Predictive scoring

Predictive scoring keeps three numbers current on every contact: churn probability, predicted lifetime value and purchase propensity. They are written onto the contact record, which is the part that matters — they appear in the segment builder as ordinary numeric fields ("ML churn probability", "ML purchase propensity"), so a high-risk audience is a filter you set in the UI and then hand to a journey, exactly like any other rule.

There are two engines behind the scores, and this is the honest version. Out of the box, the platform uses always-available heuristic scoring derived from the behaviour it already tracks — recency, engagement, order history. Deploy the Python ML sidecar and it trains per-account models on your own data and takes over the same three columns. The distinction matters because it means the feature is never a blank screen waiting for a data scientist: it works on the first day and gets better later, and every scoring run is recorded so you can see which engine produced the current numbers.

A summary endpoint reports coverage and distribution — average churn and propensity, total predicted value, and how many contacts sit above the high-risk and high-intent thresholds — so you can sanity-check the scores before you build a campaign on them. Scores also flow into warehouse exports for your own analysis, and they are erased when a contact is anonymised for a GDPR request, like every other derived attribute. Predictive scoring is a Business-and-above feature; see the docs for the field-level detail.