A few years ago I participated in a kind of AI method using near infrared spectroscopy to monitor a process. It was a success but required constant revalidation with "known value" substance using up to a few dozen algorithms at user-selected wavelength regions. There is no reason this process of changing weights and methods couldn't have been automated. That my friend is a real world example involving experience. But it could have been programmed into a learning model and validated with every inplementation