The argument is you can design a model, check that it runs, then submit the job to train / evaluate the model on real data. The data scientist doesn't necessarily need to touch the real data ever (especially if ML is used).