If you mean deep learning type systems, you need to train them on vast training sets of similar data so they can learn very general patterns, similarities and connections. Like giving them 10,000 pictures of dogs, so they can learn to recognise pictures of dogs.
These banking systems are all massively complex one off systems, often with very different sets of functions or even definitions of the problem, and the implementation details are different in almost every way possible. They’re the exact opposite of the sorts of problems modern trained AI systems are good at dealing with.
