A hunch, tested

Everyone on the team has a theory about why customers leave. This tests them. Each warning sign is a fact about a customer’s first two weeks (“didn’t use it in week two”, “connected nothing in week one”), counted on every past customer, with the cancel rate for those who had it and those who didn’t. The strong ones become points in Monday’s risk score. The ones that never help are retired.

What it does

• Reads the stories of customers who left and customers who stayed, and finds what the ones who left did differently.
• Writes a sign for each one it can count from your usage, and says plainly why the rest can’t be a sign.
• Tests every sign each week and prints one table: in their first two weeks, the customer did this, and this share cancelled later, versus that share of everyone else.
• Retires signs still unproven after four weeks.
• Grades the score honestly: only on customers it did not learn from, replayed in time, and against a one-line rule anybody could write.

The reveal

Customers who used it on two days or fewer in their first two weeks cancelled 55% of the time. The rest, 18%. And every customer who asked what it costs stayed.

Configuration

• The stories: from Why They Leave and Why They Stay.
• Your usage source: whatever holds what customers did, day by day.
• How often: when you ask, or every Sunday, if an available scheduler can run the work and you approve it.

What it doesn’t do

• It never lets a sign look at how a customer ended. A sign must be something you could see while they still pay.
• It doesn’t trust a sign on fewer than six customers.
• It doesn’t write to customers.

Triggers

Runs when you ask (“what are the warning signs”, “is the risk score any good”). A weekly test needs a verified scheduler and your approval; otherwise it runs one test now and saves where it got to.