Which customers should we call?

How it works

The model gives every customer a score: how likely they are to leave next month, or to move up a plan in the next 90 days. The score reads what the product already records. For leaving: days with a login, seats in use, support tickets, failed payments, the Growth plan, the Scale plan, tenure and seats on the account. For upgrading: the plan, tenure, seats on the account, days with a login, days at the plan’s limit and support tickets. The list calls the highest scores first.

Each model is fitted on earlier history and tested on customers whose outcome it never saw: the leavers from April to September 2026, and the upgrades in the 90 days after 30 June 2026. Every figure on this page comes from those test periods.

Reading the chart

The chart is a cumulative gains chart. Along the bottom is how far down the list you call; up the side, the share of all the leavers, or upgrades, the list has reached by then. A list in random order follows the dashed diagonal: a tenth of the calls reaches about a tenth of them. The further a line bows above the diagonal, the better the ranking. Lift is that gap as a ratio: the top 10% of each month’s churn list held 3.6 times its share of leavers, and the top 10% of the upgrade list 4.5 times its share of upgrades.

The rule it is measured against

The baseline is the rule a team uses without a model. For leaving, it calls the customers with no login for 30 days. For upgrading, it calls the customers at their plan’s usage limit on 5 or more of the last 30 days, the threshold at which that rule reaches the most upgrades on the top 10%. A rule flags customers but does not rank them, so its list calls the flagged customers in random order, then customers picked at random; its figure is the average over every such order, worked out exactly. At every depth from 1% to 50%, the model’s list reaches more customers, and more revenue, than the rule’s, for both goals.

What this shows, and what it does not

Revenue at risk is what the leavers on the list paid a month; a call saves only the part it keeps. Upgrade revenue is what the customers on the list added by moving up, and the upgrade history holds no sales calls, so every one of those upgrades happened without one. On your own customers, a random part of the list left uncalled, compared with the called part at the end of the period, measures what the calls add. Where to stop calling depends on what a call costs, which is not priced here.

The figures show how the method behaves on simulated customers. On your own data, the model is fitted to your history, and putting it live carries the analytics director’s sign-off.

The full analyses: the example churn model readout, with how well the model ranks customers and which signals drive the score, and the example propensity-to-upgrade readout, with how the model was built and the list scored today.