The account that never looked sick
Every health metric was green. Then they were gone. Here's what the sequence actually said.
The account scored an 82 out of 100 in their platform's health model the week before they cancelled. Logins were consistent. Feature adoption was broad. NPS response from their admin, three months prior: 8. By every composite metric, this was a satisfied customer.
They churned on a Tuesday, no warning ticket, no downgrade, no pricing conversation. Just a cancellation form with "no longer needed" selected from the dropdown.
We went back and read the sequence. Not the scores — the actual journey, session by session, over the final 90 days. What we found was a pattern we've started calling "the clean exit," and it's one of the hardest things for a health model to catch because it doesn't look like distress. It looks like competence.
Here's what happened. Around day –87, a second user from the account — not the admin, someone more junior — started visiting the integrations settings page. Not configuring anything. Just viewing. Three times in two weeks. On day –64, the admin exported their full data set for the first time. The export feature existed, they'd had access for over a year, and they'd never touched it. On day –59, the admin logged in, spent 40 minutes in the product, touched six features, and logged out. It was the longest session in months. On day –31, both users stopped logging in on Mondays. Their prior pattern had been Monday-Wednesday-Friday, almost like clockwork, for nine months. Now it was Wednesday-Friday. Then just Wednesdays. Then the cancellation.
None of these moments, individually, would trigger an alert. The integrations page view isn't an error state. The data export is a feature working as designed. The long session looks like engagement. The Monday drop-off is a scheduling change. Each event has a boring explanation, and boring explanations are what automated systems prefer.
But read as a sequence, the story is different. Someone on the team started quietly evaluating whether their data could move somewhere else. The admin pulled a backup — not because they were being cautious, but because they were preparing. That long session wasn't engagement; it was a farewell tour, someone making sure they'd gotten what they needed before they left. And the cadence shift wasn't a calendar change. It was a team that had already found an alternative and was winding down usage gradually, the way you stop going to a gym before you cancel the membership.
The health score missed all of this because it was built to measure what happened, not in what order and for the first time. A data export by a power user who exports monthly is routine. A first-ever data export by a 14-month-old account is a completely different signal. The model treated them identically.
We've seen three variations of this pattern in the last quarter across different accounts we observe. The details change — sometimes it's a permissions audit instead of an export, sometimes the cadence decay is faster — but the shape holds. A team that has decided to leave becomes briefly, unusually thorough. They visit corners of the product they'd ignored. They make sure they have everything. And then they go quiet in a way that's easy to mistake for a holiday week.
One counter-note: we also found a sequence that looked almost identical — integrations browsing, first export, cadence shift — where the account didn't churn. They were migrating to a new internal system and re-integrating the product into a different workflow. Same events, different story. The distinction was that their junior user configured an integration after browsing; the churned account's user only looked. That's a thin line. We're not sure a rule could reliably catch it.
The uncomfortable part isn't that this account churned. Accounts churn. The uncomfortable part is that the system said they were healthy while they were packing their bags, and no one read the sequence closely enough to notice the suitcases by the door.
