How would you design retention and deletion for AI application data?

Retention policy is encoded in lifecycle metadata rather than inferred from a table name. Deletion is idempotent, observable, and tested across partial failure, provider outage, backup expiry, and tenant closure. Evidence records what was requested, which stores completed deletion, exceptions, and final verification without retaining the deleted payload itself. Product teams reassess retention when they enable new provider state, human review, analytics, or model-improvement sharing because those can create a new purpose or deletion limitation.