AI in anti money laundering

Most transaction monitoring still runs on rules written by hand: amounts above a threshold, transfers to certain countries, many cash deposits in a short time. Rules are easy to explain, but many of the alerts they raise turn out to be harmless, and every alert costs an analyst's time. Google Cloud quotes the UN Office on Drugs and Crime estimate that 2 to 5 percent of global GDP, up to USD 2 trillion, is laundered each year.

Google's AML AI shows the model approach. It trains on the bank's own core banking data and its history of suspicious activity, produces monthly risk scores for retail and commercial customers with an explanation for each, and can replace or complement legacy monitoring. HSBC reported that the system found two to four times as much suspicious activity with 60 percent fewer alerts.

Money laundering moves through networks of accounts, so graph methods that look at who pays whom find patterns a score per customer misses. A recent study on AI in anti money laundering adds retrieval over a knowledge graph to support customer due diligence, federated learning so banks can train on shared patterns without pooling customer data, and human review built into the workflow so analysts can follow and correct what the system proposes.

Generative AI helps at the end of the chain, where an analyst writes up a case. A draft narrative still needs the analyst's judgment before it becomes a report to the Financial Intelligence Unit. Transaction monitoring in AML covers how scenarios are tuned and what happens to an alert.

Section 25h of the German Banking Act requires banks to run data processing systems that detect suspicious transactions and business relationships. The law does not prescribe rules or models, but a bank has to be able to explain to BaFin and its auditor why the system finds what it should. A model that cannot say why it raised or suppressed an alert is hard to defend in that review.

The EU anti money laundering package moves supervision of the largest cross-border firms to AMLA in Frankfurt, and the single rulebook is described on the EU AML package. AML in Germany covers suspicious activity reports under the German Money Laundering Act.

That is the main reason banks adopt it. A model that scores the whole customer relationship can rank alerts by risk and suppress the ones rules raise for harmless reasons. HSBC's figure of 60 percent fewer alerts with more true findings is the best-known example.

A person. The score directs the analyst to a case and explains the main factors, but the decision to file a report stays with the bank's staff and its money laundering reporting officer.