That would be on the companies (or their auditors), not Microsoft, in general. Clearly, no-one should ever _use_ this, should it ever make it out of research-land, but there's not that much obvious risk to _making_ it as long as they're honest about the risks.
If it could spit out the analysis as spreadsheets that used standard formulas and only use the LLM to generate the formulas, it could be verified. Errors would slip through, but no worse than people applying the wrong formula based on a quick internet search that calculates a close but incorrect answer.
That's not the marketing message at the moment. I see ads for AI (LLM) powered services and they all say the same thing, "Stressed? Not enough time? Let AI do it faster so you can do more." AI is sold as a tool that can do things faster than a human can and since LLMs do not provide reference information, there is no telling where they got the data from and no way to verify it.
Validating that a complex spreadsheet is correct is notoriously extremely difficult at the best of times; unfortunately they are about the closest thing to a write-only language in common use, and you really have to front load a lot more care than you do in conventional modern languages. The usual safeguards of testing and code review are essentially absent.
I’m sceptical that anyone really _should_ be using generative AI for anything where correctness matters at all, but spreadsheets in particular seem close to a worst-case scenario.
There may not be much risk from a legal culpability perspective if they make the appropriate disclosures somewhere in the depths of a EULA, but even so it is a failure of professional ethics to build tools which are "dangerous at any speed" and inflict them upon the world.
Oh, I don’t disagree, and I don’t think burying it in the EULA would necessarily be sufficient (especially in Europe, where the courts and regulators have tended to take a dim view of “but we told you, in three-point type on page 473 in the middle of the trademark acknowledgements”). But ultimately the blame for using known-unreliable tools is largely on the user.
