This is the first post of a series on machine learning. In this article we will look at:
Machine learning is one of the most powerful tools of the 21st century. Over the past 10-20 years, many amazing new products have been powered by machine learning, and many traditional products have been significantly enhanced by it.
When we search for anything on Google, we are magically given results that are most relevant to our search. When we step into a car that doesn’t have a driver, it somehow takes us to our destination without needing human instructions. This is the power of machine learning. Much like how Google integrated machine learning to information search, content browsing, email, digital advertising, and so on, Ophelos is injecting machine learning into the debt collection industry with the goal of improving every aspect of the customer experience.
Over a series of blog posts, I will explain in layman terms various machine learning applications, and how Ophelos integrates them into the debt collection process.
I’m sure everyone has heard of the term “machine learning”, or “artificial intelligence”, but what is machine learning, really?
First of all, machine learning is considered to be a subfield of artificial intelligence. However, because machine learning has been so successful, it has become the most well-known aspect of artificial intelligence, and as a result the two terms have become synonymous.
You can think of machine learning as a piece of computer code that identifies mathematical patterns in data based on some predefined constraints.
For example, machine learning algorithms that try to predict whether an image is a cat or a dog will first convert images into lists of of numbers (i.e. to represent the colors of individual pixels), and then solve a mathematical formula to map the numbers representing the image into a binary outcome (e.g., 1 = cat, 0 = dog).
Traditionally, the mathematical formula used to do the mapping was purely constructed by humans. With machine learning however, the human only needs to provide the mathematical framework to use, usually called the “model”, and an algorithm is used to determine the details of the model by trial-and-error on historical images with known cat/dog labels. This process of determining the details of the model is called “training”. It turns out that letting machines determine the details is often much more reliable than relying on humans to do all the work.
Machine learning has proven to be highly effective on a variety of problems across many industries. However, its use is still nascent in the debt collection industry. A major goal we have at Ophelos is to build machine learning solutions that solve a wide range of problems that arise in the debt collection process to improve the overall customer experience.
Our view at Ophelos is that customers in debt should be treated with the exact same degree of care and respect that they experience in the earlier stages of customer journeys. We are working hard to build the same level of technology as the most advanced tech firms in other industries (such as Google, Amazon, and Tesla).
Below is an overview of some applications that Ophelos has already developed machine learning solutions for. Over the next few blog posts, I will describe these problems in detail and explain how machine learning is used to solve them.
Propensity to pay (and forecasting)
Many debt collection strategies depend on the customers' ability to repay the debt. Traditionally, they find the customers with the highest debts and propensity to pay, and then call those customers regularly while paying less attention to customers with lower debts.
At Ophelos, we believe that everyone should be treated with the same amount of care, with the focus being on finding positive outcomes for the customer - not on making the most profit. Thanks to machine learning, we are able to predict customers’ propensity to resolve, allowing us to tailor communication strategies and offer personalized options to customers.
It is important for us to treat vulnerable customers with special care. While our customer support agents are amazing and are trained to detect vulnerabilities, we used machine learning to develop a tool that helps to detect vulnerabilities from text conversations, ensuring that our customers receive the best service possible.
Encouraging customers to engage
We contact customers to let them know about the debt that is owed, tailor the experience to each individual’s situation, and provide flexible repayment options. It is in everyone’s best interest to send communications at a time when customers will be most likely to engage with their debt. To maximize the chances that we reach customers, we rely on machine learning to help us determine when is the best time to contact customers, while minimizing the amount of annoyance or inconvenience we cause.
While machine learning can undoubtedly improve the debt collection process, getting there is challenging. Two major hurdles to successfully implementing machine learning solutions are: 1) understanding how to convert a problem into a machine learning problem; and 2) having access to the relevant data used to train models.
For example, it would be amazing to have a conversational AI tool (i.e. a chatbot) that customers could use to get instant help without having to actually speak to anyone. This would also free up time for customer support staff to provide more help to those who want it. However, customers have unique needs and situations can vary greatly from person to person, so building a fully automated conversational AI tool for debt collection is an incredibly complex task.
Nevertheless, our team are working hard to solve problems such as these every day.