Over a decade ago, I worked on a project that involved categorizing and then replying to e-mails using – as the sales person touted – “artificial intelligence based on hand-crafted knowledge bases.” (Not a verbatim quote; something along these lines.)
The e-mails to be categorized were, of all things, incoming support requests by end users, and the results were, of course, crap. This was partly because the underlying “AI” was just a bunch of dumbed-down regular expressions and rules that tried to match the writer’s intent, but mostly because of a completely flawed approach to the whole idea.
“Software is good at pattern matching,” someone in the product department must’ve thought, “so we are going to write software that matches patterns in incoming e-mails.”
So far, so good. The problem, however, is that people are not good at all in producing consistent patterns. — Trying to match patterns that are not there in the first place? Good luck with that.