You're implicitly assuming that manual reviews would have a 1% or better error rate. I sincerely doubt that's realistic based on what I've seen of assembly line style processes.