Machine Learning in Transaction Monitoring: The Prospect of xAI

Explainability needs vary based on who's responsible: The paper finds that requirements for AI explanations in transaction monitoring depend on whether the system is automating decisions (AI decides) or augmenting human judgment (AI assists). This suggests a one-size-fits-all approach to AI transparency won't work in financial crime detection.

Clear explanations can reduce human bias: The research indicates that when AI provides "context-relatable explanations"—understandable in the specific situation rather than just technical outputs—it helps investigators audit decisions more effectively and may prevent their own biases from influencing judgments about suspicious activity.

Banks face competing pressures on AI adoption: Financial institutions must balance regulatory requirements to prevent financial crime with the need for trustworthy, explainable AI systems. The paper suggests this tension requires tailored solutions rather than generic ML implementations.

Current research gap on AI in financial crime is significant: The authors note that "little research is available due to its sensitivity," indicating that despite ML being increasingly proposed for transaction monitoring, there's limited public understanding of how these systems actually affect the process and the people using them.

Different stakeholders need different explanations: The study reveals that various parties involved in transaction monitoring (investigators, auditors, compliance officers) have distinct requirements for understanding why an AI flagged a transaction, emphasizing the importance of context-specific design in responsible AI deployment.