Now, in the medium or long term, I expect there to be AIs that will be able to do this sort of thing just fine. As I like to say I expect future AIs will not "be" LLMs but merely use LLMs as one of their component parts, and the design as a whole will in fact be able to accurately and reliably relay corporate policies as a result. But the stock market is not currently priced based on "AIs will be pretty awesome in 2029", they're priced on "AIs are going to be pretty awesome in July".
LLMs are a huge step forward, but they really aren't suitable for a lot of uses people are trying to put them to in the near term. They don't really "know" things, they're really, really good at guessing them. Now, I don't mean this in the somewhat tedious "what is knowing anyhow" sense, I mean that they really don't have any sort of "facts" in them, just really, really good language skills. I fully expect that people are working on this and the problem will be solved in some manner and we will be able to say that there is an AI design that "knows" things. For instance, see this: https://deepmind.google/discover/blog/alphageometry-an-olymp... That's in the direction of what I'm talking about; this system does not just babble things that "look" or "sound" like geometry proofs, it "knows" it is doing geometry proofs. This is not quite ready to be fed a corporate policy document, but it is in that direction. But that's got some work to be done yet.
(And again, I'm really not interested in another rehash of what "knows" really means. In this specific case I'm speaking of the vector from "a language model" and "a language model + something else like a symbolic engine" as described in that post, where I'm simply defining the latter as "knowing" more about geometry than the former.)