I can give one example where the ML/AI toolkit is useful, even if only the largest providers in the world may need to work with real Big Data[tm] volumes. Fraud

I can give one example where the ML/AI toolkit is useful, even if only the largest providers in the world may need to work with real Big Data[tm] volumes. Fraud detection. And you don't need to be a particularly big operator to get sufficiently large & varied data sets. In a way I think it helps that fraud is an ever-evolving, fiercely adversarial domain. New avenues are explored all the time. Occasionally old tricks are revived for a while, because they may work in the margins but become distinguishing features as soon as they see more use. And even if your ML is based on nothing more than a random forest, you can still get surprisingly useful feature combinations out of it. (Or as our data scientists said: individually meaningless features may become a valuable signal when enough of them occur at the same time.)