Because often no one remembers what the initial process was supposed to do in the first place, and if the original objective was a bad one, automating it makes it far worse. Example: a publication system vendor left a QA procedure that said "put a space character in the front of all these attribute values when they start with alpha". So the team methodically and wrist-slashingly hand-coded thousands of XML files so that they all had leading whitespace in their attributes. When I came on board, I didn't automate that process out of the gate, because leading whitespace in attributes is f@#ng dumb. If I did, I would have caused a lot more problems downstream - so this was a simple process that needed to really die, or at least get explained. Which, turns out, it did have an explanation: the vendor didn't understand how the FOSI print formatter worked, but they had this workaround that seemed to work, so they handed it down. The real problem was being caused by a mishandling of the entire element in the data structure, which was a much bigger deal.
