“my experience so far has been, you have to be an expert to spot the AI mistakes”
Your post reminds me of the early days of spreadsheet software.
I had a job (long time ago) where all the number crunching was done of the green accounting paper—by hand—using calculators.
I told my old-timer boss I could turn all of it into (then Lotus 1-2-3) spreadsheets and radically improve speed and accuracy including the ability to quickly generate “what if” scenarios.
It took about a year.
He manually checked every single spreadsheet calculation for a full year.
Then one day he started using the computer. He had not found even one single Lotus 1-2-3 spreadsheet error in more than six months.
So, you could argue that Lotus 1-2-3 was a waste of time during the transition time.
However, those spreadsheets (later converted to Excel) are probably still used today, decades later (I am retired so cannot confirm that.)
This is the ugly transition period for AI. When it is over it will easily pay for itself.
I don't disagree at all, but he cited legal and medical, and having tested GPT-4 on the former, not a chance in hell is it robust enough for a professional services contract.