ML modelling where the output affects the DGP If an ML model predicts item-level demand for an e-commerce company and this output is used to re-distribute stock spend, then without understanding the underlying causal structure of the variables which affect item-level demand, then the correlation-based patterns recognized by the ML model would be compromised as the model itself changes the DGP and nature of the distribution of the input data.
As such, is ML recommended in decision-making processes as the model itself is involved in the manipulation of the nature of the environment in which it operates?
Edit: My question was indirectly asking whether causal inference modelling methods that understand the underlying cause-and-effect relationship between variables that affect item-level demand would be more robust and thus more accurate than traditional ML methods in decision-making. Whether such tools exist today and therefore whether it is feasible to replace ML with causal methods (asking pedantically as the notion of causality isn't mathematically rigorous / well defined). For example, how can one quantify the set of features/variables that are canonical in a system that has a complex chain of causality? In theory, this then delves into the realm of philosophy as causality in this sense would imply determinism.
 A: That's a common problem with training ML models using data generated from a set-up where a ML model is already part of what happens. E.g. with search engine data, the results shown on the first page (and even better in first spot) are much more likely to be clicked than those shown further down. Additionally, there can be odd feedback loops in this kind of set-up, so one needs to be careful about such problems.
One way to avoid getting no information on the results that are too far down on the list of search engine results (or in inventory management things that are out of stock), one solution is to occasionally randomly move some results up.  In the case of inventory management that would mean instead of stocking 0 of a product predicted to not sell, occasionally randomly order some to find out what happens. However, that might need some safeguards or that could get embarrassing (e.g. large quantities of perishable Christmas food newly in stock in February, a pile of summer clothes at the start of winter etc.).
Additionally, in some settings you still get useful feedback even if too little stock was ordered, if you obtain the right data. E.g. if you track whether stock-outs occur or even how many people searched for a product (but found it to be out of stock).
In summary, it's definitely a problem to be aware of and that one can try to address in many ways.
A: 
is ML recommended in decision-making processes which inherently changes the nature of the environment?

What is the alternative, given that we have to make decisions? Would HL (Human Learning) work better? Or NL (No Learning)?
I work in retail forecasting, and our forecasts are used in automatic replenishment in supermarkets. If our forecasts are too low, then not enough stock gets sent to the supermarket, so sales will be constrained by deficient stock ("censored" sales), and subsequent forecasts that learn from these censored sales will be biased low, leading again to lower stocks. A vicious circle that could end in a "silent delisting", when the system thinks there is no demand for a product and essentially does not restock any more.
However, the exact same mechanism could happen if there was a human rather than a machine doing the forecasting. Thus, the solution is not to take ML out of the loop, but to understand the feedback loops that tie the ML results back into new ML training data, and mitigate any adverse effects. In my case, that means monitoring stockouts and removing censored sales from the training data.
