I’ve seen fraud detection used in a SaaS product, and the great thing about a weighted rules approach, is professional services can understand it well enough to adjust it without help from engineering or data science, and they can explain to customers how it produced the results it did in a particular case, and the tradeoffs of adjusting the weights or thresholds, and the customers can understand it too. Whereas, a machine learning model, is much harder to understand and adjust, so issues are much more likely to be escalated back to engineering.

(This isn’t protecting the SaaS vendor against abusive signups, it is a feature of the SaaS product to help its customers detect fraud committed against themselves within the SaaS product’s scope.)

I once did a machine learning project at Intel. The end result was that it was no better than simple statistics; but the statistics were easier to understand and explain.

I realized the machine learning project was a "solution in search of a problem," and left.