The IRR of AI is the worst of any computer industry wave.
Depends entirely on how it is applied. For example, Spotify, Netflix, Apple Music, Tidal, etc. are simply not possible without machine learning. Drug discovery has benefited from ML already as have other uses in medical. Fraud detection is an example of ML. Face detection and naming of the faces in photographs on my iPhone are only possible with ML. Let’s not forget driver-less cars. And I used the term ML for a reason, rather than saying AI. AI is an over-used term.
ML is a more appropriate term in my opinion. It refers to a fundamental difference in how the code works. In “programming” you outline the steps to take to solve some problem. But some problems are impossible to solve in this manner. Instead, ML programs learn from experience, rather than being told. This is exactly how humans learn to walk, for example. We are not “told” by our parents what to do, we learn what to do by experience. This is what training data is all about.
Consider a programmer writing an algorithm to sort data. Something computers do a lot. Something sophomore Computer Science students do in class work. If my sort algorithm is slow I have to re-write the code to find a way to sort faster. An ML version of the same thing would simply sort lots and lots of data sets and find the fastest way to sort data. And have the chance to continuously improve on its own. But this is not necessary, or applicable, to all problems that we need computers to work on of course.
The more extreme example is, of course, the ML programs that learned to play (and master) the games of Chess and Go by being told only the rules and the definition of win/loss. As well as the ML programs that learned to drive cars.