What it does
Churn is usually handled as a Customer Success problem, which means the product root cause never gets named. This agent goes after that instead: given a cohort of churned accounts with their per-feature usage over the sixty days before cancellation, it finds which feature's usage dropped first, how widely that pattern repeats across the cohort, and how much warning it gives. Each finding carries the numbers a product team can act on — the share of churned accounts showing the decline, the average decline percentage, and the median lead time in days before the drop crossed half of baseline. A result reads like "usage fell about 74% on average, 75% of churned accounts show it, typically first dropping below half-baseline around 34 days before churn". That lead time is the useful part: it tells you how long a save window you actually have. Alongside each finding it recommends a product-side intervention rather than a CS play. The honesty control is unusually strict and worth noting. Every analysis tool returns its data plus a `caveat` field carrying a correlation-is-not-causation note, and the agent recommends validating with an experiment before acting. That warning is not a footer on the report — it is attached to each individual tool's return value, so it cannot be lost when a client renders only one slice. Each analyzer seam is callable alone. `critical_features` returns just the worst offenders. `early_warning_signals` returns the lead times. `correlation_confidence` returns the strength of the pattern, explicitly labelled as a correlation signal rather than proof. `cohort_summary` gives the rollup. `run_full` does everything. The decline analysis is fully deterministic; a language model only writes the root cause and the intervention, with a template fallback when none is configured. No connectors and no credentials — it never calls BigQuery, GA4 or a CRM. You pass the cohort in, and anything you leave out is reported empty with a warning rather than filled in on your behalf. Built for product managers, customer success and growth teams.
Example prompts
- Which feature do accounts stop using before they cancel?
- How many days of warning do we get before churn?
- Show me only the critical churn signals in this cohort
