I wrote this blog post after having the same conversation a few too many times, so sharing it here in case it's useful.

The quick test I've started using: can you pull one member's renewal status, event history, and payment record in under a minute? Usually not, and that's the whole story. The model isn't the issue. The data's sitting in a renewals spreadsheet, a payment processor, an event tool, and a CRM that don't talk to each other. Feed that to an AI, and it can't tell a lapsed member from a renewed one.

The unglamorous stuff has to come first: de-duped records, renewals and payments pointing at the same member, engagement captured as data instead of living in someone's head.

If you want a starting point that's easy to measure, renewal risk flagging is usually it. Small data footprint, and staff can tell you immediately whether the output looks right.