Small anecdote that might help. The real cost driver was reprocessing, not inference. Push data contracts upstream so APIs don’t drift.