In many cases people greatly overestimate the effort of labelling data. For the price of just a single engineer man-month you can generally label a lot of data, sufficient for many tasks; especially if you can do bootstrapping (the humans don't label all data, but correct mistakes of a previous insufficiently good automated solution) or some transfer learning from a somewhat similar task or an unsupervised one.

Like, if you need a solution for a some niche of image classification, even a single afternoon of labeling data might be sufficient to adapt an ImageNet classifier to your particular labels and get reasonable accuracy.