I remember an IBM study that showed you get a new bug for every two you fix.

Banking software is extremely regulated and controlled, both by governments and within the industry itself (one small mistake can wipe out entire institutions within a day: https://en.wikipedia.org/wiki/Knight_Capital_Group, banks are extremely risk-averse with technology). So likely they were comfortable using agents to _read_ their code but not to write it.

Have you ever discussed something like this with an executive? I think the most honest response I got on a project I was involved in that required some legitimate legwork was:

'We can get some cheap remote resources for this'

And the funny thing now.. those resources will likely be using tokens anyway..