Résumé
Expansive clay soils undergo seasonal moisture fluctuations, swelling when wet and shrinking during dry periods. These volumetric changes induce differential ground movements beneath buildings, often resulting in structural damage, particularly in low-rise residential buildings. In France, this phenomenon has become increasingly costly, with average annual insurance claims of 1 billion euros between 2016 and 2021, rising to 3.5 billion euros in 2022 due to severe droughts and heatwaves driven by climate change.Despite its growing impact, damage assessment remains challenging due to the scarcity of empirical data and the localized variability of soil-structure interactions. To address this, we propose a scalable, data-driven methodology that combines expert knowledge and advanced natural language processing to enable empirical risk analysis.This study introduces the Clay Shrink-Swell Damage Severity Scale, constructs a structured database of building and environmental characteristics, and identifies damageability factors based on observed damage levels. A total of 10,325 loss adjustment reports from Generali (2000–2021) were analysed. A subset of 155 reports was manually annotated and scaled to the full dataset using Large Language Models (LLMs) under a Retrieval-Augmented Generation (RAG) framework.The resulting tools provide a solid foundation for post-event assessments, support rapid field diagnostics, and enable the development of personalized prevention strategies. This work contributes to improved vulnerability modelling and more effective risk reduction in a changing climate