Probably because 99% accuracy isn't good enough.
For humans, much can be deciphered through context. That's much more complex to do with OCR.
If I give you 1000 handwritten numbers, do you think you'll make less than 10 mistakes?
Hide some graduate students. Tell each to transcribe, and to mark the difficult spots. Give each page to two students. Next, have someone else process the page pairs and resolve conflicts and the marked trouble spots.
As long as one of the students notices that a particular spot is difficult to read, the error is discovered and can be handled by someone who isn't numb from transcribing pages of simple numbers.
As you start labeling data everything has to reach a consensus. I'm not sure exactly how it works, but it does have multiple people verify each piece of data.
Maybe they should split in lines each figure, make a coordinates map of the entire figure or so, make a copy and applying a bulk search with a machine for crossing the map and annotate as many undoubtely identifiable numbers as possible. Then paint it in a different color easy to filter or hide it and remember its position.
And then add humans to focuse only in the remaining dificult cases and outliers armed with a reference sample chart. This way an human would need to focus its eyes in 20 characters/image (instead 60 or 100). They would accumulate more completed figures faster and obtain a bigger sense of reward. If the human can't recognize the number, could put a ? in the chain and move on.
Thus the machine would annotate for example 1_3, 4_67_, _8 and the human would write: 5?01 for the same line.
Just an idea, don't know if designing that is easy-peasy or really defiant but the number of possible values is limited in any case, so should be possible to train a machine to recognize some single characters or even entire numbers. Specially if written for the same people.