Very insightful post. I would add the following points.
- Prediction value is not always positive. In many cases you lose value for low accuracy predictions. These use cases are, in fact, perfect for building defensible moats because until one can get the accuracy to be high enough generate positive value, the model is not very useful.
- It is often useful to segment the value-accuracy plot. For example, in an ecommerce recommender system, the value-accuracy plot for high value buyers will be quite different from that of chronic browsers.