An LLM can only be derivative. It can come up with "unique" patterns, but only based off it's training data. For something like your use case, it would have very little data to work on given that you claim the documentation is poor. Your best case would be to use a high reasoning model and feed the documentation into it's context before asking anything - but it will only give you answers based on that input. If the documentation is wrong or outdated, you'll get answers based on wrong and outdated data.

As far as "who could get left behind?", I feel that is more people who's job can be impacted by LLMs - programming is a big one, and the reason is largely in my first paragraph. The amount of "training data" on the internet for coding is absolutely massive and incredibly well organized because programmers are nothing if not incredibly pedantic. The result is that the task of writing code can be largely offloaded and with recent models, the quality of the code being produced is extremely good, especially in tasks that are well solved/documented.