In my case, it helps improve my ADHD-affected executive function. It lowers the barrier to entry to ingesting information into a system I know I can extract it from later. It gives me peace of mind to know I can ramble semi-coherently at the speed of conversation and know that the salient points are being captured.

Local models will continue to improve, if your concern is privacy -- already they do a decent enough job at interacting with a well-schema'd PKM