> Business minded folks are already starting to deeply question where the value is relative to the amount of money spent on training

Kinda. In my experience, the bigger issue is the skillset has largely been diffused.

Overtraining on internal corpora has been more than enough to enable automation benefits, and the ecosystem around ML is very robust now - 10 years ago SDKs like Scikit-learn or PyTorch were much less robust than they are now. Implementing commercial grade SVM or <insert_model_here> is fairly straightforward now.

ML models have largely been commodified, and for most usecases, the process of implementing models fairly straightforward internally.

IMO, the real value will be on the infrastructural side of ML - how to simply and enhance deployment, how to manage API security, how to manage multiple concurrent deployments, how to maximize performance, etc.

And I have put my money where my mouth is for this thesis, as it is one that has been validated by every peer of mine as well.