How Our Churn Prediction Model Drove a 90% Reduction in Customer Attrition

Key Takeaways: Tesseract Academy helped a major device insurance provider predict and prevent customer churn using machine learning. By identifying at-risk customers before they left, the company could proactively reach out with retention offers, significantly improving customer retention and protecting revenue. The AI-driven churn prediction model correctly flagged nearly 90% of customers who would have churned, with minimal false alarms. This case study illustrates how an AI solution can turn around a churn problem and deliver substantial ROI through higher customer loyalty and lifetime value.

The Problem: Customers Were Leaving – and It Hurt

Ever wondered how to spot at-risk customers before they leave? That question was at the heart of a challenge faced by one of the world’s largest electronic device insurance companies. This insurer noticed a worrisome trend: many customers were churning – i.e. canceling or not renewing their device insurance policies. Customer churn is a silent revenue killer for any subscription business. It’s far more expensive to acquire a new customer than to retain an existing one. High churn meant lost revenue, higher marketing costs, and a hit to the company’s growth. The insurer needed to understand why customers were leaving and how to intervene before it was too late.

Two reasons were suspected: 1)Customers buy device insurance (for example, on a brand-new smartphone) to protect their investment, but they might leave if they feel the product isn’t useful – say, if a claim isn’t covered or 2)if they never end up needing to use the insurance. 

In the insurance business, churn comes in two forms: 

active churn (when a customer cancels their policy before its expiration) and 

passive churn (when they simply don’t renew at the end of the term). 

Both were happening, and each lost customer hurt the bottom line.

Data Challenges: To tackle churn, the insurer had a trove of historical data – but with some limitations. Due to privacy regulations, personal demographic data (age, gender, etc.) wasn’t available. Instead, the team had to rely on behavioral and product data from each policy. Key data points included:

• Device details: the type of device insured and its technical specs (e.g. a high-end smartphone vs. a basic model).
  
• Tenure: when the customer joined (how long they’ve had the insurance).
  
• Geography: the customer’s country or region.
  
• Usage patterns: other proprietary data on how the policy was used (e.g. claims history, service interactions).

This data held clues about churn. The challenge was to sift through these signals to find early warning signs of churn. The goal was clear: predict which customers were likely to churn, ideally when they might churn, so the business could act in time to keep them.

Using AI to Predict Churn Before It Happens: Building an Effective Churn Prediction Model

Tesseract Academy’s data science team partnered with the insurer to develop a solution that would transform raw data into actionable churn predictions. The approach was two-pronged: first, analyze the data to uncover factors driving churn, and second, build a predictive model to flag high-risk customers in advance.

The team began by exploring the historical data to identify patterns. They investigated questions like: Do certain device types have higher churn rates? Does churn risk spike at a particular policy age? Are customers in some countries more likely to cancel? Through statistical analysis and visualization, the data scientists identified which factors correlated most with customers leaving. For example, they found that churn risk tends to increase as a policy ages – in other words, the longer a customer has the insurance, the more likely they are to eventually drop it. This makes intuitive sense: when the device is new and expensive, customers are motivated to insure it; as the device gets older (and loses value), some customers reconsider the need for insurance. Insights like these gave the insurer a clearer picture of why churn was happening. In fact, Tesseract Academy delivered a ranked list of the top features and signals that put a customer at higher risk of churning, providing valuable business intelligence to guide retention strategy.

With these insights in hand, the next step was building the machine learning pipeline to predict churn. There are multiple ways to model churn, and Tesseract’s experts experimented with two popular approaches:

• Classification Models: A traditional classification model outputs the probability that a given customer will churn in a certain time frame. For example, it might say “Customer A has a 80% chance of canceling in the next month.” This approach is straightforward and provides an easy-to-understand risk score for each customer (the higher the score, the higher the churn risk).
  
• Survival Analysis Models: Going a step further, the team also leveraged survival modeling – a technique borrowed from the medical field that estimates how risk changes over time. Survival models are powerful for churn because they don’t just ask “Will this customer churn?” but also “When is this customer likely to churn?” Using survival analysis, the model could compare the relative churn risk between customers and track how that risk evolves month by month. For instance, the team plotted survival curves for different contract types to see how retention dropped over time for each. A common pattern emerged: the “survival probability” (likelihood a customer remains) decreased steadily as time went on – meaning the probability of quitting grew higher with each passing month. This temporal insight was key to prioritizing interventions.

In the end, the solution combined the best of both approaches. Tesseract Academy developed a robust machine learning pipeline that ingested the insurer’s data and produced two outputs for each customer: a churn risk score (probability of churn) and a time-to-churn estimate. The model was trained on past customer behavior, with churn (canceled or not renewed) as the target variable. It was tuned and validated to ensure it would generalize well to new customers. To evaluate the model’s performance, the team used standard metrics and visualization tools. We tested predictions on a hold-out sample of historical data, comparing predicted churn vs. actual outcomes. A confusion matrix summarized the results – showing how many true churners were correctly identified and how many false alarms (non-churners incorrectly flagged) the model produced.

The figures shown here are for illustrative purposes only – they use mock data to demonstrate the format and style, not actual client metrics.

The confusion matrix told a positive story: the model was catching the vast majority of true churners while keeping false positives low. The Tesseract team iterated on the model until it met high accuracy standards, then deployed this solution for the client as a proof-of-concept tool.

The Results: Proactive Retention and Measurable ROI

The AI-driven churn prediction model had a transformative impact. With the model in place, the insurer could proactively target customers at risk of leaving – turning churn from a surprise into a manageable risk. Tesseract Academy’s work delivered three key outcomes for the client:

1. Identified Why Customers Churn: The analysis pinpointed the top factors contributing to churn, giving the business clear insights into customer behavior. For example, policy age, device type, and usage patterns were found to significantly affect churn risk. This meant the company could address underlying issues (like coverage gaps or customer engagement) that were driving people away.
   
2. Predicted Who Is Likely to Leave (with High Accuracy): The machine learning model proved remarkably effective at predicting churn before it happened. In testing, it correctly identified about 89% of customers who would churn, with up to 90% precision. In plain terms, if the model flagged 10 customers as high-risk, about 9 of them would indeed end up leaving, and it would catch roughly Nine out of ten actual churners in the data. This level of accuracy gave the company confidence in the predictions. They could now compile a list of at-risk customers each month, knowing the model would reliably capture the vast majority of true churners while avoiding too many false alarms.
   
3. Estimated When They Might Churn: Thanks to the survival analysis component, the model didn’t just produce a static risk score – it also indicated how soon each customer might churn. The output could rank customers by urgency of risk. For instance, Customer X might be likely to churn within the next 30 days, whereas Customer Y (with the same risk score) might be more likely to churn in 90 days. This temporal ranking let the insurer prioritize outreach, focusing their retention efforts on those customers who needed attention immediately. It’s one thing to know who is at risk; knowing when they might churn is a game-changer for scheduling timely interventions.

Equipped with these deliverables, the insurance provider was able to launch targeted retention campaigns. Now, instead of a blanket one-size-fits-all approach, their customer success team could focus on the 15–20% of customers flagged as high risk and offer them special incentives, personalized messaging, or improved service. Even if only a fraction of those outreach efforts succeeded in saving a customer, it directly improved the company’s bottom line. In subscription businesses, keeping just a handful of customers from churning can translate to significant revenue retention over the long term.

The insurer’s team envisioned a dashboard where each customer is listed with a churn prediction model risk score and predicted churn date. Such a tool enables managers to quickly see “Who do we need to call this week to prevent cancellations?” For instance, a product manager could filter for all customers with >80% churn probability in the next 60 days and initiate a win-back campaign for that segment. This kind of data-driven prioritization ensures that limited retention budget and resources are used where they matter most, maximizing ROI.

The figures shown here are for illustrative purposes only – they use mock data to demonstrate the format and style, not actual client metrics.

Conclusion: Turning Churn Prediction into Business Value

This case study showcases how an AI-driven solution to customer churn delivered clear business value in an approachable, actionable way. What started as a high churn problem for a device insurer ended with a customized churn prediction engine that boosts retention and revenue. The project highlights a few takeaways for any tech decision-maker:

• Strategic Value of Data Science: Tackling churn with AI isn’t just a tech experiment – it’s a strategic move that directly impacts ROI. By retaining more customers, the insurer protected its revenue streams and reduced acquisition costs, yielding a tangible financial payoff.
  
• Ask the Right Questions: A key to success was framing the problem in business terms (e.g. “How can we know who might leave and when?”). This kept the project focused on outcomes, not just algorithms. Throughout the process, the team posed questions like “What factors are driving customers away?” and “What would it take to change their mind?” – ensuring the analysis stayed aligned with actionable insights.
  
• Cross-Industry Applicability: While this story comes from the insurance industry, customer churn is a universal challenge in telecom, banking, SaaS, e-commerce or any subscription-based business. The methods used here – from classification models to survival analysis – can be adapted to any sector where keeping customers is key. The success in this case demonstrates that with the right data and expertise, predictive analytics can significantly improve customer retention anywhere.

In today’s competitive markets, understanding and preventing churn can be the difference between stagnation and growth. AI and data science proved to be the insurer’s ally in this story, turning a mountain of data into a crystal ball for customer behavior. The Tesseract Academy team not only delivered a technical solution, but a new mindset for the client: churn went from an inevitable cost of doing business to a preventable outcome with clear signals and timely action.

Ever wondered what it could mean for your business if you knew ahead of time which customers were about to leave? This case shows that it’s possible to find out – and to make a difference. By leveraging your data and the power of machine learning, you can keep more of your hard-won customers happy and loyal. In the end, the best way to grow your business might not be gaining more customers at any cost, but rather loving the ones you have – and using AI to help make sure they stay. 

📧 Get in touch for a free 30-minute consultation and see how AI-powered churn prediction can boost your customer lifetime value.