I think it is soft to the extent that you need an expert to define precisely the bounds of your problem and the AI solution. However, consider the following sit

I think it is soft to the extent that you need an expert to define precisely the bounds of your problem and the AI solution. However, consider the following situation today: you have millions of images that are fairly related (perhaps on a real estate listing site) and tags. You can now straightforwardly build a product for predicting hashtags and describing the photos without knowing hardly anything about AI. There are plenty of products to help engineers do this already, one of which is Google cloud's ML products. Now imagine that at some point a lot of the AI problems get to a point where they can figure out the correct training procedures automatically. People are working on this with varying amounts of progress but I think the future looks like we'll be able to do enough of this to sell/open source something as well-defined as a SQL database.