HYBRID AUTOMATION OF FINANCIAL MONITORING BASED ON LARGE LANGUAGE MODELS: A REVIEW OF ARCHITECTURES AND CRITERIA FOR FULL COST OF CASE PROCESSING

Abstract

This article examines the relationship between large language model architectures and the economic outcomes of automating anti-money laundering workflows. A critical review of 28 Russian- and English-language scholarly sources compares approaches to sanctions matching, information retrieval, investigation narrative preparation, and analyst interaction. The proposed taxonomy links the case-processing stage, the allocation of work between software and humans, quality criteria, and cost components. A comparative evidence map distinguishes computational acceleration, expert-rated usefulness, changes in task effort, and workflow outcomes. An analytical framework for the full cost per incoming case incorporates repeated activities and independent quality assurance. The analysis explains why fewer model calls and shorter generation times are insufficient to establish lower financial-monitoring expenditure. Conditions for comparing architectures and limitations on transferring published findings to Russian financial institutions are formulated. The results constitute a methodological synthesis of the literature; the return on investment of a particular deployment is not empirically estimated.