Several extensive investigations around the globe into the activities of banks have revealed massive failings in the fight against money laundering. The research showed that more and more criminals are exploiting the financial system and laundering their black money by means of quick and immediate remittances.
Banks face the challenge of keeping their ever-increasing compliance costs under control. To keep the balance between security on the one hand and cost efficiency on the other, financial service providers need to improve their compliance systems, but how?
Fortunately, supervisory authorities are adapting to advances in analytics, which will help financial institutions. There is a growing willingness to trust new technologies such as artificial intelligence (AI), machine learning or robotics.
Authorities actively encourage banks to consider, evaluate and, if necessary, implement these innovative solutions.
This does not mean that the existing risk-based approach, which is based on good compliance knowledge in defining ‘detection scenarios’, is no longer valid. In fact, many regulators insist that Anti-Money Laundering (AML) controls need to remain explainable and auditable, they see its application in supplementing existing controls. Hence, a kind of coexistence is emerging – a mixture of the existing scenarios and the AI mechanisms.
Thank you very much, Mr. Roboto
Robotic Process Automation (RPA) enables banks to massively streamline and automate investigations and alert processing within their Know Your Customer (KYC) and AML solutions.
Until now, entire crowds of investigators have often been hired only to do the initial processing of alerts, which typically consists of at least 90 to 95 percent of false positives.
This is especially true in Watchlist Screening, where a lack of data quality and the nature of the matching process itself create massive amounts of alerts, which to the human eye are obviously false.
In doing so, clearly defined alert and case rules – which are specifically adapted to the situation of the financial institution, products, customers, etc. – can cover many of the repetitive activities, thereby minimizing duplication of efforts, while still retaining full audit trails. RPA should therefore be integrated into company-wide alert and case management.
This also ensures that investigators time is freed up from mundane tasks, so they can focus on more complex work. RPA and analytics-based alert prioritisation even helps to reduce costs in the short term.
Above all, however, the combination of both ensures enormous efficiency and efficiency improvements.
Experience from the everyday life our customers has shown that the number of suspected cases increases by up to 20 percent through appropriate implementations, while at the same time increasing efficiency by up to 30 percent in alert and case management.
Deus ex Machina
However, automation and analytics are not the only means to effectively combat money laundering.
As early as 2017, a study by McKinsey showed that machine learning could reduce the number of ‘false positives’ by 20-30 percent. At the same time, machine learning in the AML area can improve the conversion rate ‘from alarm to suspicion’ by a factor of three thanks to tighter segmentation.
Examples of finer segmentation include isolating the fact that a customer has financial relationships outside the country, is a wealthy private individual or is a small business owner.
While a risk-based segmentation into such customer groups can be done via configuration of expert systems, given the complexity of all our individual lives, it is also a mammoth task when done manually, and very hard to maintain over time.
In this way, machine learning calls into question the status quo of KYC processes, as it can take over such an ongoing segmentation by using real-time behavioural analysis based on financial transaction activities.
It furthermore can add value by sifting through banks mountains of data, exploring, and finding unknown connections. Again, a boon for KYC operations by providing more information, while reducing effort.
One of the most basic requirements in AML is to ensure that we are keeping audit trails of our own actions, as well as the underlying machine behavior.
This is very easily done with rules-based systems, since we can point to the underlying typologies we have used when defining them and explain the thresholds we have chosen. Hence, machine learning models need to be similarly explainable in the event of an investigation, or an external audit.
For example, the following illustration shows how the different variables of the model (V1-V6 on the far left) are incorporated into the machine learning algorithms, the results of which are processed by a Reason Reporting and Ranking algorithm.
The reasons are sorted by importance and relevance to explain how the model came to the evaluation. Despite the model not using fixed thresholds like a rule would, an analyst can therefore easily see why an alert has been created.
Another way for financial institutions to reduce costs and improve business outcomes is to connect AML and fraud detection solutions, as what we have done for our clients.
These are relatively similar and meet many common requirements, such as detecting unusual behavior. Unfortunately, most banks still operate them in a silo environment.
The use of a fully scalable IT environment that simultaneously meets both fraud detection and AML requirements would not only provide a cost advantage – it would also allow financial institutions to act ‘cross-border’ in detecting illegal activities.