Most importantly, what they really do, 80–90% of their time, is data cleaning, joining, formatting, augmenting, then cleaning again.
The data science ecosystem: industrial edition
Balázs Kégl
322

This, conversely, is what academically trained, theoretically or domain-wise informed (depending on their training) specialists are very bad at. Speaking from firsthand experience — although I am probably unusual bad programmer to begin with.