One would hope that IBM’s selling a product that has a higher success rate than a coinflip

Again, my point really doesn't have anything to do with specific percentages. The point is that if some percentage of it is broken you aren't going to know exactly which parts. Sure, some problems might be obvious but some might be very rare edge cases.

If 99% of my program works, the remaining 1% might be enough to not only make the program useless but actively harmful.

Evaluating which parts are broken is also not easy. I mean, if there was already someone who understood the whole system intimately and was an expert then you wouldn't really need to rely on AI to port it.

Anyway, I'm not saying it's impossible, or necessary not going to be worth it. Just that it is not an easy thing to make successful as an overall benefit. Also, issues like "some 1 in 100,000 edge case didn't get handle successfully" are very hard to quantify since you don't really know about those problems in advance, they aren't apparent, the effects can be subtle and occur much later.

Kind of like burning petroleum. Free energy, sounds great! Just as long as you don't count all side effects of extracting, refining and burning it.