I’m wondering what happened to the people in meeting rooms who suggested that this could be automated if it needed no deep understanding of the COBOL language or the banking domain..?

EDIT (for clarity): that’s what I would say, and I’m pretty sure many people smarter than I had said it - so I’m just curious what objections made business push forward with the “relatively unskilled” human labour intensive practices instead, because that’s the history we have now.

There were tools available, but it was increasing the risk (in a risk-averse atmosphere) to trust them to catch all cases. Try explaining to your superiors why you were given the task of fixing the issue and instead of completely fixing it you saved some money. Probably unacceptable.

Also, much of the remediation could not be automated anyway (for example you may not be able to assume the century is '19' for a given date field which can be in the future. You may need to do a comparison to a cutoff date, or you may only have to enlarge the field knowing that the value is supplied from elsewhere which will itself get fixed), so it fell back to people looking at the code, understanding it just enough, making the required changes, and planning/executing the testing.

I'm not a COBOL expert, but I suspect you misunderstand both the difficulty of teaching a machine to understand the problem well enough to automate fixes for it and the available computing resources at the time. This was still the era of Windows 95, the classic Mac, and expensive workstations.