I wonder if this also applies to computer vision? There are certainly problem spaces where heuristics are well established, but many approaches around object detection/segmentation seem much easier/robust to implement with machine learning.
I myself tried several time to “not use ML” for some easy computer vision tasks where traditional CV methods are supposed to work. Well I always end up in situations where they don’t work well without fine parameter tuning, and tuning the parameter for a situation breaks the model in other situations, so you start adding layers of complexity to automatically tune the parameters, but the parameter tuning system also has its own parameters... While a simple neural net is trained easily and is much more robust, saving a lot of time and complexity.
Another proof of that is that CV products only started meaningfully entering the market after ML became applicable to CV (after 2015 for complex tasks, or earlier for simpler stuff like MNIST).
That being said, you are right to say that ML changed the game entirely for CV in industry at large.
It revolutionized the field in a little over a decade and brought forth new frontiers.
I do believe that CV is an area ML will excel well into the next century.
Perhaps, we will find a way to chain together ML systems dynamically, overseen by a procedural system that makes real time decisions in understanding its input.