My org is paying api prices for Claude because it’s convinced that we cannot under any circumstances let them train the model on our source code, even though they could put everyone on Claude ultra subs. How do I explain to them that the model can write a better app from scratch