One of the most expensive problems a company has is customers who quietly leave. In this project I looked at 7,043 customers of a telecom company, from a public sample dataset by IBM, and asked the three questions a customer team would ask: who are we losing, what does it cost, and what should we do about it?
What the numbers say
About a quarter of the customers (26.5%) cancelled, and they took nearly a third of the monthly revenue with them: $139K of $456K.
The interesting part is who left:
- People without a contract. 42.7% of month-to-month customers left; on two-year contracts it was only 2.8%.
- Newcomers. Almost half of the customers in their first year (47.4%) left.
- People who pay by hand every month. 45% of them left, against 15 to 17% of those on automatic payment.
Most of the loss sits in one small group
Customers with all three traits (month-to-month, on fibre internet and in their first year) are only 13% of the base. Yet they account for 38% of the lost monthly revenue, and 70% of them left.
So rather than trying to fix everything at once, it makes sense to start with that small group.
What could be done
The data points to three simple ideas:
- Offer people who pay by hand a one-off credit for switching to automatic payment.
- Offer month-to-month customers a discounted one-year contract.
- Give fibre customers a free tech support trial. Internet customers with tech support leave at 15%, those without at 42%.
I also estimated roughly how much revenue each idea could keep; those are rough numbers built on assumptions. All three meet in the same small group, which makes it the natural place for a first try.
Numbers anyone can check
One thing mattered to me while building the report: under every chart there is a panel showing how that number was calculated. Anyone who doubts a figure can see where it came from.
I also kept the 11 customers that looked incomplete in the data. A closer look showed they had joined that month and had not been billed yet, so they were real customers.
The report itself is here (opens in a new tab), and the code is on GitHub (opens in a new tab) for the curious.