and heterogeneity/mediation results consistent with the main estimate.

Implications for AI Economics

• Policy design
  • Targeted AI pilot programs can produce measurable resilience gains at the firm/supply-chain level; outcomes are larger where digital infrastructure and capital investments are stronger—support complementary investments (broadband, data platforms, firm-level digitalization).
  • Spatial spillovers imply national/regional coordination can amplify returns; policymakers should consider geography when allocating AI pilot status and complementary support.
  • Timing matters for shock response: AI-driven capabilities are particularly valuable in later recovery/adaptation stages, so policies should support both pre-shock preparedness and post-shock recovery tools (e.g., digital twins, logistics scheduling).
• Firm strategy
  • Investments that raise absorptive capacity (skilled labor), resource-integration (data-sharing, platformed logistics), and innovation (R&D, patents) translate AI policy exposure into stronger supply-chain resilience.
  • Firms in lagging regions should prioritize building digital foundations and securing capital to capture policy benefits.
• Methods for research and evaluation
  • DML (with flexible learners like random forests) is a promising approach for policy evaluation in AI economics when treatment assignment and outcomes may be influenced by many correlated, possibly nonlinear covariates.
  • Cross-level designs (matching firm-level outcomes to city/policy variables) are informative for studying technology-policy interactions.
• Broader economic implications
  • AI policy interventions can produce system-level benefits (reduced fragility of supply networks), which strengthens arguments for public support of AI diffusion beyond firm-level productivity gains—i.e., AI policy as an instrument of industrial resilience and national economic security.

Potential caveats (to consider when interpreting results) - Pilot-city designation may still be endogenous to unobserved regional characteristics despite DML adjustments; external validity to non-listed firms or other countries may be limited. - The supply-chain-resilience metric and mediator proxies are operational choices; alternative measurements could affect magnitudes. - Complementary infrastructure and capital are important modifiers—policy alone is not sufficient without enabling conditions.