How often ML fails?

I am just wondering how often ML fails in real world with real data. Is there domain where ML has no use because all algorithms have poor performance? For example there are some academic papers for using ML to malware detection with good results. But I know guy who works in AV company and he told me that they are not using ML algorithms for detection because it gives too many false positives. Do you know other domains where ML fails? It will be good to know what are nowadays limits of ML.