In a typical classification problem you need some examples/observations of the positive class (customers who have already churned). You can start by defining a time threshold for what you would consider to be a churn event, e.g., a customer not having purchased in the last 90 days, or not having visited the website in the last 30 days, etc. This logic can help you construct a target variable with boolean values (only two values) for each customer, i.e., churned vs. active. The next steps are feature engineering and churn modelling in order to predict which customers are likely to leave so that you can engage in proactive campaigns to prevent this.
You can also build a separate model to see what type of promotional offers and campaigns tend to result in recapturing lost customers. By definition, this activity can only start after you have some reactivation campaigns launched and responses collected. For example, you can find if email vs. phone call or a $50 coupon vs. a 2 for 1 offer tend to perform better. Combinations of input variables such as communication medium, promotional offer, marketing channel or timing of the message can multiply in numbers making the model insights much more powerful than any manual analysis methods. This type of closed loop feedback data can in turn help you set up response models on previously churned customers, as a multi-class classification model.
A third option is to do cluster analysis on the churned customers to see how they are similar or different. Based on the degree of differences you may want to create a separate response model per cluster instead of one generalized one.