The math that was going to end the company
Seventy-five percent annual cancellation means you replace three quarters of your customer base every year just to stand still. Every dollar of acquisition spend buys a customer who leaves before they pay for themselves. Growth becomes a treadmill that gets faster the harder you run.
The team was good at acquisition. That was the trap. Strong top-of-funnel numbers made the business look healthy right up until you looked at the second year of any cohort, and by then the spend had already been committed.
The mandate I was given was blunt: fix this or the platform is not viable.
It was never a features problem
The instinct in that situation is to ship. Customers are leaving, so build what they are asking for, and build it fast.
I started somewhere else. I mapped the full customer journey, interviewed customers who had already cancelled, and compared usage patterns between the accounts that stayed and the accounts that left.
The divergence did not happen at month eight when people cancelled. It happened in the first two weeks.
Accounts that reached a specific moment of real use early stayed. Accounts that did not, left, and they left regardless of what got shipped afterward. Every feature built for the churning cohort was being built for people who had already effectively decided, they just had not filed the paperwork yet.
So the cancellation rate was not a cancellation problem. It was a first-mile activation problem showing up seven months later, which is exactly the lag that makes this kind of failure so hard to diagnose from inside. By the time the number moves, the cause is a long way behind you.
The rebuild
The roadmap stopped serving feature requests. One hundred percent of it went to retention. This was the decision that mattered and it was the unpopular one, because a roadmap full of retention work looks, to everyone outside product, like a roadmap where nothing new is happening.
Onboarding was redesigned around activation, not around explanation. The goal was not to teach the product. It was to get a shop to the moment where the product had visibly done something for them, as fast as possible, because that moment was the actual predictor.
Engagement loops and personalization were built to turn early use into habit. Software that serves a working repair shop has to earn its place in a daily routine, and habit is not a feature you ship, it is a property you design toward.
A customer advisory board put real operators in the loop, which changed what got prioritized and, as much, changed how fast a bad idea got killed.
Instrumentation for leading indicators. This is the part that made it durable. Tracking cancellation tells you about a decision that has already happened. Tracking the early-use signals that predict it tells you months ahead, while there is still something to do about it.
What held
Cancellation fell from 75% to under 30% within six months.
Then the platform sustained 10 to 12% annual subscription growth for three consecutive years.
That second number is the one I would point to. A six-month turnaround can be a burst of attention. Three years of compounding growth after the attention moved on means the system changed, not just the effort level. The instrumentation kept working. The activation design kept working. Nobody had to keep rescuing it.
Churn is a lagging indicator of something else
Cancellation is a downstream symptom. It shows up at the end of a chain that starts with decisions about onboarding, workflow, and platform architecture, made months or years earlier by people who were not thinking about retention at all.
Which means optimizing the cancellation moment is close to useless. Save-offers and exit surveys operate on a customer whose experience already told them what they needed to know. The rate does not move until you go back upstream and change the decisions that produced it.
That is the harder work and it is slower to show results, which is why most teams do the other thing. But it is the only version that holds after you stop paying attention to it.