Using machine learning in practical projects

How is the learning process practically used in real world cases? My specific doubt is that it all starts from big data or at least thousands or millions of records with an already defined output.But in a brand new project I wonder if this is a realistic starting point. Let’s say that I have a source of images and I want to recognize the physical bodies in them or whatever (assuming that neither I have an already trained model nor all the needed expected output available but only the input data, of course). Should I really train my new model from scratch, I would have to do the work (of producing the output) - manually! - on all the training and validation set before… It looks like practically unfeasible in a lot of common project scopes, as far as I can see. Usually humans "learn or guess a new algorithm or function"from very few examples. (just to do a bit off topic comparison between humans and machines) Back to the main question. Any clue about real business projects where you are actually building, i.e. training your own model?
Thanks