How Predictive Churn Models Are Improving Subscriber Retention for a B2B SaaS Company

The Challenge

Renewal risk was often discovered in the last few weeks before contract end, when a champion had already left or usage had quietly dropped months earlier.

The existing approach made it difficult to:

• Spot declining engagement early enough to change the outcome.
• Prioritize CSM time by both churn probability and annual recurring revenue at risk.
• Separate seasonal usage dips from real disengagement.
• Explain to CSMs why an account was at risk and what to do about it.
• Measure whether retention plays were working or simply following accounts that would have renewed anyway.

From a technical perspective, the data told the story in fragments. Product events lived in an analytics warehouse, contracts and seats in billing, tickets in the help desk and relationship notes in the CRM, each with different account identifiers. Churn events were relatively rare and tied to annual renewal dates, so the model had to handle class imbalance and censored accounts that had not yet reached renewal, while avoiding leakage from signals recorded only after a customer gave notice.

The Solution

InnovationM began by resolving account identities across systems, defining churn as non-renewal or a significant downgrade, and setting the existing health score as the baseline to beat. The team built features at weekly grain and evaluated models on time-based back-tests across renewal cohorts. The solution included:

• Unified account data model: product events, seat utilization, billing, support tickets and CRM activity joined in the warehouse with InnovationM's data engineering team.
• Behavioral feature set: usage trends, feature adoption depth, active-user ratios, admin logins, ticket sentiment, payment issues and contact changes such as a champion leaving.
• Churn and survival models: a gradient-boosted classifier for 90-day renewal risk paired with a survival model that estimates time-to-churn for accounts not yet near renewal.
• Explainable risk drivers: the top three drivers for each account shown in plain language, linked to suggested plays such as an admin training session or an executive check-in.
• CRM and CS platform sync: weekly risk scores, revenue-weighted priority and drivers written into account records and task queues through AI integration, so CSMs work from one list.
• Uplift measurement: holdout groups and play tracking to measure which retention actions actually change renewal rates.

The key focus was not simply predicting which accounts would churn, but telling each CSM where to spend the next week and why.

The Impact

Customer success moved from renewal-date firefighting to a weekly, risk-ranked workflow that surfaces disengagement months earlier. The solution supported:

• Higher gross revenue retention in at-risk segments.
• Earlier detection of at-risk accounts before renewal.
• Better precision than the previous health score on held-out renewal cohorts.
• More CSM time spent on high-value, high-risk accounts.
• Clear evidence of which retention plays work, through holdout-based measurement.

By bringing together unified account data, explainable churn and survival models, CRM integration and uplift tracking, the company can extend the same approach to expansion and upsell prediction.