How to ensure only metadata is sent to AI, not data?

[dbcandle]

How do I confirm that only metadata is sent to the AI chatbot/API, not data?

Note: I see that I'm currently rather verbose in my posts here, but I'm hoping I can make AI Assistant work. I had been writing my own tool, but when Anthropic, OpenAI and ChatGPT Copilot all drastically decreased token-counts, I couldn't afford agentic software development.

[otykier]

We require explicit consent before providing certain types of information to the LLM. Data that's being queried from the model requires its own separate consent: https://docs.tabulareditor.com/en/features/ai-assistant.html#consent

Currently, users will have to manually click "Deny" each time this prompt appears, to make sure that the LLM never sees any actual data.

For GA we are considering adding more fine grained control at a policy level; i.e. today, IT admins can disable the AI Assistant using Windows group policies, but we want to make it more granular so admins can permanently deny things like DAX query execution (where the LLM sees the result). Moreover, we have plans for adding more control to the user / semantic model authors, for example such that data in specific columns is always hidden from the LLM (think of it as a kind of Object-Level Security, that applies only to queries made by the LLM).