You see people trying to shoehorn ML into many such system though, and there is money in it, which is why people are chasing it to have it on their CVs.

Like the noSQL hype of a some years ago it'll settle down and people will gravitate back more towards the right tool for the job (which will sometimes be ML based, but often not, just as "noSQL" is sometimes the right tool or right enough). ML will survive where it is the best tool for the job, or at least where it can be genuinely useful and not significantly sub-optimal.

> our problems are CRUD at modest scale.

I see some of our client base looking into ML and "Big Data", and I despair a little because they often fail badly at getting "little data" correct. It is actually part of the sales pitch for ML: let the AI filter out the crap in your inputs and give you something approximating a decent answer as output. They'd be much better served working on fixing the data sources or using more traditional cleansing methods, but that seems like harder work compared to the new magic some consultant is extolling. ML isn't a magic bullet, but it is currently being sold as one.