Every analytics leader at a $100M+ business eventually inherits the same fight: should the data team sit in the middle (hub-and-spoke) or live inside the business units (federated)? The honest answer is that neither model is correct in the abstract. Both work. Both fail. The question worth asking isn’t which is better — it’s which set of failures you are currently living with, and which set you would rather trade for.
The two shapes, briefly
Hub-and-spoke puts a central analytics team in charge of the warehouse, the semantic layer, the reporting platform, and the standards. Domain analysts may exist in the business, but they consume from a curated middle. The benefit is one source of truth. The cost is a ticket queue.
Federated analytics pushes ownership into the domains — finance owns its models, supply chain owns its forecasts, marketing owns its attribution. A small platform team keeps the lights on for shared infrastructure. The benefit is speed and context. The cost is divergence: three teams, three definitions of “active customer.”
Most real organizations are somewhere on the slider between these poles, often unintentionally. The signals below tell you which direction to push.
Signals that you should centralize
Multiple dashboards disagree, and nobody can adjudicate. When the CFO’s revenue number, the CRO’s revenue number, and the board pack’s revenue number all differ by 2–4%, you do not have a reporting problem. You have an ownership problem. Decentralized teams have each defined the metric correctly for their own use case, and there is no authoritative source. This gets worse, not better, as the company grows. Centralization solves it by making one team accountable for the canonical definition.
Audit or compliance findings on your numbers. External auditors, SOX controls, or regulatory inquiries that flag inconsistent calculations or lineage gaps are unambiguous signals. Federated models struggle here because audit trails span teams. A central platform with enforced semantic-layer governance makes the audit response a tractable engineering problem instead of a cross-functional negotiation.
Leadership is making decisions on bad numbers. If you can’t trust the deck in the executive review, centralize. Period. The cost of a wrong strategic call dwarfs the cost of a slower analytics team.
Tooling sprawl is destroying margin. Six BI tools, four warehouses, redundant ETL jobs all pulling the same Salesforce extract — this is what federation looks like when nobody is steering. A central team can rationalize the stack and reclaim the spend.
Signals that you should decentralize
The central team has a 6+ week ticket queue, and the business has stopped asking. This is the most dangerous failure mode of hub-and-spoke, because it’s invisible. Demand doesn’t disappear when the central team can’t service it; it goes underground into spreadsheets, shadow Tableau instances, and gut feel. By the time the central team realizes what’s happening, half the organization has routed around them.
Domain decisions are being made without domain context. When a central analyst produces a supply-chain forecast without understanding lead-time variability by SKU class, the output is technically correct and operationally useless. Domains starved of analytics context will make worse decisions than domains with mediocre embedded analysts. Embed.
Innovation has stalled. Centralized teams optimize for consistency and reliability — virtues that, taken too far, become risk-aversion. If the data team hasn’t shipped a genuinely new analytical capability in 18 months, the model is calcifying. Pushing ownership outward reintroduces creative pressure.
The central team is the bottleneck on hiring. If you cannot hire central analysts fast enough to meet demand, stop trying. Hire domain analysts into the business units instead and give the central team a smaller, sharper mandate: platform, governance, shared semantic layer.
What to actually do
Treat the model as a dial, not a switch. The strongest analytics organizations we see run a federated execution layer (analysts embedded in finance, ops, commercial) on top of a centralized platform layer (one warehouse, one semantic layer, one set of canonical metrics, one governance regime). The central team owns the substrate; the domains own the questions.
When you’re debating which way to move, the signals above are the right input. When the disagreements are about what the numbers mean, centralize the substrate. When the bottleneck is who can answer the question fast enough with enough context, decentralize the execution.
The mistake is picking a model on principle and defending it through the failures. Pick the model that fixes today’s failures, and be honest about which new failures you’ve just bought.