Despite many examples of successful Artificial Intelligence (AI) applications in the finance industry, the back office is still forgotten. Credit and collection professionals still spend a lot of time and effort in manual and mundane work. We designed YayPay to have predictive capabilities with the goal of leveraging AI to bring process improvements and insight to free the back office.
AI is a tool that helps people break free of mundane work and switch to useful, creative, and value-generating responsibilities in their day-to-day jobs. From fraud detection, credit decision making, risk management, trading, conversational services, and insurance underwriting, AI start-ups are applying machine learning techniques to well-known inefficient processes in finance and improving them. We are creating a product that makes collections intelligent, but it also manufactures intelligence for the back office.
So how did we go about doing this for our product? We initially developed our prediction algorithm for precision. Our first version attempted to predict an exact date of the payment of an open invoice based on the historical data and some other behavioral characteristics. We achieved an accuracy of around 80%, which meant that our predicted invoice full payment date for an open invoice would be plus-or-minus three days from the actual date in 80% of the cases. Well, the problem with this approach was that if you needed precision, you are making sizable errors, and the customers won’t really like that.
Fortunately, the most important goal for foresight was accuracy instead of precision. We discovered a different but more practical use case. Sometimes what our customers need is not the exact date an invoice is going to be paid, but rather in which phase of the invoice life cycle the invoice will be paid. Will the invoice be paid before the due date, go overdue, or go 60 plus? Knowing this, we changed our algorithm. Today, our product estimates whether the invoice is going to be paid by the due date with an accuracy that over 90%. Then, once the invoice goes past due, we estimate whether it will be paid in 30, 60, 90, or more than 90 days with varying levels of accuracy. In this instance, the predictions for the first 30 days is the most accurate at around 90%, and at 80% accuracy for the 90+ bucket.
We expect these models to improve significantly in performance as we integrate more internal and external sources of data into our models. Right now, we start with invoice-level data and payer/payee-level behavior data. Then we include contextual information such as industry, revenue, and company size. When leveraging this data, we are careful to choose the appropriate algorithms that can correctly leverages contextual information by learning their interactions and dependencies.
In the near future, any area that requires decision making and has large amounts of structured data can become handled by AI. We envisioned a product that, with sufficient data, can look at trends across multiple levels. For example, how are sales distributed across individual buyers and sectors? How are the sectors performing relative to each other and over time? We also want our product to provide insights to the underlying trends. For example, what sectors are growing or shrinking? What do the Account Receivables (AR) assets really look like from the expected cash flow perspective? When will you get paid? Can we expect changes in payment behavior? Can we use payment behaviors as an early indicator of economic shifts in key sectors?
