The two ways people leave look nothing alike
Slowing down before churn and speeding up before churn are opposite stories. Most models flatten them into one.
Watch two users in their final week before they cancel. One logs in less, clicks less, stays for shorter sessions. The other logs in more, exports data, downloads invoices, copies integrations into a spreadsheet. Both churn on Friday. Your dashboard calls them the same thing.
They are not the same thing.
The first pattern — deceleration — is what most churn models are built to detect. Activity slopes downward. Sessions get shorter. Features that used to get daily use go untouched. The interpretation is straightforward: this person lost interest, lost the habit, or found something else. The signal is absence. The story is disengagement.
The second pattern — acceleration — looks almost like engagement if you squint. Pageviews go up. Time on site might even increase. But the texture is different. These users aren't exploring new features. They're visiting settings pages, export tools, billing history. They're extracting. Gathering what they need before they close the door. The signal is presence, but the story is departure with luggage.
We looked at a dataset recently where roughly 40% of churned accounts showed acceleration in their final 5 days. Not slight acceleration — meaningful spikes in activity, concentrated in administrative and data-export areas. The other 60% decelerated in the way you'd expect. The two groups had almost nothing in common behaviorally, except that they both ended up cancelled.
Here's what that distinction costs you when it's invisible. Decelerating users might respond to re-engagement. A prompt, a check-in, a "here's what you missed" email — these make sense when the problem is drift. The person forgot, or the product slipped out of their routine. There's a plausible path back.
Accelerating users are past that. They've already decided. The spike in activity is the execution of a decision, not a signal that a decision is being made. Sending them a re-engagement email doesn't just fail — it confirms they made the right call, because you clearly weren't paying attention to what they were actually doing. The intervention for this group, if one exists, happened two weeks earlier, when something broke the relationship in a way that didn't show up as declining usage.
Most "at risk" models compress both patterns into a single score. They look for deviation from baseline. And acceleration is a deviation — but it gets scored inconsistently. Sometimes the model reads it as positive engagement. Sometimes it reads it as anomalous. Rarely does it read it as what it plainly is: someone packing boxes.
The reframe is this: churn isn't a single behavior. It's at least two, and probably more. Disengagement is the slow leak. Extraction is the clean break. Treating them identically — with the same alert, the same playbook, the same urgency — is like treating a cough and a packed suitcase as the same symptom because both preceded someone leaving the house.
One smaller thing worth noticing. The accelerators often have higher lifetime value. They used the product seriously enough to have data worth extracting. They built workflows, saved configurations, accumulated history. These aren't casual users drifting away. These are committed users who made a deliberate decision to leave. The loss is different in kind, not just in timing. And the cause is almost never "they forgot to log in."
If your churn model can't tell the difference between someone fading out and someone backing up the truck, it's answering a question that's too simple to be useful.
