From Netflix recommending the next TV series to binge, to Apple unlocking iPhones through image recognition, artificial intelligence (AI) has become ingrained in day-to-day life. So, if a refrigerator can provide notification when it is out of milk, why are anti-money laundering (AML) platforms unable to eliminate monotonous activities and uncover bad actors with better accuracy? For many financial institutions, AI is doing just that. While there is still a way to go on this modernization journey, there are many current examples of financial institutions doing innovative things with AI to fight financial crimes and be more compliant.
Artificial Intelligence vs. Machine Learning: What is the difference?
Before diving into how banks are leveraging these techniques, two terms that seem to be used interchangeably in this discipline need to be defined. While AI is the broad science of mimicking human abilities, machine learning (ML) is a specific subset of AI that is based on the idea that systems can learn from data, identify patterns and make decisions with minimal human intervention. In short, AI is all about automation of tasks, and ML uses data to learn on its own.
AI and ML are not new. Research in this discipline dates to the 1950s when the Turing test was developed to prove that machines can “think.” Why all the buzz now if these principles have been around for 60+ years? These algorithms thrive in environments with lots of data; therefore, innovations that led to the availability, speed and storage of big data has driven the boom. Now ML can iterate the learning process thousands, if not millions, of times in a very short time frame, fueling innovation.
How are financial institutions leveraging artificial intelligence?
AI Helps Mitigate Trade Finance Risk
Trade finance provides liquidity to the global economy but can be exploited to hide illegal movements of funds. Trade finance has unique complexities due to banks serving as intermediaries in a primarily paper-based operation with an obligation to uncover risks in AML, fraud, military and dual-use goods, boycotts and sanctions. While document imaging has been utilized for some time, deciphering trade documents is still a tedious, manual process that has led to high cost and potential for human error. AI has been leveraged to scan documents, perform classification by document type and extract essential data elements such as locations, names, descriptions of goods and prices. For example, if the document is classified as an invoice, it is then possible to monitor the critical data elements for risk factors such as over/under invoicing, misrepresenting the quantity or quality of goods, and/or phantom shipping to determine if it can be automatically cleared or escalated for further analysis. By using AI, financial institutions can more accurately detect risk and significantly reduce the time it takes to manually process documents. In one instance, a bank reduced the time from several weeks to a few seconds.
AI Automates Case Enrichment for Investigators
It is essential in the AI workflow to make sure information developed by AI is “worked” and validated
Investigators often spend between 60 and 70 percent of their time collecting data from many systems to support case investigations. A key challenge for case enrichment is the accumulation of data that is relevant and necessary for efficient investigating. Traditionally, in-house and third-party data is disparate, available from a complex matrix of big data sources and legacy information platforms causing investigators to fetch it manually or not have access at all. Deployment of AI to automate this process brings opportunities and challenges. When an investigation is initiated, it launches the need to embellish the data about that person, entity, location and transaction. A closed investigation vertical that does not enable third-party data or enterprise-wide data limits and reduces their ability to see connections hidden in the data. In addition, by not having all the data that is needed to investigate business content or entities, a financial institution may fall short in thoroughly investigating a matter, resulting in incomplete investigations or compliance issues. It is essential in the AI workflow to make sure information developed by AI is “worked” and validated. It would be problematic if inaccurate information is looped back into the AI processing and/or used in the decision-making process resulting in false data being accepted. This “fruit of a poisonous tree” event could result in long-reaching inefficiency. High quality, comprehensive and complete data is essential.
A benefit to case investigators using AI processes to enrich additional data is a higher-level identification of information that has been identified in the past but not with enough information to complete processing. AI fills in gaps within an overall picture. While the data it brings may be insufficient on its own, when added to the overall information available, it brings clarity and insights. AI leveraged to identify information can be used to reach a decision quicker and triage or channel the work for follow-up or classification as a closed matter.
How are financial institutions leveraging machine learning?
Machine Learning Reveals What Is Not Known
Financial institutions have discovered that while rule-based transaction monitoring systems can be useful at detecting known risk patterns, they are ineffective at uncovering unknown or emerging behaviors. With vast amounts of data to monitor, how can the entities in a bank that may be hiding below transaction monitoring thresholds be pinpointed? One method leveraged by financial crimes units is to improve upon their existing system by creating an effective segmentation model. Without proper segmentation, thresholds are often set too high, leading to non-productive investigations focused on the largest customers. Unsupervised ML algorithms can be used to group customers into cohesive groupings based on their behavior. Once customers are grouped into a segment, transaction threshold settings can then be applied more appropriately based on how each group transacts. This allows institutions to tailor their monitoring uniquely to each group, resulting in greater coverage of their customers.
A second method that is beginning to gain traction is leveraging ML as a detection method in addition to traditional rule-based transaction monitoring. Supervised ML algorithms can learn from investigators’ historical case decisions related to what was suspicious vs. not suspicious, and use this information to predict new suspicious transactions in the future. This method has had great success in detecting fraud, and institutions have begun to explore the application of supervised learning and semi-supervised learning in the detection of money laundering and other compliance risks.
Machine Learning for Alert Scoring and Hibernation
Many financial institutions have their highly specialized and skilled investigations resources focused on low-value alerts due to the high level of false positives generated by transaction monitoring systems. These resources would be best allocated to spending more time investigating the riskiest activity within the institution. Money laundering and terrorist financing red flags can often be more clearly identified when a pattern of activity is present, rather than an individual alert. To better reveal these patterns, some financial intuitions are leveraging ML to risk-rate alert groupings at the customer level using a scoring model and hibernating or escalating them based on the overall likelihood of risk. This method is not to be confused with suppression or auto-closing alerts. Instead, the process of hibernation has allowed banks to get a comprehensive view of compliance risk and escalate it to investigations once it breaches their programs’ risk tolerance. Leveraging ML to automate level one triage using hibernation frees up resources to spend more time on the riskiest activity and provides them with more information to make their final disposition.
Words of Advice for Those Starting This Journey
Build on Current Investment
Think of AI and ML as a tool in a toolbox, rather than a cure-all. Amazon’s Alexa will not be replacing investigators and robots will not be running entire transaction monitoring systems any time soon—if ever. Time is best spent focusing on supplementing or augmenting the current process.
Start With a Well-Defined Use Case
Tackling AI for the sake of AI will not help compliance organizations better detect compliance risk. Start with a clearly defined business problem that would benefit from AI. Be open-minded and challenge traditional techniques with more modern ones. Keep in mind that when all things are equal, always pick the method that is easiest to explain.
Data Matters More Than Ever
Remember when Twitter users corrupted the Microsoft Chatbot Tay in less than 24 hours by training it with racist and misogynist tweets? Data quality, quantity and completeness are instrumental in adequately training ML models. Do not lose sight of ethical considerations when setting up models to ensure the desired results are being achieved without disparate impact.