Goal
Help a leading Latin American water utility increase detection of fraudulent customers, and improve automation and efficiency across the fraud inspection scheduling process.
Use data and analytics to improve prioritisation of customers, detect new fraud types, and help front-line teams execute inspections more efficiently.
Approach
Developed a robust data engineering pipeline infrastructure that enables the business to perform hypothesis testing across data sources and increase automation in how daily inspection lists are generated.
Gained in-depth understanding of end-to-end fraud inspection operations.
Reviewed past fraud inspection data, identified key high fraud propensity features (e.g. irregular consumption, irregular payment behaviour, neighbourhood fraud activity levels).
Developed a machine learning model that demonstrated good results in better inspection prioritisation in simulations with historical data.
Interventions
Designed a pilot for a specific concession to test model performance on daily prioritisation, and surfacing of new fraud cases in the broader population, before adapting the solution to scale to multiple concessions and rolled out to second concession.
Demonstrated an overall 1.8x uplift in fraud and anomalies caught, which translated into an estimated of $12m in revenue collected per year.
This uplift was driven by:
- Increase in detection accuracies (up to 1.6x) and exposing new types of fraud (outside existing rules) based on model recommendations
- Improvements in operational efficiency and productivity (up to 2.3x)
- Value-based prioritisation of customers with higher reconnection rates and recurring revenue
Trained and up-skilled internal team to be able to run end-to-end solution independently and roll-out to other concessions across the country.