Uncovering the risks of digital supply chains: A large language model framework for semantic identification and validation

This study proposes a semantic framework powered by Large Language Models (LLMs) to detect and quantify latent, technology-induced risks embedded in unstructured corporate disclosures under ongoing digital transformation. While existing methods often fail to capture key supply chain risks such as external shocks, operational disruptions, and digital system failures, structured prompts guide GPT-based models to extract standardized and interpretable risk indicators. These indicators are validated through panel regressions and behavioral inventory simulations using the classical (s, S) policy. Empirical results show that LLM-derived measures explain 10 %–15 % more variance in firm-level market volatility, digital investment, and inventory decisions than traditional Bag-of-Words (BoW) methods. External risks significantly predict stock return volatility (β = 0.173) and international revenue share (β = −0.477), while digital risks are associated with higher management costs and increased patent output. Simulation results show that inventory buffers expand and total inventory costs rise by up to 18 % under elevated risk exposure. The findings provide a theoretically grounded and practically viable framework for semantic risk identification, advancing intelligent supply chain analytics.