The idea of an LLM is fundamentally flawed for business intelligence. ChatGPT won't even reproducible follow the same exact prompt the same exact way across multiple runs, which is generally a good thing for avoiding too much repetition and improving its creative and summarizing abilities.

But because it will make stuff up (and has no idea that it's doing so) it's completely useless. For example, I ask it to reformat a transcript of a voice recording, and it adds or changes sentences in unpredictable and highly-inaccurate ways... Often, it's trying to make them more polite and corporate.

I'm sure that these fundamental flaws won't stop it being used for BI, since I've sat through many meetings where entirely wrong (i.e. bugged) or misleading data (small sample size, etc.) are presented as "truth" at work when they wouldn't have passed my sniff test back when I was in high school because the stats are so wrong.