The level of verification should be proportionate to the consequences of error. AI-assisted brainstorming for internal use requires less verification than a regulatory submission. Code deployed to a customer-facing system requires more review than a personal productivity script. The standard is: Would you be comfortable defending every element of this output as accurate and appropriate if asked?
