Churn metrics

Day 17: Churn

Churn sits at the bottom of your dashboard, boring, until the month it isn't — then it's the only number anyone in the room cares about. If you can't say why people left, you can't stop the next ones from leaving. And in AI products, people leave faster, and a lot more quietly, than you'd expect.

What churn actually is

Churn is simply the rate at which customers or users stop using your product over some fixed window — a month, a quarter, whatever cadence you track. It matters because every growth number you love — signups, MAUs, revenue — can lie to you. You can add users every single month and still be bleeding out if you're losing them faster than you bring them in. Churn is the number that forces you to look at the bucket instead of the tap.

• ▹Customer churn (logo churn) — the % of accounts that cancel in a given period
• ▹Revenue churn — the % of recurring revenue you lose; it weights big accounts more heavily than logo churn does
• ▹Voluntary churn — the user chose to leave: wrong fit, found something better
• ▹Involuntary churn — a card expired, a payment bounced. Looks like churn, but it's really a billing-ops problem
• ▹Gross vs net churn — net churn nets out expansion revenue from existing accounts, and can actually go negative (a very good sign) even while logo churn is still positive

Churn rate = (customers lost in period) / (customers at start of period)

Example: 1,000 customers at start of month, 40 cancel
Churn rate = 40 / 1,000 = 4%

That shape — a steep drop in the first weeks, then a flattening — is what you want to see in any product's retention curve. The flat part is your real number; that's who actually stays. If the curve just keeps sliding and never levels off, you don't have a churn problem, you have a product problem: nothing about what you built is sticky enough to hold on to.

Churn in the AI era

AI products churn differently than the SaaS products most churn playbooks were written for, and most teams are still measuring it the old way. Here are the four shifts I actually watch for, building agentic tools and running the teams that ship them:

• ▹Trust churn — in traditional SaaS, a bug annoys someone. In an AI product, one confidently wrong answer, or one destructive action an agent takes on its own, can end the relationship in a single session. Trust doesn't fade here, it cliffs.
• ▹Silent churn — the seat stays paid, but usage of the actual AI feature quietly drops to zero while the user goes back to doing the task by hand. Subscription-level churn metrics miss this completely — you need feature-level and task-level usage as your leading indicator, weeks before the cancellation ever shows up.
• ▹Cost-driven churn — the moment you reprice around inference cost (rate limits, tiered models, usage caps), you'll get a churn spike that has nothing to do with product quality. Teams that don't separate this from 'organic' churn end up chasing the wrong root cause for months.
• ▹Task churn in multi-agent systems — if you're running agent fleets or multi-step workflows, track abandonment per task type, not just per account. One agent role that keeps failing — say, a code-review agent that nags people with bad suggestions — can drag down retention for the whole product even while every other agent works fine.

When to use it — and when not to

Treat churn as a confirmation metric, not a steering wheel. By the time it moves, something already went wrong weeks or months ago — it's a mirror, not a dashboard light. For the decisions you make day to day, pair it with a leading indicator instead; in AI products that's usually task-completion rate, or what I'd call 'time to trusted first result,' because that's what actually predicts churn before it happens. And don't grade a single feature launch or a single agent change against raw logo churn — the sample's too noisy and the lag is too long. Use short-term engagement or task-success deltas for that, and let churn confirm the trend a few months later.

• ▹Pitfall: lumping voluntary and involuntary churn together — fixing a broken payment-retry flow and fixing a product-market-fit problem are not the same job
• ▹Pitfall: only measuring at the account level — in AI products, usage can die inside a still-active account long before anyone cancels
• ▹Pitfall: reacting to one bad month like it's a trend — check whether a pricing change or a model-version bump happened before you panic
• ▹Pitfall: treating all churn as bad — some of it is users who were never a fit, and losing them can be net-positive for your support and infra load
• ▹Pitfall: not splitting churn by task or agent type in multi-agent products — one weak agent's failures get averaged away in the aggregate number and never actually get fixed

Extend your knowledge

• ▹Read a cohort retention writeup (Mixpanel's or Amplitude's docs are good starting points) and map that Week 0–4 curve shape onto a product you use every day — where does it actually flatten?
• ▹Find out how your team currently reprices or rate-limits AI features, and check whether churn spikes around those dates get separated from 'organic' churn in your reporting
• ▹If you run any agent-based workflows, instrument task-level success and abandonment separately from account-level usage — that's your earliest churn signal, full stop
• ▹Lenny's Newsletter is a solid place to go deeper on SaaS metrics like retention curves and NRR, to pair with the AI-specific angle in this lesson