Built In, Not Bolted On: Operationalizing System Safety For Ai