As an example of why this is difficult: when integrating with legacy systems you often have (1) no API (the concept didn't exist / contractor didn't bother to isolate one), (2) no comprehensive documentation of the legacy system's internals (often, not even source), & (3) no human expert available with experience on the legacy system.

Turning any kind of genetic / ML system loose on that? How would you even write test cases, short of backfeeding production data? And expecting the result to pass muster with auditors?