Recent advancements in natural language processing techniques have significantly expanded the possibilities for integrating unstructured textual data into a quantitative framework. We use large language models (LLM) to construct credit sentiment indicators, relying on financial news articles sourced from the Dow Jones Factiva database, and incorporate them the Bank of Italy’s ICAS, part of the collateral framework employed by the Eurosystem. We find that LLM-driven sentiment analysis enhances the ability to distinguish solvent firms from insolvent ones, offering a robust approach to incorporating unstructured textual data into credit risk evaluation.