Make your Models Talk
Here is a scenario playing out right now inside large CPG organizations. A pricing team builds a model. An advertising team builds a model. A demand planning team builds a forecast. All three are reasonably well-constructed. None of them are connected. The pricing team optimizes their number. The advertising team optimizes theirs. And demand planning, working from neither, makes its best guess.
The result: a decision gets made on pricing that never makes it into the forecast. A marketing pivot gets executed that nobody told supply chain about. Products go out of stock on the exact SKUs the market just started pulling for.
The models were right. The outcome was wrong.
A Director of Strategic Growth Management at a Fortune Global 100 CPG company describes this as one of the most consequential failures she has seen in various AI-enabled organizations. Pricing and advertising models were each doing their jobs in isolation. Demand planning, sitting in a separate silo, had no visibility into either. Decisions optimized within each function were actively contradicting each other at the enterprise level. The work wasn’t throwaway because it was bad. It was throwaway because it was unconnected.
A different Global Director of Marketing and Commercial Intelligence at a Fortune Global 400 company, frames the solution in terms of architecture: a single source of truth that every function draws from and every model feeds back into. Not a reporting layer on top of disconnected systems. A foundation underneath everything, built upstream, that makes the outputs of one model available as inputs to the next.
The SGM Director is direct about what this requires organizationally: someone elevated enough to see across all of it. Not a committee. A person with accountability for how the models connect, what the end-to-end architecture looks like, and where the gaps are creating decisions that contradict each other. Without that role, every team optimizes locally and the enterprise pays for it globally.
AI that optimizes one function at the expense of another isn’t intelligence. It’s a more sophisticated way to work in silos.