I think automatic moderation is one of those golden use cases for LLMs. You can use cheaper inference models, and maybe some clever sampling techniques to limit the token expense.

Thinking out loud, I'd be surprised if this isn't a startup already.