I hear you. But in this case, it seems like the author was mostly referencing academic uses of LLMs for either writing assignments or reviewing (academic) papers. Enterprise communications have their own carefully set requirements, but often they aren't meant to be instructive to the person writing them (assignments) or enhancing an existing corpus of knowledge (academic papers, optimistically).
There's too much information in the World for it to matter, I think is the underlying reason.
As an example, most enterprise communication nears the levels of noise in its content.
So, why not let a machine generate this noise, instead?