Well, it certainly would depend on the use case and the quality would be influenced by both the scoring threshold and extent/quality of training materials. Using large data sets, with a particularly difficult lowest sufficient score, could produce a small quantity of highly relevant and substantive content. If the purpose is to discover a handful of articles in a specific subcategory (ex. theoretical physics) that a user could confidently share to social media - then there is a significant utility to the service - not just a business case.
What would this service have to offer for you to actually pay for it? Of course you could say you wouldn't pay for it, simply curious...