What is Propensity Model?

Propensity models are the upgrade path from rule-based lead scoring. A rule-based score says "VP title plus 100-500 employees equals 80 points." A propensity model says "based on 12 months of historical conversion data, prospects with this signal combination convert at 14% within 90 days." The model uses what predicted outcomes in the past instead of a sales leader's guess about what should matter.

For GTM Engineers, propensity models show up in three contexts. The most common is conversion propensity for inbound leads, where the model scores form-fill leads on likelihood to convert to opportunity. The second is churn propensity for customer accounts, scoring usage and engagement signals to flag at-risk renewals 60-90 days before the renewal date. The third is expansion propensity, predicting which existing accounts are likely to upsell or cross-sell within a quarter.

Building a propensity model from scratch requires three ingredients. You need historical outcome data with at least 500-1,000 examples of the event you're trying to predict (closed-won deals, churns, expansions). You need feature data about each example at the time of decision, not after the fact. And you need a modeling tool, which can be as simple as a logistic regression in Python or as sophisticated as a gradient-boosted tree in scikit-learn or XGBoost.

The cheap version uses a CDP or RevOps platform with built-in propensity features. MadKudu, 6sense, and HubSpot's predictive lead scoring all offer pre-built propensity models that train on your CRM data. Setup is days, not weeks. Accuracy is usually 60-70% as good as a custom model, which is fine for most use cases. Custom models matter when your deal economics justify the engineering investment, typically deals above $25K ACV with sales cycles long enough for misallocated SDR effort to be expensive.

The trap with propensity models is treating them as truth instead of guidance. A 14% predicted conversion rate means the prospect is more likely than baseline to convert. It does not mean the prospect will convert. Sales teams that get propensity scores wrong assume high-score leads close themselves and low-score leads should be ignored. Both behaviors damage pipeline. The right operational model uses propensity to prioritize effort, not to substitute for it.

Model maintenance is the part most teams skip. A propensity model trained on 2024 data starts losing accuracy in 2026 as your ICP shifts, your product changes, and market conditions evolve. Retrain quarterly with the most recent 12 months of outcomes. Monitor predicted-versus-actual conversion rates monthly. If the gap exceeds 20%, your model has drifted and needs retraining or feature updates. A propensity model nobody monitors is a propensity model nobody should be using.