How to automate churn analysis with AI

Fix the definition before you touch a query

Churn is four different numbers depending on whether you count logos or revenue, when a downgrade counts, and what happens to an account that failed a payment and returned nine days later. Settle it in writing once and every later run means the same thing.

Prompt: “Remember our churn definitions. Logo churn: an account with no active subscription at month end that had one at month start. Revenue churn: net of expansion, on a monthly basis, in the currency the account bills in. A downgrade counts as revenue churn only, never as logo churn. Involuntary churn, cancelled after a failed payment with no explicit cancellation, is reported as its own category and never mixed into the headline. Cohort by signup month, not by cancellation month. If a query cannot honour one of these rules, say so instead of approximating.”

Where is the number, and where is the reason?

Split the work by what each system knows.

Metabase runs the finance-grade questions your team has already modelled and validated. PostHog and Amplitude answer the behavioural ones.

Prompt: “In Metabase, find our saved questions about subscriptions and cancellations, run the ones matching our churn definitions for the last 12 months, and give me the results as a table with the question name and its collection next to each figure. If no saved question covers involuntary churn, tell me that rather than substituting a similar one.”

The Metabase integration runs saved cards and reads dashboards. Board-facing churn figures then come from a query a human wrote and reviewed, which is the boundary you want.

What did the leavers do differently?

Now the behavioural half, where a companion writing its own queries fits, because the output is a hypothesis rather than a reported figure.

Prompt: “In PostHog, write and run HogQL to compare two groups over their first 60 days: accounts that cancelled in the last quarter, and accounts from the same signup cohorts that are still active. For each group give me median days to first key action, share that ever invited a second user, sessions per week in weeks 2–4, and share that hit any error event. Show me the query text with every result. Then in Amplitude, run the onboarding funnel segmented the same way and tell me which step differs most.”

Ask for the query text every time. A gap between two cohorts stays a hypothesis until someone has read the query that produced it.

Read what the leavers actually said

Behaviour tells you what happened before someone left. The why is usually sitting in text somebody already wrote.

Prompt: “For every account that cancelled last quarter, pull their Front conversations and Pylon issues from the six months before cancellation. Summarise each account in two lines: what they contacted us about most, whether anything went unresolved, and the last thing they said. Then group all of them into themes by underlying problem, with the account count and two verbatim quotes per theme. Do not merge two themes because they sound similar. Keep them separate and say they may be related.”

Read the two halves together. That is where a theme with eleven accounts behind it shows up, the one no behavioural query would ever surface because the product worked exactly as designed.

How do you make it a standing monthly pass?

Prompt: “On the third working day of each month, re-run the Metabase questions, the PostHog cohort comparison, and the theme grouping for the month that just closed. Write it into the ‘Churn’ Notion page under a dated heading, with a section listing what changed since last month and a section headed ‘What we still cannot explain’. Post the headline numbers to #growth with a link, and add an Amplitude chart annotation on the first of the month naming any pricing or onboarding change we shipped.”

Six months in, every chart carries the changes that might explain its shape. It sits with your research and operations routines, reading PostHog and Metabase directly.

Frequently asked questions