Accurate classification of crime narratives is vital for reliable public safety statistics. In Ecuador, the Comisión Especial de Estadística de Seguridad, Justicia, Crimen y Transparencia (CEESJCT) manually categorizes robbery incident reports, which is a time-consuming process. While transformer-based models have shown success in natural language processing tasks, their application to Ecuadorian legal and security texts in the Spanish language remains underexplored. This study addresses this gap by developing an automated classification system using a BERT model tailored to Spanish robbery narratives. Utilizing transfer learning and subsequent fine-tuning on an expanded labeled dataset, the system significantly improves classification performance. Initial transfer learning achieved moderate accuracy (80.5%) but faced difficulties with semantically similar categories. Fine-tuning notably increased minority-class recall (up to 30%) with an improved accuracy (90.3%). The final implementation, which increased the number of categories to 11, achieved 95.5% accuracy with robust and consistent results on both police and judicial narratives. Collaboration with Ecuadorian institutions, including the Fiscalía General del Estado (FGE) and Instituto Nacional de Estadística y Censos (INEC), ensures model credibility.