Abstract
<jats:p>The digitalization of the financial sector leads to an exponential growth in the volume of unstructured data, which creates serious operational challenges for the compliance risk management system. Traditional automation methods cannot cope with the semantic analysis of complex documents, causing increased costs and workload for employees. The purpose of thearticle is to consider a practice-oriented methodology for the safe implementation of large language models in the compliance processes of a credit institution. Based on the analysis of international experience and systemic risks, an architectural framework based on the principles of an isolated contour, mandatory human control and contextual specialization of models is proposed. The results show that the integration of LLM models in the form of AI agents makes it possible to automate a significant part of routine operations, creating a technological foundation for proactive risk management and enhanced economic security.</jats:p>