This feels like the issue with reproducing research in the softer sciences.

From my understanding businesses operate by optimizing over some gradient similar to gradient decent. Correlation or causation is really not the objective (but can be helpful to optimize over).