To be honest, before the modern machine learning approach, this was known as a decision tree and was thought to be a valid way to approach "artificial intelligence". Lots of "AI" hype in the 80s was based around "Expert systems" and "Decision trees".
And there are even modern tree based approaches, that beat some of the modern artificial neural network approaches! It's not like it has become an absolutely unusable class of algorithms.
People seem to think that the presence of neural nets and deep learning means that most of types of models are practically superseded whereas in my experience if some non-deep-learning model even gets you mostly there, then the efficiency and explain-ability wins make it worth it.
At my previous job I had a ML-based service that used a basic random-forest model instead of a neural net because it was faster to train and operate, not to mention easier to maintain and had equivalent accuracy with little to no effort required on my part. It was a solid little service.
And you can even do both. Decision tree to get to a smaller problem space, then NN on the output. You end up with a bunch of neural nets, each of which performs better than a single monolithic net due to solving a simpler problem.
Decision trees/behaviour trees are still the most widely used way to build video game AI. Of course video game AI just has to appear smart, really you're solving a different problem a lot of the time.
