Churn Prediction Model — Series A SaaS Startup
A B2B SaaS company providing project management software to mid-size professional services firms across the USA. 450+ active clients, 12,000+ end users..
Challenges
The startup was losing $192,000 in ARR every month to churn — and only discovering it when customers cancelled. By the time the customer success team reached out, the decision was already made. With investors watching retention metrics closely ahead of a Series B raise, solving churn was critical.
Finding Out Too Late
By the time customers submitted cancellation requests, their decision was effectively final. The customer success team had no system to identify dissatisfaction before it reached that point. Every month, dozens of accounts silently moved toward cancellation with no early warning signal reaching the team.
$192,000 in Monthly ARR at Risk
An 8% monthly churn rate on $2.4M ARR meant $192,000 was leaving the business every single month. Even a small reduction in churn would produce significant ARR recovery. The financial impact of solving this problem was clearly defined and measurable.
Series B Pressure on Retention Metrics
Investors evaluating the company for Series B funding were scrutinising retention and churn metrics closely. The startup needed to demonstrate not just that churn was being addressed, but that a systematic, data-driven solution was in place.
What We Built
We trained a machine learning model on 18 months of customer behaviour data — login frequency, feature usage, support tickets, billing interactions and NPS scores. The model assigns every active customer a daily churn probability score. A Slack integration automatically alerts the team when any account crosses a risk threshold, with a tailored intervention playbook for each risk level.
Multi-Signal Behavioural Model
The model was trained on behavioural patterns across all 1,200 accounts, using customers who had already churned as labelled negative examples. By analysing the combination of signals that preceded past cancellations, the model learned to identify the same patterns emerging in active accounts weeks before cancellation intent became visible.
Daily Risk Scoring and Slack Integration
Every morning at 7am, the model re-scores all active accounts and pushes a ranked list of at-risk customers directly to a dedicated Slack channel. Each alert includes the customer name, churn probability score, the primary risk signals driving the score and a suggested intervention playbook tailored to the risk level.
Automated Personalised Outreach
Accounts crossing a high-risk threshold automatically trigger a personalised email from the customer success team — drafted by the system and reviewed by a human before sending — referencing specific product features the customer has not engaged with and offering a dedicated onboarding call.
Implementation
Development Process
Data extraction and preparation took the first week, pulling 18 months of customer event logs, support data and billing records from the client's PostgreSQL database and Stripe account. Model training and evaluation was completed in week 2, testing multiple algorithms including XGBoost, Random Forest and Logistic Regression. The Slack integration and automated outreach workflow was built and tested in week 3.
Technology Stack
GPT-4 · RAG Architecture · LangChain · Python · Pinecone Vector Database · React.js · REST API
Model Validation
The final model achieved 89% accuracy on the holdout test set, with a precision of 87% and recall of 91% — meaning the model correctly identifies 91% of customers who will churn, with very few false positives that would waste the team's intervention time.
Deployment
The final model achieved 89% accuracy on the holdout test set, with a precision of 87% and recall of 91% — meaning the model correctly identifies 91% of customers who will churn, with very few false positives that would waste the team's intervention time.
Conclusion
The churn prediction model delivered exactly what the startup needed — a systematic, data-driven approach to customer retention that produces measurable ARR recovery and gives investors clear evidence of operational maturity. By identifying at-risk customers 30 days before cancellation and automating the intervention workflow, Xor Solutions transformed churn management from a reactive scramble into a proactive, scalable system. The model continues to improve as it learns from new customer behaviour data every month.