no confidential client information is to be sent to any cloud AI service

If the code has been crafted correctly, none should be.

Client info should be 100% data, and remain 100% in production. If you have any dev-side data to test and debug with, it should be 100% procedurally-generated pseudo-data that only imitates client data in terms of gross patterns, and not specific points of data.

Like customer names. Those names should be broadly representative of what is found in the regions the software operates in, in order to demonstrate acceptable data input validation and effective storage compatibility with those names. For example, you don’t want to be storing European names in a DB field that is not UTF-8 compliant, or validating those names by restricting characters to US-ASCII. But any one generated name that just so happens to match an IRL name should be a statistical occurance, and any extended data attached to that entry should (generally) not match any other IRL data for any IRL occurance of that name..

Multiple companies have that approach, I'm not denying there ARE companies doing things right (or at least better).

I work in a company using multiple agents. One of them is supposed to be off Internet, and even with that one we're not to share sensitive IP.

But Proton is not an isolated case. I bet some execs have invested and promised so much they need to report some good numbers on usage, and you can't make up productivity improvement if there are people left getting as much done without AI.

For the LLMs killer feature, software development, you don't really need confidential information 99% of the time.

It continues to annoy me how both haters and glazers will over-simplify the topic by calling everything AI. There's a giant difference in usefulness, ethics, and quality between LLMs that produce code guided by an experienced developer, and "art" slopped out for greed by some random schmo.