Answers are grounded in this paper and cite their sources.
This white paper sets out an engineering design for transaction monitoring in anti-money laundering (AML) programs, for U.S. banks and for fintech lending programs run with a sponsor bank. It starts from a public enforcement case. The U.S. Justice Department found that one bank left about 92% of its transaction volume unmonitored for more than six years. The paper argues that most monitoring failures happen before any model is involved: products that were never connected to monitoring, data that was never reconciled, alerts that can’t be rebuilt later, and testing that measures new models against filings the old system produced. This is done by Muoro's team. The design works in a fixed order. First, reconcile the data, so the institution can prove which transactions its controls checked. Second, run the scenarios regulators require in a protected lane, where no model can switch them off. Third, add machine learning in a separate lane to rank the optional alert queue and to find patterns the rules miss. Every alert keeps its source events, its known data gaps, and the logic that raised it, so an investigator or examiner can see what the system knew at the time the alert fired. The paper covers: mapping regulatory obligations to engineering requirements the difference between what a bank and a fintech lending program can each see a reference architecture and a typical tech stack a worked example in SQL: three loans repaid from one shared funding account, and how the picture changes when one of those payments is returned how to choose training labels, and why past suspicious activity reports (SARs) are not ground truth testing a new model in shadow mode, and comparing it with the current system at the same investigator capacity where generative AI can help investigators questions examiners and model risk committees will ask a phased rollout with acceptance criteria and rollback steps The worked example and the capacity comparison use synthetic, illustrative figures. This paper is not legal or regulatory advice.
No takes yet. Share an insight, caveat, or question.
Mukul Juneja (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: