I would give the LLM statistical tools that work to find general anomalies, distributions etc. from the spreadsheet. Then it's about an LLM interpreting (or not) those results in natural language.

Of course this is a can of worms for a product because we can still not guarantee accuracy.

That makes sense to me in terms of tech, but I don't get how the economics will work. If one company has an agent that's a great statistical tool, another is great for making charts, another is great for grammar and spelling, etc., I'm guessing it'll cost a bazillion dollars in subscriptions.