Unlike Excel, AI truly can deliver a competitive advantage when used judiciously. In the right situation, it can deliver that small catalytic boost that lets you see just a little farther than before. And incremental advancement is the basis for almost all scientific or commercial advancement.

That said, deep CNNs and LSTMs have delivered only a fairly narrow set of catalytic advancements to date, even fewer of which have changed the status quo in big businesses, where R&D ROI really adds up. Accurately identifying the next use of AI that will meaningfully move the needle is beyond hard.

But hasn't that always been true of all forms of innovation? That's why so many are keen on AI today; as hard as it is to wield, it still seems to be the sharpest arrow in our quiver.

I agree with you. I transited from computing to drug discovery. A lot of projects attempt to harness the power of ML but there still seem to be a lot of misunderstanding. A lot of the time the valuable ML methods are supporting / assisting hypothesis generation rather than 'doing it all'. The domain experts will judge the value of ML methods by their usual merits in the drug discovery world. If the more sophisticated methods can predict hit compounds no better than the traditional standard deviation based cut off, or if the models can not explain key insight of any indication of mechanism of action, then it may not be a good method after all. But meanwhile if it opens up a new way (e.g a new phenotype based screening) then it is worth pursuing.