Math and Information Science failures
If you teach a machine to analyze the behavior of millions of people, it just might learn to predict the behavior of millions of people. At the individual level, that means your machine maybe has a one-in-a-million chance of better prediction than manually researched market segment data. It's useless. Unlike manual research, big data doesn't even tell you why results are poor. The same goes for analyzing millions of low quality DNA tests for diseases, whether or not somebody is beautiful, or whatever the VC scam du jour is.
The kind "big data" needed for accurate prediction per individual is generally not legal or ethical to collect.