AI adoption in Chinese firms seldom directly boosts measurable innovation; instead, innovation gains run through capability building—skills, organizational learning and networks—and vary sharply by region, policy intensity and platform ecosystems.

Main Finding

Across 38 empirical firm‑level studies of China’s AI industry (2019–2026), direct effects of AI adoption or digital inputs on firm innovation performance are uncommon. The dominant pathway is indirect: AI-related assets and investments improve innovation primarily by building firm innovation capabilities. Core proximate drivers are organizational learning, knowledge management, human capital, and social/relational capital; many studies model these as mediators (often an innovation-capability layer) rather than as simple covariates. Effects are strongly conditioned by China‑specific ecosystem and institutional features (regional hubs and gaps, policy intensity/design, capital cycles and platform ecosystems, shortages in talent/compute/data quality). The review proposes an integrative, testable framework linking OL–KM–HC–SC → innovation capability → innovation performance and highlights gaps in measurement, causal identification, and multi‑level linkage.

Key Points

• Evidence base and outcomes
  • Final sample: 38 peer‑reviewed, English‑language journal articles (33 WoS + 5 Scopus).
  • Outcome measures: patents (n=15) and survey‑based innovation scales/indices (n=10) are most common; others include green innovation (n=3), product/new‑product outcomes (n=2), and mixed/other (n=8).
  • Publication surge: majority of included studies published in 2025, indicating a rapidly evolving literature.
• Dominant mechanisms
  • AI adoption and digital inputs rarely show robust direct effects on innovation outputs.
  • Recurrent finding: capability building (innovation capability) mediates the relationship between inputs (AI, digital tech, resources) and innovation outcomes.
  • Key capability drivers: organizational learning (OL), knowledge management/sharing (KM), human capital/upskilling (HC), and social capital/networks (SC).
• China‑specific boundary conditions
  • Spatial heterogeneity: innovation concentrated in major AI hubs (regional gaps matter).
  • Policy: intensity and instrument design shape both quantity and direction/quality of innovation (risk of quantity-oriented patenting).
  • Finance and platforms: capital cycles, platform ecosystems, and institutional financing constraints condition firm-level strategies.
  • Resource bottlenecks: limitations in high‑level AI talent, specialized compute, and high‑quality data influence both capabilities and outcomes.
• Methodological profile & limitations
  • Methods: regression/econometrics (50% of studies), with smaller shares using network/patent analytics, DID/quasi‑experimental designs, PLS‑SEM, CB‑SEM, and fsQCA.
  • Data sources: archival patent/databases dominate (65.8%); surveys account for ~24%; few mixed designs.
  • Heterogeneity: inconsistent operationalizations of innovation performance impede comparability; the field is recent so constructs/outcomes are not standardized.
• Review contributions
  • Systematic consolidation of firm‑level evidence and alignment of constructs/outcome families.
  • Proposal of an integrative, testable mechanism framework (OL/KM/HC/SC → innovation capability → innovation performance) with explicit China‑context modifiers.
  • Identification of measurement, design, and multi‑level gaps and a focused research agenda.

Data & Methods (of the review)

• Search & scope
  • Databases: Web of Science Core Collection (primary; search date 2026‑01‑26) and Scopus (supplementary; search date 2026‑02‑08).
  • Time window: 2010–2026; English language; peer‑reviewed journal articles.
  • Search string combined AI terms, innovation/outcome terms, and China context.
• Screening & inclusion
  • PRISMA‑style workflow with predefined eligibility criteria (firm/organizational level, China context, AI industry relevance, innovation performance as focal outcome).
  • Exclusion reason taxonomy applied at full‑text stage (EX‑F1 … EX‑F5).
  • Final included studies: 38 (33 from WoS, 5 additional from Scopus).
• Extraction & synthesis
  • Structured coding: bibliographic info, context, unit of analysis, design, data sources, theoretical lens, constructs and operationalizations, outcomes, findings, robustness notes.
  • Given heterogeneity of measures, synthesis used qualitative thematic (narrative) aggregation rather than meta‑analysis.
  • Thematic grouping: organizational learning (OL), knowledge management (KM), human capital (HC), social capital (SC), innovation capability (IC), and innovation performance (IP).

Implications for AI Economics

• For empirical research
  • Prioritize causal and longitudinal designs to unpack the mediating role of innovation capability (e.g., panel data, natural experiments, difference‑in‑differences, instrumental variables).
  • Standardize outcome measurement: move beyond patent counts alone to validated innovation indices, product commercialization success, and quality‑adjusted measures (including green and service innovation).
  • Adopt multi‑level approaches linking firm capabilities to regional ecosystem and policy variables (cross‑level interactions and moderated mediation).
  • Improve measurement of compute, data quality, platform ties, and talent (these resource constraints are central but often proxied crudely).
• For theory and modeling
  • Incorporate capability formation as the primary mechanism in models of AI investment → innovation (explicitly model OL/KM/HC/SC as state variables).
  • Account for non‑linear and asymmetric effects (e.g., thresholds of compute/talent where returns accelerate; diminishing returns from quantity‑focused policy incentives).
  • Model platform and finance dynamics (capital cycles, platform ecosystems) as system‑level modifiers that shape firm incentives and outcomes.
• For policy and industrial strategy
  • Design policy to promote capability quality, not only quantity: incentives should favor capability building (training, knowledge sharing, R&D quality) and high‑value commercialization rather than raw patent counts.
  • Address spatial and resource bottlenecks: targeted investments in compute infrastructure, data governance/quality, and talent pipelines can raise the marginal return on AI inputs.
  • Monitor and mitigate quantity‑bias: strong policy pushes can produce high patent volumes with unclear quality—use balanced metrics and support absorptive capacity.
  • Strengthen ecosystem coordination: policies supporting platform openness, cross‑firm collaboration, and shared infrastructure (e.g., compute, labeled data) can amplify capability formation across regions.
• For investors and firms
  • Investment in AI should be paired with investments in organizational learning, knowledge management systems, and deliberate human capital strategies to translate AI inputs into innovation outputs.
  • Firms should evaluate platform positions and network ties as strategic assets for accessing complementary capabilities and markets.

Limitations of the review to note - English‑language, WoS/Scopus limited coverage may omit Chinese‑language or gray‑literature studies that could alter findings. - The field’s recency (concentration of studies in 2025) implies evolving constructs and potential publication lag/bias. - Heterogeneity in outcome measures and methods prevents a pooled quantitative meta‑estimate.

Actionable next steps (research agenda highlights) - Test these hypotheses with causal designs: 1) AI adoption → innovation performance is mediated by measured innovation capability (OL/KM/HC/SC). 2) Regional policy intensity and platform ecosystems moderate mediation strength. 3) Minimum thresholds of compute/talent are necessary for positive innovation returns to AI inputs. - Build standardized, multi‑dimensional innovation performance metrics (quality + commercialization + sustainability). - Conduct cross‑language synthesis including Chinese publications to broaden evidence and validate China‑specific boundary conditions.