The problem is that there is no money in that middle layer you speak of, so its hard to justify putting in all that work to keep it updated to an every changing landscape of models and user expectations.
I do know that Anthropic offers a "Zero Data Retention" contract for their enterprise api that could work. Thats still kind of a "trust me bro" but at least its a legally binding "trust me bro". I can't image summarizing / categorizing documents on a file server is beyond the scope of the latest claude models.
