In 2009, the United Nations Office on Drugs and Crime estimated that up to $2 trillion is laundered globally in one year, but less than 1 per cent of this illegal activity is caught. Almost ten years on, this trend still held true in the UK – in 2018, although as much as hundreds of billions of pounds in laundered cash was thought to have washed through the City of London, only 40 arrests were made from 22,196 flagged cases.
The difficulty in identifying suspicious activity is fundamentally a problem of volume: a combination of the enormous amount of transactional information received, the expert money laundering attacks that circumvent published AML transaction scenarios, and the sheer speed of criminal activity. With the advent of near real-time bank transfers coupled by advances in cloud computing, criminals are now able to disperse illicit funds through different coordinated bank accounts so quickly that they are almost impossible to intersect.
The current state of anti-money laundering (AML) efforts
To combat the tide of illegal activity, regulatory bodies are penalising institutions found to be involved in money laundering-related offences, even if these were carried out unknowingly. As a result, global financial institutions have been levied with fines amounting to $26 billion over the last ten years, a development that has seen financial institutions expanding their compliance teams to try and beat the criminals.
Although this has caused the number of alerts to grow through softening of transaction rule thresholds and bringing in armies of human talent, 95 per cent of alerts are false positives, and nearly 98 per cent never result in a suspicious activity report (SAR). Further, of those that lead to SAR filing, there has been a rise in ‘defensive SARs’ that create confusion as the analyst erred on the side of safety vs. truly identifying a suspected money laundering event. This is a low rate of return and often quality of additional SAR filing that is incongruous with the investment being ploughed into AML efforts.
Fight money laundering more effectively with AI
Current AML processes can be streamlined and optimised with AI, the business use cases for which are growing by the day. AI and advanced analytics can reduce the costs of AML efforts while catching more suspicious transactions.
During a proof of concept at a European bank, these AI AML models identified only ten false positives for every 1000 alerts generated for 42 per cent of the highest scoring consumer transactions. Compliance officers, on the other hand, would have had to work through more false positives than real alerts.
With the superiority of AI in identifying genuine AML cases proven, the question now is: why are they so much more effective? This stems from the two new technologies these models are built with: unsupervised learning machine learning and explainable AI.
The most well-known applications of machine learning and AI involve supervised learning, that is, using an algorithm to learn how the input is best transformed to map to the output, which is analogous to that of a teacher supervising learning processes. Unsupervised learning, on the other hand, has no know output to match, which removes the need for a teacher.
Applying unsupervised learning for AML involves exposing the AI system to raw uncategorised data. During this interaction, the computer system identifies patterns that indicate money laundering by using soft clustering on behavioural archetypes to check for customer behavioural transaction anomalies within archetypal clusters (see Fig 1). By being able to boil down transaction behavioural analytics into archetypes of behaviour indicating normal behaviours, and measuring abnormality from that, the AI models can detect suspects very rapidly.
Although a very useful tool, explainability is not AI’s strong suit (due to the typical pursuit of performance over transparency), – hence its reputation as a black box technology. Explainable AI (XAI) is a field of science that attempts to remove this black box and deliver AI performance while also providing an explanation for the “how” and “why” a model derives its decisions. With explainability increasingly demanded by new laws such as the General Data Protection Regulation (GDPR) not to mention AML regulators world-wide, incorporating XAI into AI models has become a question of how, and definitely not when.
Explainable AI opens doors in the banking industry to the advantages of AI, historically these financial institutions have been wary of deploying AI solutions owing to its often inexplicable nature, a quality incompatible with highly-regulated industries. By incorporating XAI in AML applications, financial institutions will be able to explain and justify the model’s decision making, thus satisfying the regulators and making money laundering cases easier to investigate.
AI models can be deployed to supplement existing rules for AML systems, to enrich current data with scores that can be used for both enhanced detection and alert prioritisation, given that the majority of alerts are false positives. With AI, compliance officers will have more time to improve their investigative analyses on the one per cent that matter and importantly those that go undetected by current transaction monitoring. This will lead to faster response rates, the reduction of reputational risk from regulatory fines and a lower cost of compliance. Having AI and humans work in tandem with each other is most effective way to significantly improve current AML outcomes for financial institutions.
AI has huge potential to significantly improve AML processes, particularly due to its ability to reduce the number of false positives and zone in on the truly suspicious cases and find so many more that are currently missed by traditional transaction monitoring. With the development of XAI, it is likely that more and more financial institutions will use AI to bolster their AML efforts, especially as the levels of financial crime continue to rise.
Frank Holzenthal and Dr. Scott Zoldi, FICO