In Part 1 of this series, we explored how we built our trigger system, using behavior-aware thresholds to detect early signs of churn in our clients. That work laid the foundation for everything else we discovered in part 2 of this series. It gave us a way to listen to the data and catch the moment that things start going wrong.
This next part builds directly on that foundation and focuses on what came after we had identified our behavioral signals. Instead of simply analyzing risk scores, we started paying attention to the behavioral combinations that led there. Over time, we began to notice consistent patterns. For instance, certain triggers often fired together, meaning customers at risk were following recognizable paths and patterns.
Two things helped us make sense of those paths. The first was the churn model.
Discover what AI can really do for your business by exploring Impact's AI Consulting Services.
The Churn Model
A churn model, like the one we called Version 6.3, estimates the likelihood that a customer will leave based on past behavior. It looks at a wide range of features and assigns a contract renewal probability score. The higher the score, the lower the probability of renewal. But a churn model only answers part of the question. It tells us who is at risk and how high that risk is.
The second tool was our survival model, Version 7.
The Survival Model
Survival models do something different. Instead of predicting whether churn will happen, they estimate when it’s most likely. This is especially useful in the context of contract renewals, where time is a major factor. It’s one thing to know a customer is at risk, but it’s much more powerful to know that they have a risk score of X, indicating they’re likely to churn in the next 45 days based on historical data and pattern recognition.
That kind of insight creates urgency and helps prioritize outreach.
Together, these models gave us something new. A way to describe customer behavior not just in terms of risk, but in terms of timing and context. When combined with the sequence of triggers that fired, each contract became a behavioral profile.
For example, some customers looked risky but were likely to stick around for a while. Others seemed quiet but had urgent churn windows. These combinations created a richer view of the customer journey.
To make sense of all these patterns at scale, we used a clustering technique called K-Prototypes. This method is well-suited for real-world customer data because it can group customers based on both numbers and categories.
Think of clustering as sorting customers into buckets, not just by how much risk they have, but by the type of risk and how it formed. K-Prototypes compares contracts based on any number of metrics, including but not limited to:
- Scores
- Department names
- Service types
- Specific triggers activated
Instead of assigning customers to categories based on a single feature, the algorithm looks across all the features and finds natural groupings where customers share similar patterns. It then learns from the data to identify which combinations matter most.
For example, it might find that contracts with low engagement scores and a high number of billing flags tend to behave similarly. That becomes one group. Another group might contain long-term customers with no recent activity and a high survival risk. These become segments with shared traits.
Clusters Becoming Clearer
Clusters began to take on shape and meaning. One group showed billing issues with lots of time remaining to intervene, while another revealed accounts that had disengaged over time. Some customers were silent but stable. Others were active, but the signals said they were already shopping around.
These weren’t just groups on a chart. They became personas, or behavioral identities.
For instance:
- The billing group became a cue for early finance reviews
- The disengaged needed re-engagement campaigns
- The stable-but-silent accounts called for low-effort outreach
- The loyal ones might be perfect for a referral ask
Each cluster told a different story about how risk accumulates. But more importantly, they gave us a practical way to act.
Instead of creating one-size-fits-all strategies, we could plan outreach based on what each group needed most.
These clusters allowed our sales and customer success teams to speak to each customer not as a statistic, but as a story. They provided the foundation for contract-specific planning, but they also opened up something bigger: department-level pattern recognition.
Because clustering considered service lines and internal department assignments, we began to see where specific teams were running into trouble. Some departments had high churn among newer accounts, while others struggled with billing-related triggers. This insight let us go beyond customer conversations to internal alignment.
There was now a practical reason to loop in operations, support, and product teams. Because the clusters told us more than a churn risk, they told us what was broken and where. In doing so, they set the stage for forecasting that goes beyond customer loss and identifies organizational opportunity.
Unleash the full potential of AI in your business with the help of Impact's AI Consulting Services.
