What it does
Rank a book of customer accounts on the signals you actually measured — usage trend, support volume, days since login, NPS, tenure — reporting each account's rank on each signal among the accounts that reported THAT signal, so the denominator is honest: "worst of the four who told us" is a different claim from "worst of twelve". A rank is ordinal. It needs no coefficient, no baseline and no threshold, so nothing borrowed from another company's book of business can distort it. Direction stays, because direction is definitional rather than invented: more support tickets is worse, a higher NPS is better, and that is a property of what the number measures. How much each signal matters is yours — supply weights and you get your score, echoed back, computed over the signals each account actually reported; omit them and nothing is scored. A MISSING SIGNAL IS MISSING, NEVER NEUTRAL. In a weighted model an absent feature contributes its weight times zero, so an account nobody has data on lands mid-table and reads as unremarkable — and the accounts with the thinnest data are exactly the ones nobody has looked at. Here an absent signal is named per account, counted, and an account reporting fewer signals is marked not comparable to one reporting more. An account carrying no signal at all is reported unread rather than ranked last in every column, which would read as the healthiest account in the book. WHAT IT WILL NOT GIVE YOU, and why that is the product. The agents this consolidates all produced a churn probability, and every one looked rigorous doing it. One carried a logistic regression whose six coefficients were typed by hand, pushed through a sigmoid and printed as "0.73" beside an account name — beside a model-health tool quoting validation AUC for a model that was never fitted. Another weighted four signals 35/30/25/10 against a hardcoded monthly-active-user baseline. A third multiplied a risk tier by contract value to produce expected lost revenue in dollars, the kind of figure that goes in a board deck. None of it came across.
Rank a book of accounts on the signals you measured — no churn probability
Example prompts
- Here are 40 accounts with usage trend, support tickets and days since login. Which ones are worst on more than one signal?
- I only have support ticket counts for half the book. Rank what you can and tell me which accounts I am missing data on.
- Rank these accounts, weighting usage trend twice as heavily as ticket volume.