Modern Data & Analytics architecture and organization: In your experience, what has been a good way forward in terms of centralized / decentralized data management. e.g. establishing data analysts/scientists' team within a specific functional domain rather than centralized team across the board. Have you established any best practices / guidelines or any sort of hybrid organization? Thoughts/comments?

In two past lives, as well as my current one, there is a strong desire to federate the science and analysis practice outside of IT and align to the Lines of Business (LOBs) being supported. In my experience, each unified standalone org had LOB specialists to capitalize on the nuances in each domain (Customer, Sales, Supply Chain, etc.). There were two primary benefits to keeping a unified analysis & science practice: 1) Shared wisdom of the org created cross-domain data science opportunities; and 2) Easier alignment on common practices and the underlying engineering (which was still handled by the technology organization). 

Assuming you are referring to a large org with multiple LOB, here is what I have seen at my previous employers. You would have a global data/AI org and Data scientists looking at overall company strategy, cutting edge Data science advantages, R&D, standardized governance approach etc. Then you would have LOB level Data analysts/scientists who will take globally defined standards, understand LOB business, interact with customers, and will create real LOB level models that goes to production. Some gotcha to watch out for is ensuring that LOB data scientists truly understand state of the business, and appreciate time value of the money. On the other hand Global team may get bit disconnected with ground realities of business and wonder too much into R&D land. For a mid size setup like my current employer, it simply makes sense to pool all talent into one team, but Hybrid is the way to go for large org in my view.