Lot of words, not a singe real use case. Do they have a working prototype? Is the AI better than humans?

The idea of A/B testing is that humans are not good judging which version is better, so the only solution is to run an experiment, so in many cases you will get better results with an A/B test.

And there are also variations where the split is not constant, like the mutiarmed bandit. If A/B test is better or worse than multiarmed bandit is also a good flamewar topic.