Centralize everything and the work stops fitting the business. Push it all into the units and you lose leverage and consistency. The hybrid wins for a reason worth understanding.
Every company standing up an AI and data capability faces the same structural choice early, and gets asked to decide it before it has enough information. Do you keep the talent in one central team, or push it out into the business units? The answer that holds up almost everywhere is: both, on purpose.
A centralized center of excellence keeps all the specialists in one team. You get consistency and leverage, and you get clever work that the business never adopts, because the people building it are too far from the floor to know what would get used.
A federated model pushes people out into the units. You get relevance and adoption, and you get six teams solving the same problem six ways, with no shared platform and no standards.
The hub-and-spoke hybrid does both at once, and for any company large enough to have distinct business units, it almost always wins. A central core owns the platform, the standards, the governance, and the scarce specialist talent. Translators and analysts embed in the units and sites where the domain knowledge lives. The center gives you consistency and leverage; the embedded people give you relevance and adoption. Neither works alone.
The scarce specialists sit at the hub so their work carries further. The domain experts sit in the business so the work is worth carrying at all.
The spoke that matters most is the one people try to cut first. Every delivery team needs someone who knows how the business actually works — fraud rules in a bank, clinical workflow in a hospital, how freight moves through a logistics firm. Different knowledge, same rule. A model built without that person is technically correct and operationally useless. It is the difference between a science project and a capability.
This is also why the reporting line matters more than it looks. The value in AI lives in the business, not the technology function, so the office has to sit close to it and carry enough credibility to change how the work is done. Bury the hub deep inside IT and even a strong central team produces work nobody runs.
Zoom in and the same shape repeats at the delivery level. One small cross-functional squad owns a use case from intake to run — a product owner who holds the value case, the data and ML engineers who build it, an MLOps engineer who runs it in production, an embedded domain expert, and a change lead who gets people to use it. The same faces the whole way through. That continuity is what stops value leaking out of handoffs between separate teams.
Around the pod, the hub shows up as a service rather than a gatekeeper: the platform team, governance, and finance support the work without sitting in its way. Every use case moves through the same gates in the same order, so the tenth is faster and calmer than the first — the first version is crafted, the tenth is assembled from reusable pieces the hub maintains.
Centralize the leverage. Distribute the relevance. Keep the two wired together through a pod that owns the outcome end to end. That is the whole argument for hub-and-spoke, and it is why it keeps outlasting the tidier alternatives on either side of it.
Bring us where you're stuck — a mandate, a stalled pilot, or the whole build. We'll tell you where we'd start.