A churn propensity model is a scoring system that evaluates every customer against a set of risk dimensions and produces a single number — a churn propensity score — that represents the customer's likelihood to cancel their contract.
Done well, a churn propensity model is one of the most operationally powerful tools a B2B customer success team can have. Done poorly, it's an expensive spreadsheet that CS reps ignore because it doesn't match their intuition about which customers are at risk. The difference between the two usually comes down to what inputs the model uses and how it weights them.
The Core Architecture of a Churn Propensity Model
Every churn propensity model follows the same basic structure:
- Inputs: Raw data from multiple sources — support systems, product usage, CRM, survey responses
- Dimensions: Groups of related inputs combined into meaningful signal categories
- Weighting: Each dimension contributes to the final score according to its predictive weight
- Score: A final composite number (typically 0–100) that represents churn risk
- Threshold: Cutoffs that categorise customers as High, Medium, or Low risk
The model needs to update continuously — weekly at minimum, daily ideally — because customer risk can change quickly. A customer who was Low risk last month can become High risk this week if they receive a series of SLA breaches or give three consecutive low CSAT scores.
The Five Dimensions That Matter Most
Dimension 1: CSAT and Customer Sentiment (highest weight)
CSAT data from support interactions is typically the strongest single predictor of churn in B2B environments. The most predictive CSAT signals are:
- Average CSAT below 3.5 over the last 90 days
- Declining trend (score dropping over consecutive rolling periods)
- Recent very low score (1 or 2 in the last 30 days from any interaction)
In a weighted model, CSAT-related signals typically account for 35–45% of the total churn propensity score.
Dimension 2: Case Pressure
The volume, severity, and recency of unresolved support cases is the second most predictive churn dimension:
- Overdue cases (past SLA deadline)
- Cases at SLA risk (approaching deadline, not yet resolved)
- Cases containing negative sentiment in the interaction history
Case pressure scores typically account for 35–45% of the total churn score, reflecting the fact that support experience is the dominant driver of B2B churn.
Dimension 3: Engagement Level
Engagement measures the volume and recency of the customer's interaction with your product and team:
- Number of open cases (3+ is a stress signal, 5+ is a significant risk signal)
- Days since last resolution (90+ days without a resolved case may indicate disengagement)
Engagement typically contributes 10–15% of the churn score, acting as an amplifier of other signals rather than a standalone risk driver.
Dimension 4: Customer Value
Customer value is not a churn predictor per se — a high-value customer is not inherently more likely to churn. But it modifies the urgency of the churn risk. A High-value customer with a medium churn score deserves more attention than a Low-value customer with the same score. In a propensity model, customer value contributes to the score to reflect the business impact of the risk, not just its probability.
Dimension 5: Renewal Proximity
Renewal proximity converts latent dissatisfaction into active risk. A customer with a churn propensity score of 45 and a renewal in 14 days is in immediate danger; the same score with a renewal in 300 days is a monitor situation. Renewal proximity typically contributes 15–20% of the churn score, with exponential weighting as the renewal date approaches.
How the Score Is Calculated: Weighted Euclidean Method
There are several ways to aggregate dimension scores into a final churn propensity score. A common approach for multi-factor churn models is the weighted Euclidean L2 method:
- Each dimension contributes a raw score normalised to 0–1 (by dividing the raw value by its cap)
- Each normalised score is multiplied by its dimension weight
- The weighted scores are squared, summed, and then square-rooted
- The result is scaled to 0–100
The Euclidean method has an important property: it is non-linear. A customer who scores highly on multiple dimensions simultaneously gets a significantly higher churn score than a customer who scores the same total across dimensions one at a time. This reflects the real-world observation that churn risk from multiple simultaneous signals is greater than the sum of its parts.
Threshold setting: Common thresholds are High Risk above 60 and Medium Risk above 30, but these should be calibrated to your actual churn data. If all your historical churns had scores above 55, your High threshold should be lower than 60 to maximise recall.
Why AI Makes Churn Propensity Models Better
A rules-based churn propensity model uses fixed weights: CSAT counts for 40%, case pressure counts for 40%, and so on. These weights are set based on expert judgment and adjusted manually as the team learns what predicts churn in their customer base. This works — but it has three limitations:
Limitation 1: Fixed weights don't adapt. What predicts churn for your Q1 cohort may not be the same as what predicts churn for your Q4 cohort — different customer segments, different products, different market conditions. A fixed-weight model doesn't evolve.
Limitation 2: Human-set weights reflect past assumptions. If your team believes CSAT is the top predictor, you'll weight it highest. But your actual data might show that renewal proximity interacting with case pressure is more predictive. Human weights encode biases that the data can correct.
Limitation 3: No individual customisation. A fixed-weight model applies the same formula to every customer. But different customer segments may have different churn drivers — enterprise customers may churn due to support quality while SMB customers churn due to price sensitivity. An AI model can learn segment-specific weight profiles.
AI-powered churn models address these limitations by learning from outcomes. When customers churn, the model observes which signals preceded the churn and increases their weights. When high-scoring customers renew despite the prediction, the model adjusts. Over time, the model's weights become a learned representation of your actual churn dynamics rather than a human guess.
What AI Can't Replace
AI-powered churn models are powerful, but they're not a replacement for CS judgment — they're an amplifier. There are things CS reps know that no model can capture: a customer champion who just left the company, a procurement change that signals budget pressure, an informal conversation at a conference where a customer expressed dissatisfaction.
The most effective CS teams use churn propensity models as a first-pass filter — surfacing the accounts that need attention — and then apply CS judgment to determine the right intervention. The model finds the risk. The CS rep understands the context. Together, they determine the action.
Getting Started With a Churn Propensity Model
If you're building a churn propensity model for the first time:
- Start with the data you have. CSAT, open case count, and renewal date are enough to build a functional v1 model.
- Set initial weights based on your team's intuition about what drives churn in your customer base.
- Run the model against your current customers and check whether the highest scorers match the accounts your CS reps are most worried about. If they do, your model is calibrated. If they don't, investigate the discrepancies.
- Track outcomes. After 90 days, review which high-scoring customers churned and which renewed. Use this data to refine your weights.
- Automate. Once the model is stable, connect it to your support data and run it continuously so scores update as new interactions arrive.
A churn propensity model built into your support workflow
SignalHOT calculates a weighted churn propensity score for every customer across five dimensions — updated continuously from your support data — and surfaces your highest-risk accounts automatically.
Request a Demo