Context

"Before, we noticed churn too late. Now, we act preemptively." Using neural networks on Vertex AI, one of the largest loyalty programs in Ukraine learned to analyze the individual behavior of 1.5 million customers and automatically detect churn risks before the person actually leaves.

challenge

Before the project started, the team defined churn using a simple rule: if a customer had not visited the restaurants for 3 months, they were considered "lost." This approach gave a signal too late – when the person had already effectively left, turning any reaction into resuscitation rather than retention. At the process level, the difficulty lay in the fragmentation of data: customer behavior was scattered across various sources, and gathering a coherent picture from them manually was impossible at the scale of over a million profiles. An automated response did not exist: even if the analytics detected an alarming signal, it did not reach the CRM system in time.

Why now?

LOKAL program grew to a scale where manual or rule-based churn management was no longer sufficient. With 1.5 million participants, even a few percent of uncontrolled churn means tens of thousands of lost customers. An instrument was needed to analyze each customer individually and send a signal to the CRM before the person made the decision to leave.

solution

We built an end-to-end churn forecasting pipeline based on Google Cloud. Data from various sources is gathered and pre-processed through a Data Fusion layer in BigQuery. Based on this, a recurrent neural network (RNN) was trained, which models the sequence of each customer's behavior over time – visits, activity, and transactions. After training, the model was deployed via Vertex AI Model Registry and Vertex AI Endpoint, ensuring stable industrial serving without manual intervention. Additionally, we developed a custom Python module that automatically synchronizes the model's results with the LOKAL CRM system: every customer receives an up-to-date "propensity to churn" score, allowing the team to react proactively. Separately, based on the internal user embedding constructed by the same model, we laid the foundation for a personalized recommendation system.

result

• LOKAL team transitioned from reactive logic to predictive: instead of the "3 months without a visit" rule, there is an individual assessment of each of the more than 1.5 million participants.

• The churn risk signal now reaches the CRM automatically, before the customer "cools off," enabling the launch of retention campaigns at the exact right moment.

• The churn rate decreased significantly – decisions are made based on data, not on intuition or delayed signals.