I used to build fraud models where we had insane amounts of negative data and precious positive data. Our performance metric was well-reasoned; as you mention,

I used to build fraud models where we had insane amounts of negative data and precious positive data. Our performance metric was well-reasoned; as you mention, we cared about the happiness of our customer, and baked all sorts of useful information into our evaluation metrics. The problem is there is not always a cheap way to encode a complex evaluation metric into a loss function for many machine learning algorithms. In fact most of the really popular algorithms are popular because they work well mathematically - for example, generalized linear models are convex on a simple quadratic loss function, and neural networks have gradients when their loss functions are differentiable. Also, I found optimizing for simple but sensible loss functions often gave pretty good performance. From a practical perspective, most of the gains of understanding the performance metrics we cared about came in the data preparation, evaluation and calibration stages of modeling, not model selection or training.