The bulk of the work is setting up experiments to test how well the AI generalizes to unseen data, debugging stochastic systems, and designing good metrics.

In my experience, this is missing a big part of the work: confirming what the data actually is, sometimes despite what people think it is.