Of the many uses of artificial intelligence (AI) in financial services, finding and documenting fraud and money laundering crimes stand out. In general, AI’s value proposition lies in automating tasks that are extremely cumbersome for people. Some processes are slow and expensive for people and can be made faster by AI. The real game changers are processes that are so slow and expensive for people that they are not done at all—people could do them but do not. Through automation, AI can deliver a net-new use case here.
Financial crimes are often an example of this. Some prominent examples of financial crimes have come to light through accidental discovery or through an outright mistake of the criminal. The bulk of the investigation does not lie in the discovery of the crime but rather in the documentation in preparation for legal proceedings.
Fraud Detection
There are many types of fraud, and each leaves its own fingerprint. Generally, money is either transferred to someone other than the intended recipient, the money is transferred without the approval of its owner, or the service for which the money is spent is not provided. We are familiar with the old credit card check that if a credit card is used in two countries at the same time, one of the two transactions must be fraudulent. Patterns such as this are constructed by experts in the subject and implemented in computer systems for detection. The complexity arises due to the vast amount of transactional data—most of which is legitimate—and the short approval times (fractions of seconds) available to greenlight a transaction.
AI helps primarily by learning these complex fingerprints of fraud from a library of examples and then executing the detection rapidly in line with the transaction processing system. As the library of examples evolves continuously and quickly, the AI models must be retrained often, and this can be done automatically. The result is a largely automated pipeline of tools that achieves a high degree of accuracy in detecting fraud. The key to make this happen is to keep the library of examples up-to-date and this is necessarily a human activity requiring some expertise and contact with the evolving nature of the industry and its bad actors.
Anomaly Detection
Determining that a financial crime has occurred is, therefore, a challenge and can be made significantly easier with AI. This falls under a more general heading of anomaly detection in the science of AI. When looking at a large collection of objects, such as financial transactions, the goal is to find those that are different from the norm. As “the norm” in international finance and trade is fairly complex, detecting deviations from it is not easy and is practically impossible to encode in hard-and-fast rules. That is where AI comes into view.
We show a good deal of data to AI that we know to be legitimate and let AI learn directly from the data to determine what is “the norm.” When new data does not fit into this complex schema, we can quantify how different it is and raise an alarm. If that happens multiple times, AI can attempt to find regularities among those anomalies by AI methods known as pattern recognition. Those patterns can then be described in human language by explainable AI systems that can alert a human analyst to the observation that multiple anomalies appear to be connected.
Data Integration
While AI is at the heart of making this work, the bulk of the work behind the scenes lies in uniting lots of diverse data sources. Apart from the obvious difficulties of managing access to multiple systems owned by an ecosystem of competing institutions that reside in many countries and are subject to different laws, these data sources must be cleaned. For instance, the same legal entity may be encoded differently in different systems. The fact that these are duplicates must be discovered and resolved. The umbrella term of master data management covers the resolution of diverse records referring to the same real-world entity, such as a company or a person.
Other facts, for example that one legal entity is a subsidiary of another, can be hard to determine and require yet more data sources to be attached. While AI delivers the final interesting outcome, it is like the icing on a cake. The substance of the cake, however, lies in data warehousing and software interfaces to grant AI essential access to all information. Beyond these aspects, the right processes must be in place to allow seamless processing to take place. Data governance is the central aspect here and combines three main elements. First, it controls who has access to what. Second, it documents all the diverse datasets so that everyone knows what they are working on. Third, it assures the quality of the data itself with a variety of checks and balances.
Risk Quantification
Practically, transactions are scored using models that compute a risk, which is the probability that this transaction is not legitimate. A probability is, of course, a number between zero and one. At some point along this spectrum, we have to draw a red line in the sand that represents our risk tolerance. Every transaction scoring a higher risk than that will then be alarmed for human attention.
As humans provide more feedback to the AI system on which high-risk transactions are acceptable, the AI system improves and the risk score becomes better and better at identifying truly illegitimate items. This approach is an example of a human-in-the-loop system where the AI automates much of the analysis but acts as a decision support tool in which a human provides the final step and decides whether to do something or not.
Risk Categories
Financial crimes fall into several categories and can arise in different mechanical ways. Distinguishing them is useful for AI as well as humans because each method has its own fingerprint in the data and consequences in real life. Fraud, money laundering, insider trading, money mules or the use of shell companies leave behind a very different trail of evidence that must be found out using different source data.
Beyond just the categorization, certain financial crimes are often local, such that all the data comes from a single jurisdiction, for example, in the case of fraud or most cases of insider trading. However, money laundering is often an international crime that involves much more complex data analysis because source data from other countries is required. This also brings various privacy and data protection laws into play that make the analysis difficult—after all, we must search through a huge amount of legitimate data belonging to innocent people to find the proverbial bad apple.
Ways Forward
Detecting and flagging financial crime is certainly an ideal use case for AI automation, as only AI has the capacity to look through the amount of data necessary and to capture the complex patterns that indicate normality or abnormality of behaviors. The speed at which it can provide risk scores and flag important events to expert human investigators is instrumental and unique. The more such systems are used, the better they will get through the feedback loop of the expert humans who use them.
Any such analysis however rests on a foundation of data obtained from many sources that is well curated, cleaned, documented and stored in a legally compliant manner involving data governance and master data management, conducting AI in an ethically responsible manner. Once all of this is done, we will be able to systematically detect financial crime instead of relying on a lucky chance to stumble across a strange transaction.
Patrick Bangert, SVP Data, Analytics and AI, Searce Inc., patrick.bangert@searce.com, 
