if the validation of their output didn’t outweigh the supposed time saved

This is the key. LLMs are (somewhat, there are style and maintenance issues)) useful where things are easy to prove correct, like mathematics (there are theorem validity provers) or programming (does it compile, pass tests).

Most of the rest of the world is not like that and tasks require judgement, so people need to validate that the LLM didn't 'hallucinate' (bullshit anthropomorphic word for 'get it wrong') again.

In subtle domains finding if an error exists can take longer than doing it yourself, with a worse failure rate, especially as skills atrophy. Same thing happened on production lines with the introduction of robots.