Good idea, I have not considered on prem yet. But I can see the need in cases where data is sensitive.

That's what I was thinking. For sensitive information or information that's under legal restriction, how are we going to train the models for on-prem AI?

I guess we could train models on similar corpus scraped from the internet - mostly old General Aviation manuals, I guess, but a) that's not enough corpus, and b) those GA docs are so old I'm afraid my AI will start chain smoking and casually using the word "broad".