Understanding Escalation Prediction: What’s Predictable, What’s Preventable, and How to Measure Both

Defining escalations rigorously, separating the predictable from the preventable, and measuring model performance without fooling yourself.

Escalation prediction has become a standard claim among AI vendors in customer support, yet the discipline of evaluating those claims remains immature. The ultimate goal of escalation prediction is escalation prevention. This distinction is not rhetorical; it is the organizing principle for every design and measurement decision in this paper. A prediction that does not change an outcome has marginal value — it is, at best, an early-arriving postmortem. A prediction that arrives early enough for a human or an agentic workflow to intervene has immense value: it converts an escalation that would have happened into one that never did.

This paper presents the framework SupportLogic uses across millions of B2B support interactions to make escalation prediction rigorous.