Stripe cohort analysis is solid AI territory — the strategic read still needs a human.

Cohort analysis follows a fixed methodology: pull events, bucket by acquisition month, compute churn/ARPU/expansion per cohort. This structure is identical every time it runs, making it highly automatable.

Churn rate, ARPU, and expansion revenue have standard definitions, but 'most at-risk' and 'strongest unit economics' require thresholds and weighting the task doesn't specify. An agent can compute the numbers but needs guidance on what counts as a red flag.

Stripe has a well-documented API and the data is structured, but the agent needs authenticated access, vertical segmentation metadata (which likely lives outside Stripe), and a clear mapping of customers to cohorts. That cross-system join is a real dependency.

A miscalculated churn rate or misattributed cohort could lead to bad retention decisions, but the output is a report — not an action — so errors are catchable before they cause damage. A human review step keeps risk manageable.

Computing metrics is mechanical, but interpreting why a healthcare cohort churns differently than retail, or what 'at-risk' means given current sales pipeline, requires business context an agent lacks. Strategic conclusions need a human owner.