https://laughingmeme.org/2011/07/23/cost-of-false-positives/
Imagine you’ve got a near perfect model for detecting spammers on Twitter. Say, Joe’s perfectly reasonable model of “20+ tweets that matched ‘^@[\w]+ http://’”. Joe is (presumably hyperbolically) claiming 99% accuracy for his model. And for the moment we’ll imagine he is right. Even at 99% accuracy, that means this algorithm is going to be incorrectly flagging roughly 2 million tweets per day as spam that are actually perfectly legitimate.
If you’ve never run a social software site (which Joe of course has, but for the folks who haven’t) let me tell you: these kinds of false positives are expensive.
They’re really expensive. They burn your most precious resources when running a startup: good will, and time.