Shile Adebimpe
Director, Data & AI Engineering · 8 min read
An enormous amount of executive energy goes into comparing AI vendors, architectures, and model providers, and a comparatively tiny amount goes into the decision rights, governance forums, and ownership structures that actually determine whether any of those choices ever scale past a single team's pilot.
The operating model is the unglamorous machinery underneath the strategy: who decides which use cases get funded, which forum reviews risk, whether a Center of Excellence sets enforceable standards or merely publishes suggestions, who owns the budget, and, critically, who is accountable for the outcome once the system is in production and something inevitably needs a decision made quickly.
The failure modes are strikingly consistent across industries. A Center of Excellence is created with responsibility for AI standards but no authority to enforce them, so business units quietly build shadow AI capability outside its view. No single executive owns AI outcomes across the enterprise, so when something goes wrong, the post-mortem becomes a search for whose job it was rather than a fix.
The organizations that get this right rarely design a perfect enterprise-wide operating model on a whiteboard before doing anything. They pilot a operating model on one business domain, learn what breaks, and scale the model, not just the technology, once they've seen it function under real conditions with real stakeholders pushing back on it.
The right operating model also isn't static. Early in an AI program, tighter central control is usually right, because governance muscle doesn't exist yet and the cost of an early mistake is reputational as much as financial. As the organization matures, federating more authority to business units becomes viable, because the governance discipline has had time to become real practice rather than a slide.
Pick whichever model architecture you like; that decision is recoverable. But if the operating model around it doesn't specify clear ownership, the initiative will fail for organizational reasons long before anyone gets the chance to find out whether the model was any good.
“No one has ever cancelled an AI program because the model underperformed. They cancel it because no one could agree who owned it.”