Privacy is an interface control.
A project that should stay private shouldn't end up as accidental context for a chat, a feed, or some outside tool.
No fence, no real work.
Get privacy wrong and users leave the same week. The ones who stay hold back: the sensitive project lives in another app, the real numbers never get pasted in, and the AI only sees the harmless stuff. An assistant nobody trusts never gets the material where it could actually help.
What to build.
Show the scope at the moment of asking: a visible list of what the model will read. Make exclusion a first-class control you can reach from the thing itself: hide this project, this folder, this account. Start new AI features with the narrowest scope that works and make widening it deliberate. A narrow scope costs a rerun. A leak is permanent.