But LLMs does not learn/train like that. They need the same thing repeated a gazillion times in the their training data to actually use it in inference.
So unless this human element get absorbed widely into literature, then I think the scenario you present is not possible, at least in the context of current LLMs.
May be there will be true AI in future, that can read a fact and grasp it brilliance, like a human do, and use it in answering future questions. Then what you said could be possible.
I'm assuming for it to work at the level of filling rolls in businesses like this they will need to generate synthetic data to train on from onsite data, which will be the best practices of you and your competitors and hence include the corpus of what makes you better/lean compared to your competitor and remove differentiators.
So unless this human element get absorbed widely into literature, then I think the scenario you present is not possible, at least in the context of current LLMs.
May be there will be true AI in future, that can read a fact and grasp it brilliance, like a human do, and use it in answering future questions. Then what you said could be possible.