Pretty much this. However I think that you can actually get pretty far with cleaning a surprisingly small subset of data and just quickly tossing together a pro

Pretty much this. However I think that you can actually get pretty far with cleaning a surprisingly small subset of data and just quickly tossing together a prototype. Talking to people helps a lot. "Ok so you have this and that data...how helpful would it be if you could predict the price of item X that you buy 10% better"...well ok it would be worth roughly Y$/year. Let's build a small prototype and see how well we can actually predict it. At this point you quite often find out that "oh we more or less predict it via one person who makes the buying decisions by experience" and it might take some convincing to actually compare the two. I'd summarize that the "big three" of ML/DL are all nontechnical 1) Finding the data/cleaning the data 2) Is there actually some reference you can benchmark against 3) It's not really all that hard and you might need less data than you think to show enough business value (broadly speaking a pretrained net + some usecase layers is good enough to get started for most business cases) #3 is interesting. I tend to be quite honest but I've seen a lot of mumbo-jumbo hand waving "oh it's very tough" which of course makes sense if you want to sell shiny new stuff. Reminds me a lot of ERP sales where I've seen basically SQL+a little webfrontend pitched as rocket science :D