The situation
Most retention activity in these organizations was reactive. A customer called to cancel, and a retention offer followed. By that point the decision had usually been made some weeks earlier, and the offer was expensive precisely because it arrived late.
The approach
Active risk triggers. Certain events reliably precede attrition: a poor customer service experience, an unexpected fee, a difficult claims interaction, a price change. Modeling these as events with a decay window allowed intervention within the period where it could still change the outcome.
Passive risk indicators. A second class of signal is quieter and easier to miss. Gradual inactivity on an account, or a demographic change such as a house move or adding a teenage driver to a policy, shifts the probability of cancellation without producing any obvious event. This behaves more like a survival problem than a classification problem, and modeling it that way recovered signal that event-based approaches had been discarding.
Win-back. Former customers are not a homogeneous population. Profiling past customers by the circumstances of their departure and their behavior beforehand identified segments where a targeted product bundle had a genuine chance, rather than mailing the entire lapsed base and accepting a sub-1% response.
The outcome
Trigger-based intervention reduced churn by 10% to 25% on a net-positive basis. Passive indicators contributed a further 3% to 7%. Win-back campaigns converted at 3% to 5%, against an industry norm below 1%.
What made it work
Holding every intervention to net profitability rather than to churn reduction. It is straightforward to reduce churn by giving away margin, and a retention program measured only on retention will do exactly that until somebody in finance notices.