My point is that even humans make fewer mistakes when they're reading an established form structure with pre-defined fields over long-form text. Hence you'd expect a machine to process them better as well, even if it could read like a human.