I think a big reason so many internal AI tools break is that they round everything down to the lowest common denominator. To make the tool safe and shareable for everyone, you strip it back and treat things as flat and discrete. And in doing that, you lose so much of the context that actually made the information valuable in the first place.
The nuance matters. The same question deserves a different answer depending on who’s asking and what context they’re allowed to hold. The fix isn’t to flatten, it’s to gate. Share information, not raw data. A good tool, like a good manager, gives you what’s useful for your level of context without dumping everything or rounding everyone to the same bland baseline.
When a tool feels generic and useless, is it actually just protecting you from the context that would’ve made it sharp?