Most stats / ml majors (or any quantitative field) can take on ML jobs. This isn’t too different from quantitative trading shops would hire any random math / physics phds from harvard / mit / etc - they too make $500k easily.
Most fresh grads study materials with a huge assumption: the data you train & build model on is static. In real world, input data changes all the time, and the model you trained yesterday now becomes completely useless today. Sometimes the brittleness of the data comes not in time dimension, but follows a random process. So that adds another modeling layer on top of the first problem you were trying to solve. So techniques you employ is different in every deployment.
“New algorithms” @hanera refers to typically assume that there is a well defined problem - e.g. object recognition in images. Once a problem has been posed (by a well known researcher, or from slew of related applications), the ML community solves those problem thru a series of publications, and they become available in open source, so anyone can run those libraries.
But there is a huge long tail of problems that are seemingly unrelated, and therefore still needs lots of engineers with experience building similar systems. This reminds me - a long time ago, @manch asked if web search is solved, every other companies can just turn on the programs and “bam you go”.
That won’t work, because searching for snpachat videos vs maps vs web documents are all different problems. 