Segmenting SAP's Churn

I led the team. I directed the analytical approach, pushed the segmentation strategy, and owned the recommendation we presented to the judges. We placed first out of forty teams.

SAP's AI investment was supposed to deepen engagement and was instead creating a new kind of churn risk. One group could not figure the features out and stopped using the product. Another group used them heavily and found the personalisation too shallow to be worth it.

Treated as a single churn number, those two populations average into something meaningless. Any intervention aimed at the average would have been wrong for both.

We started with bivariate and correlation analysis, and the finding that reframed the problem was a strong negative correlation between age and AI interaction level. Low interaction and low satisfaction were both driving churn, but in entirely different populations.

That is what moved me to push for cluster analysis rather than treating churn as one monolithic metric. Looking at the density distributions across AI interaction level, satisfaction, personalisation effectiveness and response time, the groups were clearly separable rather than one smooth population.

Three segments came out. High-satisfaction interactors: heavy users, satisfied, not at risk. Unsatisfied service seekers: younger, highly engaged, interacting constantly and finding the personalisation shallow. And aging churners: an older cohort with the highest churn rate and almost no AI engagement at all.

The third group is the one that justified the whole approach. They were not dissatisfied with the AI features. They were not using them. No amount of improving the feature would have reached them.

Each segment got a different intervention. For the aging churners, a change management support service with consultants working directly with organisations to drive adoption, because the barrier was onboarding rather than quality. For the unsatisfied seekers, a machine learning personalisation engine adapting UI, workflow and process suggestions to individual behaviour, because the barrier was depth.

We also proposed surfacing AI suggestions inline at bottleneck points in the user's real-time process analysis, rather than requiring people to go looking for them. We called the combined framework ARC, built on the principle that you cannot fix churn you have not properly segmented.

We recommended SAP evaluate its AI solutions through an ethical lens covering accountability, privacy, bias and accessibility, since the adoption barriers we found fell along age lines and had real equity implications.

Segmentation first is the whole lesson. Without splitting the problem into two distinct populations, every solution we could have proposed would have partly addressed one group while ignoring the other, and the averaged metric would have hidden which.