AI in AML: joining the party before it's too late.

Over the last years, regulators have gained extensive powers to impose fines and sanctions: US regulators have taken the lead, while European supervisors play catch-up. AML fines in Europe and the UK totalled $214M from 2014 to 2017, with those in the US at $1.96Bn. Fast forward to last year and, during the first three quarters of 2018, fines in the UK and Europe reached $918M, compared with just over $1Bn in US penalties. Hence, FIs have been starting to deploy the new generation of smart weapons rolling out machine learning innovation across different stages of the Financial Crime detection process (see figures below).

Where can AI help?

Via the replacement of rule-based models with data-driven ones, the suspicious activity detection process can be made more objective, time-efficient and cheaper. Furthermore, Big Data analysis techniques can help to spot hidden patterns (e.g. particular cases of "smurfing") and processing external unstructured information (e.g. adverse media detection).

Needless to say what the final aim of all this digitalization process is: the reduction of false positives. 99% of the FIs rule-based-generated alerts, in fact, the result being false alarms and, hence, are not escalated into SARs; involving machine learning has proved to be a winning strategy reducing to less than 70% the rates of false positives detection.

WHAT ARE THE ROADBLOCKS?

Why, though, if the upsides are so many, hasn't AI been actively implemented by the majority of the FIs yet?

• Poor data quality: a great part of the data stored in the FIs databases are outdated or missing, the lack of traceability and data lineage.
• Obsolete IT infrastructure: most FIs have still quite out-of-date IT and database systems that need modernization; machine learning routines cannot be implemented across inefficient and obsolete data architectures.
• Limited data sharing: FIs do not formally collaborate on AML ad have no legal freedom to share information about their customers to build a comprehensive network; these constrictions prevent the algorithms from spotting new possible criminal patterns in between different accounts.
• Regulatory restrictions: regulators expect models that can be clearly understood, interpreted and explained; algorithms implemented to assist AML-related issues, though, are usually not that vanilla.
• Hard to explain: there is a clear trade-off between the predictive accuracy of a model and its interpretability: a linear regression would typically be easier to interpret but does not have sufficient predictive power.

HOW CAN ALPHA REPLY HELP?

Thanks to our hands-on and industry-specific expertise, at Alpha Reply we can offer help throughout any of the stages of the process by taking complete ownership of the whole project, or via offering team reinforcement capabilities.