It seems like LLMs would be great for this: "Is this email trying to sell me something?" "Is it from an acquintance?" "I have these mailboxes, which is the most appropriate?" with some additional metadata.

Well, as long as it doesn't fall into prompt injections.

Sadly I don't know of existing solutions to do this.

> It seems like LLMs would be great for this: "Is this email trying to sell me something?" "Is it from an acquintance?" "I have these mailboxes, which is the most appropriate?" with some additional metadata.

Why LLM's, unless it's the only hammer you have? For instance, "is it from an acquaintance?" seems like a case for a classic mail rule. It's not like an LLM can read your mind to find out who your acquaintances are, and mining your correspondence to figure it out doesn't pass the smell test. If you chat with a friend about "Steve," how would it figure out who that actually is? Should it let all mail from every self-identified Steve in?

The only way to do better than a mail rule is privacy-invading metadata collection a-la Facebook (who slurped the address books of you and your friends, and tracks all your communications that it can, so it knows who your friends are who who your friend's friends are).

You skipped over the one that is perfect for LLMs, determining the purpose of intent in the message.

Would you be willing to pay for a gmail plugin that takes a stab at this? something like Mailman [1] but with an LLM layer for detecting AI slop?

[1]: https://mailmanhq.com/