Comparison of Methods for Assessing the Coverage of AML Red Flags by Transaction Monitoring Scenarios Using Artificial Intelligence

Financial institutions are required to monitor suspicious patterns of customer behavior and implement appropriate control mechanisms as part of their anti-money laundering obligations. Regulators often describe these risks in the form of AML red flags, which banks translate into transaction monitoring scenarios. This thesis compares methods for assessing whether such scenarios actually cover the relevant red flags. The comparison was conducted on a dataset consisting of 100 red flags and 16 monitoring scenarios, resulting in 1,600 red flag–scenario pairs. AML compliance expert assessment was used as the reference benchmark. Three approaches were compared against this benchmark: a reference scenario method, a method based on a large language model, and a hybrid method combining lexical and semantic scoring. The results show that the reference scenario method reproduced the SME decision profile most closely. The LLM-based method more often identified broader semantic relationships, but also generated more false positive cases. The hybrid method was more reproducible and conservative, but was less effective at capturing implicit domain-specific relationships. The low agreement in positive predictions also indicates that the methods do not assess the relationship between red flags and scenarios in the same way. The thesis concludes that automated methods cannot replace expert assessment, but can support it effectively. Reference scenarios can serve as a stable baseline, the hybrid method as a screening tool, and the LLM-based method as support for semantic assessment and explanation of borderline cases.