A couple months ago I shared an article in this newsletter by the venture capital firm Andreessen Horowitz on The New Business of AI (and How It's Different From Traditional Software). In this context traditional software is SaaS software. In this article they explained why they are seeing that very often AI based software product businesses have inherently lower gross margins than typical SaaS software businesses.

One of the underlying reasons for the lower gross margins is what is often referred to as the "long tail" in model data distributions and the human and computational work involved in addressing this dispersed part of the data. In this second article, Taming the Tail: Adventures in Improving AI Economics they give some good examples of how to try and avoid this when practical which I found helpful. They also give some practical advice of how to minimize the long tail if in fact there is no way to avoid the "long tail".

Both of these articles highlight the issue that if you can possibly avoid it, don't use AI models in your product that require training against individual customer data, use a universal model if you can.