Case Study: natural language processing & customer churn use case for Kooperativa (part of Vienna Insurance Group)
Advanced natural language processing (NLP) helps to understand customer interactions and automatically route each individual case to the right specialists. It can also detect and monitor important topics customers care about, understand what drives customer satisfaction, and predict churn risk.
The insurance company Kooperativa transcribes calls from within the call centre to determine the most common customers’ complaints. It aims to distinguish which topics are prevailing and, therefore, identify the problems their clients face. The company was also interested in the churn analysis to reduce the loss rate of their customers.
Based on machine learning and natural language processing, our solution evaluated the net promoter score (NPS) forms and churn analysis. We analysed the duration of the call, number of calls from the same customer, intonation, pitch, and speed of the client’s speech. We proved the hypothesis that it is possible to predict that the customer is dissatisfied, and if it could lead to churn.
During the three months of the project duration, we used NLP practices and worked with audio and text evaluations. In addition, we used machine learning to categorise NPS forms, AdPicker for improving AdForm advertisement, and helped optimise the cost of online acquisition.
- Creation of personal profile based on a combination of Google Analytics and AdForm increased web traffic per person and reduced costs of one visit by 40%.
- Machine learning helped to categorise 80% of the NPS forms.
- Analysis of customers’ calls predicts churn.
- To implement the identification of customer dissatisfaction quickly.
- To connect data as much as possible for a better customer experience.
- To target advertising more precisely on relevant customers.
Categorization of NPS forms by topic based on free text fields. The solution significantly facilitated manual sorting, as up to 80 % of forms can be categorized automatically. Other outputs included keyword detection and text sentiment determination.