Hemisphere did this project for a large utility company ( with 2 million+ clients). These clients have to pay monthly and annual invoices. However in about 2% of the cases people were not able to pay the invoice in time. In some cases this resulted in far going payment issues. By detecting potential payment risks using the companies service chats we were able to detect payment problems early on.
Not detecting payment problems in an early stage leads to a cumulation of problems that even further the risk of future payment. Mitigating non-payment early on helps to prevent these problems.
By applying NLP models on hundred of thousands of anonymised WhatsApp messages we were able to detect patterns (early warnings) that would lead to late or non-payment. These patterns were distilled from clients that ended up in certain situations: i.e. divorce, job loss, moving abroad etc. In a 10 week pilot we tested together with the client if a proper mitigation of this early warning group showed a lower non-payment percentage over the control group. After the 10 week period we discover that we could reduce non-payment with as much as 25%.
Better insights: With the early non-payment warning detection the client was able to detect signs of non payment early on.
Better decisions: By applying the right means at the right moment, the client was able to prevent a build-up of payment problems later on, this resulted in more invoice payments and happier customers.
400.000 anonymized WhatsApp messages
Semi-supervised clustering model
Explore, Experiment, Prototype
Realized in 16 weeks