Firms that signal stronger data-driven decision-making in their disclosures tend to perform better internationally; the gains operate through sustainability-focused value-creation channels and are larger in competitive markets and in firms with significant foreign or state ownership.

Main Finding

Using a novel firm-level DDDM index constructed from annual-report text with AI language models (BERT and ChatGLM2-6B), the authors find that data-driven decision-making (DDDM) is positively associated with international firm performance for Chinese listed firms (2007–2022). They show that this relationship operates through sustainable value creation—across four dimensions (current internal: pollution prevention; future internal: green innovation; current external: sustainability information disclosure; future external: sustainability vision co-creation)—and that the positive effect of DDDM is stronger under greater market competition, higher foreign shareholding, and state ownership.

Key Points

• Research question: Does firm-level adoption of DDDM enhance international firm performance, and if so, through what mechanisms?
• Main empirical result: Higher DDDM is associated with higher international firm performance (baseline specification uses ln(sales) as the outcome).
• Mechanisms: The paper frames and provides micro-level evidence that DDDM increases sustainable value in four quadrants (short-term/long-term × internal/external), and these increases mediate the performance benefit.
  • Current internal → pollution prevention (short-term internal gains)
  • Future internal → green innovation (long-term internal gains)
  • Current external → sustainability information disclosure (short-term external gains)
  • Future external → sustainability vision co-creation (long-term external gains)
• Moderation: The positive DDDM → performance link is amplified by:
  • Higher market competition
  • Greater foreign shareholding
  • State ownership
• Data availability: All relevant data are publicly posted on Zenodo per the paper.

Data & Methods

• Sample: Panel of 2,873 Chinese listed firms spanning 2007–2022.
• Outcome variable: International firm performance proxied by lnsales (log of sales; described as dependent variable in baseline model).
• DDDM measurement (key methodological contribution):
  • Source text: Management Discussion & Analysis (MD&A) sections from annual reports.
  • Text processing: Extraction and cleaning of MD&A sections from ~50,000 annual reports; split into ≈1.13M coherent text units.
  • Annotation: Random sampling of 50,000 paragraphs, manual labeling yielded ~3,000 positive samples (DDDM present) and the remainder negative; sampling/ balancing procedures were used to train classifiers.
  • Models: Supervised text classification using transformer-based models (BERT) and ChatGLM2-6B to score firm-year documents and construct a continuous firm-level DDDM index.
• Empirical strategy:
  • Baseline panel regressions: lnsales_it = α0 + α1 DDDM_it + controls + firm + year + industry + province fixed effects; standard errors clustered at the firm level.
  • Mechanism tests: Regressions of mediator variables (e.g., pollution prevention, green innovation, disclosure measures, co-creation indicators) on DDDM to establish links consistent with mediation logic.
  • Moderation tests: Interaction terms between DDDM and firm characteristics (market competition, foreign shareholding, state ownership) to assess heterogeneous effects.
• Robustness: The paper emphasizes model controls and fixed effects to mitigate confounding; data and code availability are reported (Zenodo). (Specific robustness checks such as IVs or lag structures are not detailed in the provided excerpt.)

Implications for AI Economics

• Measurement innovation: Demonstrates how modern language models can be used to construct scalable, semantically informed firm-level indicators of organizational practices (here, DDDM), improving over simple keyword searches or small-sample surveys. This approach can be adapted to measure other soft/capability constructs in empirical IO and international economics.
• Firm strategy and digitalization: Provides micro-level evidence that adopting data-driven managerial practices confers international-market advantages, mediated by sustainable-value creation. This links digital-capability economics with sustainability and internationalization literatures.
• Policy and organizational design:
  • Policymakers aiming to boost firms’ international competitiveness may leverage programs that lower adoption costs of data analytics/AI (training, subsidies, data infrastructure), particularly in competitive industries and contexts with foreign investors or state ties.
  • Firms should consider embedding DDDM not only for operational gains but also as a lever for sustainability practices and credibility in foreign markets.
• Heterogeneity matters: The amplified effects in competitive markets and under certain ownership structures suggest complementarities between DDDM and market/governance environments—important for targeted industrial or ownership-specific policy.
• Research agenda:
  • Use of LLMs/transformers to measure organizational attributes opens avenues for causal work (e.g., IVs, experiments) to address endogeneity between disclosure/text and actual practices.
  • Cross-country replication is needed to test external validity beyond Chinese listed firms and to investigate institutional moderators.
  • Exploration of whether LLM-based indices capture disclosed rhetoric versus realized operational change (disclosure vs. action) is crucial for interpretation.

Reference: Xu, M., & Lu, B. (2026). The data-driven decision-making, sustainable value creation, and international firm performance: Micro-level evidence based on AI language models. PLoS ONE 21(2): e0340731. DOI: 10.1371/journal.pone.0340731. Data: Zenodo repository (paper provides link).