I think the one-sentence version of this thesis is much more interesting and true than the specific build out:
"AI" will very obviously continue to make inroads into traditional human knowledge-work and adjacent domains, automating roles and (for some metrics) outperforming humans.
The lemmas are playing out in too many domains to track any longer already: there will be blowback largely defined by intentional and unintentional confusion and misplaced identification of root problems; and by chaos where historic structures (legal ethical and otherwise) which were reasonable when the only agents exhibiting agency were human, now have to contend with a spectrum of collaborative entities mixing individuals, communities, and "AI" of all kinds.
"Middle management" is a misnomer, but the premise is reasonable, that which can be quantified will be optimized.
As noted ITT what and how quantification occurs and whether that tracks "reality" (or traditional fictions at least) remains to be seen.
My own take on that is that what we will mostly see is evolutionary pressure weeding out unviable failures, accommodating for a while at least midling failures, and a slow quiet creep towards deeper and more pervasively "correct" solutions,
all trending inexorably towards even less and more existential challenges at both the philosophical and societal level, and the basic one defined by the paperclip peril of the Alignment Problem.
I am unsurprised that certain domains (like effective definition of and execution of a business plan through an organization made of non-fungible humans) will prove exceedingly hard to fully capture.
But I am now regularly surprised by ice-shelf-break level dramatic gains in domains that seemed until recently equally unassailable.
The ingenue in this domain for me is Whisper. It's just... solved that problem for good, it seems.