An AI-Agent Approach to Constructing Input-Output Production Networks

Understanding production interdependencies is essential for economic modeling, yet existing approaches to constructing large-scale input-output networks are resource-intensive and demand specialized expertise. This study introduces an AI agent-based framework that leverages Large Language Models (LLMs) in conjunction with the Harmonized System (HS) classification of goods to infer and validate production linkages. The method automates the identification of input-output relationships at both the two-digit (HS2) and four-digit (HS4) levels, reducing reliance on manual mapping. The resulting networks are assessed through structural comparison with the World Input-Output Database (WIOD) and statistical analysis of international trade data. Structural validation demonstrates high recall and strong temporal stability, while statistical evaluation confirms that the majority of inferred input-output pairs align with observed trade flows and exhibit positive import-export correlations. These findings indicate that LLMs can effectively reason about and model production processes, providing a scalable and systematic alternative to conventional methods. Overall, this work highlights the potential of LLM-driven approaches to advance the analysis of production structures and offers practical implications for applications in trade analysis, economic modeling, and industrial policy.