Pareto principal. We've seen the 80% but that last 20% is going to be really tough (and expensive). GPT5 illustrates this - it wasn't really better than GPT4o and in some ways worse.
In many situations, the work is not indivisible. If AI can handle 80% of the work, then a company can let AI handle that 80%, fire 80% of their people, and consolidate the remaining 20% still-human-work with whoever is left.
That all being said I strongly disagree that GPT-5 isn't categorically better, I just think it's less obvious because we're starting to hit human cognitive ability to even assess that limits.