This is all true if you take a tiny portion of what AI is and does (like generative AI) and try to extrapolate that to all of AI.
AI is a vast field. There are a huge number of NP-hard problems that AI is really really good at.
If you can reasonably define your problem in terms of some metric and your problem space has a lot of interdependencies, there’s a good chance AI is the best and possibly only (realistic) way to address it.
Generative AI has gotten all the hype because it looks cool. It’s seen as a big investment because it’s really expensive. A lot of the practical AI is for things like automated calibration. It’s objectively useful and not that expensive to train.
That’s a great use case. Splunk does something along those lines. Logs are particularly nice because they tend to be so large that individual companies can often have the AI trained to the specifics of their environment.