In some of these cases, as you say, the KPI was higher, e.g. in classification tasks (where previously a clerk was supposed to label some cases). An LLM produced decent results most of the time - and didn't get tired. But when I looked deeper, it turned out the folks who decided to use LLMs could have used classical ML learning methods as well with similar efficiency.

I believe we arrived to the point where people just got lazy and just use LLM for anything.