I was loosely involved with a project at a very large company where they were using Watson (all on-premise, no cloud) to classify customer feedback (verbatims)

I was loosely involved with a project at a very large company where they were using Watson (all on-premise, no cloud) to classify customer feedback (verbatims) into an existing taxonomy. This was actual customer feedback that came through service channels, not social media chatter. The taxonomy basically determined the escalation process within the company. As an example, think about issues with a cell phone. "Sometimes my phone volume goes down automatically" isn't as critical as "Sometimes my phone gets really really hot in my pocket." The Good: They felt Watson was doing a good job classifying the customer feedback. Overall, they felt they could rely on Watson and could defend using Watson instead of human raters in a legal situation. The Bad: The amount of time/resource investment required to get it up and running had been way beyond what they initially expected. They also felt various parts of the system could have been more open (e.g. Watson's back end database) to inter-operate with downstream systems. They also were exploring if other solutions (e.g. Data Scientists + Python) could do this and had started to see positive results.