César Cernuda is the president of NetApp.
I talk to business leaders around the world every week, and I often hear a version of the same story: An AI pilot went well, and the team was behind it. But when the moment came to move from experiment to production scale, the results never arrived.
The diagnosis is almost always the same, too: The model wasn’t quite right, the use case wasn’t well defined, or the team wasn’t ready. Leaders reach for those answers because they’re easy to explain and safe to put on a slide. But in most cases, none of them is the real reason things stalled.
I’ve found that the problem often sits underneath the AI, not inside it. It lives in the data pipelines that feed the model, in the rules that govern which data to trust and in systems built for an earlier era now asked to carry workloads they were never meant to handle. We keep pointing at the engine when the problem is the fuel.
The Gap Between AI Ambition And AI Reality
Every major company I visit has a genuine ambition for AI. The boardroom conversations are urgent, and the expectations are high. Yet, the gap between what leaders expect and what AI delivers in practice stays stubbornly wide. This is because those leaders aren’t looking for the answer in the right place.
Countless companies have put serious money into the visible parts of AI: the models, the compute, the use cases and the tools. But I’ve found they’ve spent far less on the data layer those models depend on. Stale training data teaches a model about a world that no longer exists. Research from MIT’s NANDA initiative found that 95% of generative AI pilots fail to generate meaningful revenue impact at scale, which says something important about where the problem actually lives. A schema change can quietly break a system in ways that take weeks to trace. Infrastructure built for reports and dashboards often can’t handle what production AI actually demands.
The companies that want to get real, lasting value from AI must make the harder call early and put time and money into the parts of the stack that don’t get the glory. That’s exactly where AI works or doesn’t.
What Getting It Right Actually Requires
The strongest AI plans I’ve seen share a common starting point: Before asking which model to use, leaders must ask whether their data is actually ready for what they want AI to do. That means clear rules about what data can be trusted, quality checks treated as a live operational need rather than a one-time fix and systems that can keep up with what AI demands in production.
None of these are new ideas. What’s new is the cost of getting them wrong. Informatica’s “CDO Insights 2025” survey of 600 chief data officers found that 92% are concerned that AI pilots are moving forward without resolving the underlying data challenges first—a sign that most organizations know exactly where the risk is and are proceeding anyway.
As AI takes on more active roles and makes decisions with less human review at each step, bad data doesn’t just create errors. It creates decisions that grow into bigger problems before anyone notices.
The Commitment Leaders Must Make
I believe the companies willing to do this work will have a distinct competitive advantage in the long run. We tend to talk about AI as if everyone is on the same playing field, choosing between the same models and chasing the same use cases. But the real edge isn’t the model. It’s what the model has to work with.
A company that has spent two years organizing its data pipelines, governing its assets and building infrastructure matched to AI demands isn’t just better set up to run AI. It can see returns that compound and real value created, unlike companies that are stuck cycling through AI pilots.
To put it simply, the infrastructure question is the AI question. The AI race is on, and there is still a window of opportunity for companies to get ahead if they get their data foundation right.
Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?
