The part when they said they spent a lot of money to port to C/C++ but couldn't get same efficiency/speed, that's the problem with most porting endeavours: you need people who grok both languages you're porting from and you're porting to, or at least understand the specs and business logic. That COBOL program has probably been optimised so much that the code makes no sense to people without understanding of the intricacies of that language.
My team spent a whole quarter on converting R code into Python, because we wanted to use Tensorflow for machine learning(). When they finally got the thing running, they found out it wasn't performing as fast as R. I thought, that couldn't be possible, they use pretty much the same linear algebra libraries. So I peeked into the code to see what went wrong and found out that they writing it the wrong way: (1) calculated on Pandas dataframe directly instead of extracting the values first when doing matrix calculation, (2) instead of plain ndarray, were using matrix instead, which is slower. Both of which someone without experience in Python wouldn't have known.
() On a hindsight, did it have to be Tensorflow? Besides, there's already an interface for R[1]. Maybe the team decided on Python anyway in case they want to try out the plethora of ML libraries available for that language.
[1] https://tensorflow.rstudio.com/