Use case assessment
Is AI churn prediction worth building?
The prediction usually works. The intervention is what's missing, and no model supplies it.
Predicting churn is a solved classification problem. Most projects fail after the prediction, because nobody defined what happens when the score goes red.
What it usually means
Scoring customers on likelihood to leave, so someone can intervene.
Is it genuinely AI?
Yes, in the ordinary machine learning sense. Given usage, support and billing history, predicting churn is a well-understood classification problem and models do it competently.
Which is why it is a strange thing to keep failing at.
Where it actually fails
Downstream. The scores get produced, they land in a dashboard, and then one of three things happens.
Nobody acts. No team owns intervention, so the list is information rather than a workflow.
Everyone acts identically. Every at-risk account gets the same discount email, which trains customers to become at-risk.
The prediction is too late. By the time usage has dropped enough to be detectable, the decision was made weeks ago in a meeting you have no signal from.
None of those are model problems, and improving the model does not touch any of them.
What the simpler version looks like
Pick the two or three signals your customer success team already believes in — usage down more than half, champion left, support ticket unresolved past a threshold. Build the alert on rules. Run the intervention for a quarter.
You will learn whether intervention works at all, which is the thing a model cannot tell you and the thing the whole project depends on.
What it costs to run
The model is the cheap part. The expensive parts are the feature pipeline, retraining as the product changes, and the human time to act on scores — which is the real cost and the one that decides whether this pays back.
A useful check: what is the value of a retained account, times the number you would realistically save, against the cost of the team acting on the list?
Where it sits under the EU AI Act
Not an Annex III high-risk use in a normal commercial context. Worth noting that creditworthiness assessment of natural persons is Annex III, so if your churn score influences credit terms or pricing for consumers, the classification question is live.
When it is worth building
When you have proven that intervention changes outcomes with a rules-based version, and the constraint is that rules miss too many accounts. That is a real and common situation, and it is the point at which a model earns its cost.
Not as the first step. Prediction without a proven intervention is an expensive way to be accurately pessimistic.
Common questions
Why do churn prediction projects fail?
Almost never on the prediction. They fail after it: nobody owns intervention, every at-risk account gets the same treatment, or the signal arrives after the decision was already made. None of those improve with a better model.
What should you build before a churn model?
A rules-based alert on the two or three signals your customer success team already trusts, and a quarter of running the intervention. That tells you whether intervention works at all, which is what the whole project depends on and what no model can answer.
Is churn prediction high-risk under the EU AI Act?
Not in a normal commercial context. Creditworthiness assessment of natural persons is an Annex III high-risk use, so if a churn score influences credit terms or consumer pricing, the classification question becomes live.
Acquiring 24,000 users for a fintech lifestyle app — retention treated as a product problem rather than a scoring one.