For heyData, a Berlin provider of GDPR and data protection software, Gemma Analytics built an AI-powered churn score. A machine learning model predicts the churn probability of every customer and is twice as accurate at identifying the top churners as the rule-based system it replaced.
The starting point
heyData offers small and medium-sized businesses digital solutions for data protection, compliance and risk management, sold as subscriptions. In a subscription business, the account management team needs to know early which customers are at risk of leaving, while there is still time to act.
Until then, heyData relied on a rule-based risk score. It measured every customer against the same fixed thresholds, which had been set once and never adapted as the business grew. The score did not learn from the customers who actually cancelled, and it was not accurate enough to show who would really churn.
What we built
Gemma replaced the rule-based score with a machine learning model. Instead of fixed rules, the model learns from history, from the customers who stayed and the ones who left, and calculates for each customer how likely a cancellation is within the next 12 months.
The score does not sit in a separate tool. It is integrated into heyData's account management dashboard, visible in real time next to every customer. Account managers can sort their accounts by risk and turn to the ones that need attention most.
The result
The AI churn score is twice as accurate as the old rule-based system at identifying the customers most likely to churn. Every customer now has a 12-month churn probability, live in the dashboard. Account managers spend their time on the riskiest accounts first and can plan retention measures before a customer cancels, not after.
Tools and setup
The churn model is written in Python and works with heyData's customer data in the Snowflake data warehouse that Gemma runs as heyData's data platform. The scores flow into Metabase, where the account management dashboard shows them for every customer. Model, warehouse and reporting form one setup, so the score runs automatically without manual steps.