Automating decisions about transactions lets them be handled in real-time, providing better customer service. If that decision fails to detect fraudulent transactions, this lets fraud into the system. Once a fraudulent transaction has been allowed, you are committed to a “pay and chase” approach to eliminating it. This has a poor success rate. Much better to ensure that your initial decision effectively considers the risk of fraud before allowing the transaction. This makes approving transactions where there is a risk of fraud one of the key use cases for digital decisioning. Digital decisioning combines the rule of approval with machine learning and AI for fraud detection.
This combination of explicit rules – based on policies, regulations and human expertise – with advanced analytics and AI is critical. As many articles point out, just relying on AI can get you in trouble. Take this one, for instance – AI algorithms intended to root out welfare fraud often end up punishing the poor instead. There’s nothing inherently wrong about automating these decisions – in this case automation makes it quicker and easier for people to get their benefits and reduces the administrative overhead, freeing up money to be used on benefits not admin. The problem is how they are designed.
It’s essential that any digital decisioning system begin with an accurate and business-centric understanding of the decision being made. Traditional requirements approaches don’t work well for decisioning systems so you need to build a decision model to get this understanding. Apply DecisionsFirst™ design thinking to make sure you capture a realistic and accurate model of the decision.
This decision will likely involve many parts including an assessment of fraud risk. Fraud assessment is not a separate thing, it should be part of the overall decision. You’re going to be more concerned about fraud, and concerned about different kinds of fraud, depending on other elements of the decision. These different elements will lend themselves to different kinds of technology for automation – some will be rules based, some might use data mining, some might need machine learning algorithms. Mix and match the technology you need for this problem.
Finally you don’t want to over shoot or end up in an article like this one on runaway algorithms. Start small and focus on continuous improvement. Don’t try and do this in one “big bang” – it doesn’t work. Focus instead on capturing data about how well the current approach is working and on regular, weekly updates to your decision-making. Test, learn, improve.
Three simple bits of advice can make your digital decisioning project a success and ensure your use of AI is going to work – in fraud detection or beyond.