The Decision Intelligence Handbook

Chapter 2. Decision Requirements

A subscription news service in the United Kingdom asked us to help it to come up to speed with using machine learning (ML).¹ This business was an insurgent in a crowded marketplace. It wanted to grow rapidly, our contact explained, while still retaining its long-standing reputation for excellent customer service. Could an ML strategy lower the cost of a high-touch relationship with the company’s most important customers? We visited the news service in person twice—once for an AI strategy planning workshop and then a second time to teach its software engineers how to develop an ML model.

To set things up in advance of our ML training, we held a few phone calls to choose a project through which we could train the team while also helping the news service accomplish its goal. The company selected to build a churn model: a system that predicts which customers are most likely to close their accounts and switch to a competitor. ML-based churn models have been widely successful and can be applied in a broad range of situations: for instance, similar systems are used to predict who will vote in an election, who is likely to continue donating to a nonprofit, and more.

As we were preparing, the director of analytics invited us to a Zoom meeting with his boss, an executive and company owner who was skeptical of the project. The executive spoke frankly:

I don’t believe in machine learning for churn prediction. And here’s why: just predicting churn isn’t enough ...