Hard real time is a thing in some systems. Also, the current approaches might have 85% accuracy -- if the LLM can deliver 90% accuracy while being "less exact" that's still a win!