While in general I agree with most of what's said, some of these feel like "Horseless Carriages".

Like for example

>It doesn't know how your company counts revenue, or which of several similar-looking tables is the real one (it is revenue_cpx_new_revised_v2_use_this_one )

This assumption mostly only holds for non-clean data ( which is mostly a result of humans being lazy and backlog being clogged up ). As AI starts building more of the pipelines upstream - this becomes a non-issue.

Even for existing pipelines, I'd expect fixes to move upstream: when an agent encounters ambiguity at query time, the useful output isn't an edit of the context,but a PR that fixes it.