This paper presents a privacy-preserving, federated learning–based AI framework designed to enhance fraud detection and credit risk forecasting for U.S. small and medium-sized enterprises (SMEs). By enabling decentralized model training across multiple financial institutions without transferring raw data, the system ensures compliance with regulations such as GLBA and CCPA. The architecture integrates differential privacy, secure aggregation, and real-time anomaly detection to provide scalable and ethical financial intelligence. Experimental results demonstrate strong model performance, robust privacy guarantees, and practical applicability for under-resourced institutions. This research supports inclusive financial innovation, regulatory readiness, and responsible AI adoption in high-stakes economic infrastructure.