The latency requirements, fragmented regulations, and ongoing risk of fraud are characteristic of modern payment networks. However, in the majority of production systems, the detection of fraud, compliance and routing remain loosely coupled functions that are not very scalable or resilient in multibank and crossborder contexts. Recent solutions based on AI enhance individual elements of fraud detection or compliance checks but do not do so on a joint basis in terms of risk, regulation, and implementation. This creates a severe void at the orchestration tier where real-time choices need to trade competing operational constraints. This paper presents a single AI-based orchestration framework of payments that involves a combination of federated fraud risk modeling, compliance-conscious admissibility screening, and adaptive routing reinforcement learning. The main innovation here is that the routing decisions are informed by probabilistic fraud and compliance intelligence and not postprocesses. Federated learning can be used to perform cross-bank risk modeling without the need to share raw data, and a lightweight blockchain-based audit layer provides verifiable accountability with minimum overhead. The results of experimental testing on large-scale synthetic transaction streams indicate that the proposed framework reduces the average settlement latency from 1290–1420 ms (baseline routing strategies) to 1050 ms, corresponding to an improvement of approximately 18–26%, while also reducing the route failure rate by more than 20% and compliance violations by approximately 45%, in comparison with rule-based routing, centralized AI fraud systems and fixed compliance baselines. The recall of fraud is 5–9% better than that of independent bank-level learning. These findings show that regulation-sensitive adaptive orchestration can be not only scaled but also necessary to create next-generation payment infrastructures that are resilient and reliable.
Pendyala et al. (Mon,) studied this question.