The rapid proliferation of Real-Time Payment (RTP) rails — including The Clearing House’s RTP® network and the Federal Reserve’s FedNow® Service — has fundamentally transformed the payments landscape for small and medium-sized businesses (SMBs), enabling instant, irrevocable fund transfers 24/7/365. However, this same immediacy eliminates the settlement windows upon which traditional batch-based fraud detection systems rely, creating a critical security gap. In 2024, RTP processed 343 million transactions valued at $246 billion, a 94% year-over-year increase, while fraud losses on instant rails are projected to exceed $12 billion by 2025. SMBs are disproportionately vulnerable, with Business Email Compromise (BEC) alone accounting for 38% of RTP-based fraud targeting small businesses. This study proposes, evaluates, and validates an AI-Driven Fraud Detection Framework (AI-RTPF) tailored specifically to SMB transaction patterns on real-time rails. Leveraging a hybrid architecture combining Long Short-Term Memory (LSTM) networks, Graph Neural Networks (GNNs), and Isolation Forest anomaly detection — scored and decisioned in under 50 milliseconds — the proposed framework achieves 95.5% recall, 96.0% precision, and an AUC-ROC of 0.978, outperforming all baseline models while reducing false positive rates to 4.2%, down from 18.2% in conventional rule-based systems. Findings demonstrate that ISO 20022 rich data enrichment and behavioral baseline modeling are critical enablers of pre-authorization fraud interception in SMB payment contexts. Implications for banking technology design, regulatory compliance, and SMB financial inclusion are discussed.
Keywords: Artificial Intelligence, FedNow, Fraud Detection, ISO 20022, Small Business, behavioral analytics, graph neural networks, machine learning, payment security, real-time payments (RTP)
