AI-Augmented Sanctions Screening: Enhancing Accuracy and Latency in Real Time Compliance Systems

Abstract

Real-time response to current sanctions lists is paramount to mitigate downstream risks in the global economy, yet legacy rule-based systems suffer from poor accuracy. This work formulates an AI-augmented sanctions-screening problem, with an embedding-based spatial similarity score as ground truth. Fine-tuning a quantile-regression forest model on a U.S. sanctions-list dataset demonstrates high accuracy for sanctions-detected pairs and low false-positive rates. The methodology is extensible to other detectors; signature-based methods gain new risk-weighting capability. Serving payloads rather than species adds minimal latency for real-time applications. Reduced model-serving overhead supports low-latency operationalization of streaming compliance workloads, accommodating large template databases and continuously evolving, error-prone source data while mitigating adverse bias. Integration of interpretability tests throughout the pipeline enables clear and auditable output control, aligning automated risk assessment closer to expert judgment while preserving speed advantages of rule-based systems.