Would 99% accuracy in ML data labeling for text be a big deal?

Hi HN! I'm a qualitative researcher pivoting to machine learning data labeling because it appears to me, as an uninformed outsider, that ML researchers are happy with 80%-90% accuracy. I would consider that bananas because I'd never give an SBS client data labeled at less than 99% accuracy.

Is this understanding of the current ML data labeling landscape correct? If so, how is the status quo so terrible? If not, what am I missing?

For reference, our beta: https://www.ameliormate.com/data-labeling