Based on my own experience working on AI-based solutions for one of the better-covered job families according to this map, I think 90% might be overstating the case. It's true that we've got a lot of published benchmarks, but they tend to cover the lowest value add tasks.

It's really easy to fall into the McNamara fallacy when rolling out these kinds of solutions. Focusing on the low value stuff is attractive. It tends to be easier to do, and also easier to quantify outcomes. But it also tends to direct attention away from important but difficult to quantify considerations, which may mean you don't notice negative externalities. So we often found ourselves in a verschilmmbessern situation where automating what we could automate made things worse overall.