In response to the problems of scattered multi-bank data and rigid decision-making in enterprises, this research proposes an intelligent financial assistant system that integrates cross-bank data integration, dynamic risk assessment, investment strategy optimization, and natural language interaction functions. The system employs dual-channel data aggregation and privacy computing techniques to break down account information barriers and comply with the requirements of the Data Security Law. It innovatively builds a “quantitative-qualitative” dual-driven risk assessment model, enabling users to obtain decision support in real time through interactive conversations and circumvent the limitations of traditional static models. Based on the localized GLM4 large model, it realizes the parsing of natural language instructions and the generation of visual reports, enabling users to obtain decision support rapidly through conversations. Research verification indicates that after its application in the case enterprise, both the efficiency of fund management and the returns on idle funds have significantly improved.