Horizons Spring/Summer 2019

Utilizing Artificial Intelligence to Detect Payment Fraud Applications of artificial intelligence (AI) are proliferating across a spectrum of consumer products, augmenting our everyday lives in ways few thought possible. Simply defined, AI describes the development of computer systems capable of performing tasks that require human-like cognition. AI can be divided into either a rules-based system or a machine learning system. The former relies exclusively on rules defined by human experts whereas the latter relies heavily on models that are ‘trained’ on large collections of examples. In the case of payment fraud, a business could employ a rules-based AI application by flagging payments to P.O. boxes or addresses outside the countries in which the organization does business. The AI application would ‘monitor’ payment activity and then automatically route the information to the appropriate parties for review. A business could also utilize a machine learning system to mitigate potential payment fraud by building a system that relies on a clustering ‘model’ or algorithm to automatically cluster or group payment activity based on certain vendor traits (such as location, transaction volume, type of payment, etc.). This could automatically detect transactions that are outside the norm. It is not too much of a stretch to build trust in the rules-based system, especially if logic is built internally (without relying on cloud-based systems). However, it may be more difficult to truly trust a machine learning application as the logic is more complex. Further, the leadership of the enterprise may be comfortable that the complex system ‘works’ based on past examples, but this creates a ‘black box’ that may not be evident until the system fails. A system that ‘works’ in terms of detecting fraud does not necessarily address privacy concerns either – however, all is not lost. Current research in machine learning is focusing on reducing the reliance on large collections of examples and developing privacy-protecting analytic methods. Imagine if you could confidently encrypt (for example) your data before feeding it to a powerful cloud-based system that was purpose-built as an engine to mitigate fraud. The enterprise would get the horsepower without having to maintain the complexity of such a system. The even better news is that the encryption step is becoming much more approachable, allowing you to easily create a ‘key’ to lock and unlock your data stores, for example. No programming skills required. The promise of AI lies in the ability to deploy automation at scale regardless of whether it is a more simplistic rules-based system or a more advanced machine learning model. In our rules-based example of mitigating payments to questionable addresses, it is easy to conceive of an automated solution that continuously monitors the payment ledger for such disbursements and then flags and routes them to the appropriate people.

Spring/Summer 2019

11