What is it about?
Banks screen millions of payments to identify transactions that may involve sanctioned entities. Most alerts are false alarms, but automatically dismissing them is difficult because even a very small number of missed cases can have serious consequences. We developed a new method that learns from previous screening decisions to identify groups of alerts that can be cleared automatically with a controlled level of risk, while uncertain cases continue to be reviewed by human analysts. Unlike many machine-learning approaches, every decision can be traced back to explicit historical evidence, making the system easier to audit and verify. Tested on 12 months of real-world sanctions-screening data, the method substantially reduced the amount of manual review while maintaining strict safety requirements.
Featured Image
Photo by Towfiqu barbhuiya on Unsplash
Why is it important?
Financial institutions must balance two competing needs: preventing prohibited transactions from slipping through while avoiding an overwhelming amount of manual review. This is especially difficult because the cases that matter most are extremely rare, making conventional machine-learning metrics such as accuracy or F1 poorly suited to the problem. Our work shows that automation can instead be designed around an explicit safety constraint: reducing manual workload only when the estimated risk remains within a predefined limit. Just as importantly, the method is inherently auditable, so decisions can be traced to historical evidence rather than relying on an opaque model. This combination of safety, efficiency and transparency is relevant not only to sanctions screening, but also to other high-stakes applications where rare errors can have serious consequences.
Perspectives
What I find most important about this work is the shift in perspective from asking “How accurate is the model?” to asking “Can we safely automate this decision, and can we demonstrate why?” In highly regulated environments, strong predictive performance alone is not enough: a system must also make its risks measurable, its decisions understandable and its failures traceable. Working on this problem reinforced for me that useful AI in high-stakes settings is often less about replacing human expertise and more about identifying where automation can be introduced responsibly, while preserving human oversight where uncertainty remains.
André Panisson
Intesa Sanpaolo
Read the Original
This page is a summary of: Safety-Constrained Classification under Extreme Imbalance: Application to Payment Filtering, August 2026, ACM (Association for Computing Machinery),
DOI: 10.1145/3770855.3818474.
You can read the full text:
Contributors
The following have contributed to this page
