Data Science & IA

Development of Intelligent Collection Management Tools

The challenge

A water distribution company with more than 500,000 customers needed to improve its collections. For this, it had five possible channels to contact its customers: calls, visits, SMS, letters and mails. The company wanted to know which customers to contact through each of the channels in order to maximize collection.

What started as a couple of proofs of concept ended up with a productive model from data ingestion to writing to the ERP.

The strategy

To achieve this goal, two complementary models were developed:

  1. Payment amount prediction model: to recommend through which channel to contact each customer, the expected payment amount considering each alternative was estimated by regression. Decision tree models, with a Catboost-type machine learning algorithm, gave the best results, with short processing times.
  2. Engagement timing model: in addition to suggesting the channel, a customer segmentation model was used to determine the day of the week with the most effective contact for each customer, as well as the optimal week of contact.

At the end of their development, these models were integrated into the AWS cloud, automating everything from data processing to writing the results in the ERP.

Monthly balance

detail of the monthly debt status of each client, including information on installments, agreements, etc.

Historical contacts

each of the contacts made to customers through the five channels described above, according to date and result of management

Customer information

category and type of customers, size, demographic classifications, among others.

The achievements

Prior to a nationwide implementation of the model, several pilots were carried out, where it was concluded that collection improved collection effectiveness thanks to the model’s recommended contacts. Once the model was validated, it was implemented nationwide, recommending more than 58k contacts throughout Chile, all this automatically, without the need of human input for decision making.

7% improvement in effectiveness

More than 58,000 recommended contacts made

End-to-end automated model execution