[pifm_guy]

Sketch of a design to solve this:

Company can upload some prompts to OpenAI, and be given 'prompt tokens'.

Then companies client side app can run a query with '<prompt_token>[user data]<other_prompt_token>'. They may have a delegated API key which has limits applied - for example, may only use this model, must always start with this prompt.

That really reduces the privacy worries of using all these third party companies.