Embracing AI & Machine Learning for AML
There’s a conversation in the industry today around leveraging machine learning for AML as AML and compliance leaders are confronted with unrelenting foes and obstacles in the form of financial crime. FIs are facing increased regulatory scrutiny and the quantity, variety and velocity of transactions continue to rise. The legacy systems which have suspicious activity prevention mechanisms embedded in them are not adequately equipped to keep up and prevent damage from money launderers. Moreover, criminals have become more sophisticated as they have access to advanced digital channels that allow them to penetrate the financial system at the slightest opportunity.
Despite additional rules and systems in place, FIs are slapped with fines which shows there is undetected activity at play that puts tremendous pressure on them to be more vigilant. According to a recent PwC report, 95 percent of alerts generated from transaction monitoring systems (TMS) are false positives. Even more importantly, 50 percent of suspicious activity can go undetected in TMS resulting in fines and expensive lookbacks from regulators.
AI & Machine Learning Applied to AML
Let’s examine the types of AI available today. One type is called machine learning. Its uniqueness lies in making computers not only as good as humans but exceeding human abilities in analyzing data and making decisions. Different machine learning technologies are a business enabler for existing workflows within the AML process.
The first approach is called supervised machine learning which is the data mining task of inferring a function from labeled training data. Training could be simulated or real data. In AML use cases, it is historical data and historical labels. The training data herein consists of a set of training examples. This approach requires historic data sets and labels that can be used to train an algorithm to calculate customer’s risk ratings as part of the onboarding process. Entity matching provides level of confidence and score to indicate how closely a name matches against a sanctions list. Where SML does well is in prioritizing alerts. Once SML models are trained with historic alerts, they can prioritize the results provided by TMS.
The second approach to using machine learning is called unsupervised machine learning (UML) which does not require large data sets or labels. The algorithm looks for clusters and patterns without any preconceived bias. It can be applied at multiple stages of the AML and monitoring process. As part of on-boarding, UML can identify patterns and segments of customers, examine account openings or find accounts that have been taken over (ATO). On the client screening side, UML is very good at identifying hidden and not so obvious links. Instead of just matching against a fuzzy logic or historic labels, it can surface what’s important in the data itself. On the transaction monitoring side, UML can identify previously undetected suspicious activity and thus reduce false negatives.
Natural Language Processing is another type of AI that allows for efficiency gains by automating language related tasks. This is particularly applicable in areas such as negative news searches, checking names against sanction lists which are usually time consuming but can be automated with NLP. Natural Language Generation can also assist investigators to file SARs reports.
From an AML standpoint, machine learning is widely used because of its transparency especially compared to black-box approaches such as neural networks, a must have requirement in a highly regulated industry.
Benefits of Adopting AI
It is no surprise that AI is being applied in many aspects of our life already. The science of Computer Vision allows Facebook to recognize human faces and pictures. The ability of Amazon’s Alexa or Apple’s Siri products to understand and respond to human speech is a result of National Language Processing (NLP) which is AI at work. Self driving cars, robots going through a fulfillment process at an Amazon warehouse are all examples of AI serving us.
While many institutions are looking for new solutions to help them reduce the burden of financial crime, the inner workings of certain types of AI are not clearly understood. For example, deep learning or neural networks are black boxes because they lack transparency, yet other types of AI such as unsupervised machine learning (UML) does offer transparency. It is also important not to look at AI in isolation, but weigh the pros and cons of various types of AI approaches, and adopt one that is complementary to the existing rules based TMS systems to allow for better coverage across known and unknown suspicious activity. The most effective approaches can be complementary, helping financial institutions retool with better technologies and meet compliance with regulatory requirements.
AI can help AML practitioners with:
Reducing False Negatives: AI can help identify suspicious activities that are missed by existing TMS. This is offered by unsupervised machine learning which does not need large historical data sets and can find the correlations between the accounts that are not obvious to human analysts. The algorithms look for new patterns that are shared across different users, accounts and transactions, and can also adapt and change to newer patterns as they emerge.
Triage of Alerts: AML investigators are faced with handling the case volume that comes through TMS and lack of an efficient way to prioritize them. AI can help by prioritizing the workflow queues for TMS alerts so the alerts can be acted on more efficiently.
Multiply Your Defense
While a change can be perceived as disruptive, leveraging multiple solutions that incorporate advances in AI and ML are important steps in the journey of FIs to reduce costs and bring greater efficiency and effectiveness. DataVisor’s technology applied for detecting suspicious activity and compliance benefits not only shareholders and customers but also the global financial system.