Something I would like to add is that most models we train are trained upon surrogate metrics.

We can’t train a neural network to optimize profits, nor accuracy, but we can train it to minimize the categorical cross-entropy. We just expect that a model that has a low categorical cross entropy will be an accurate model. And we also expect that an accurate model will be a profitable one.

Knowing the real business problem well help us to find which are the best metrics to watch for. Accuracy might just not be enough. Some problems prefer a high recall or high precision, which are much finer metrics.