The current enthusiasm for AI in large organizations skips a step. The conversation jumps straight to models, prompts, and pilots, as if the constraint were the technology. In most enterprises, it is not. The constraint is everything underneath the model, and that is where the ambition quietly fails.
AI is useful only when the data, the operating model, and the adoption path are already credible. Take any of the three away and the most capable model in the world produces confident output that no one trusts or acts on. This is not a reason to avoid AI. It is a reason to be honest about the order of operations.
Start with data. A model is only as good as what it reads. If the same metric is defined three ways, if ownership is unclear, if the logic lives in undocumented spreadsheets and legacy reporting, then pointing an AI system at that estate does not fix it. It launders it. The model will produce an answer, fluent and plausible, built on data the organization already did not trust. The failure is harder to see than a broken dashboard, because the output looks authoritative. Garbage in, eloquence out.
Then the operating model. Suppose the AI system produces something genuinely useful. Who owns it. Who is accountable when it is wrong. What decision is it allowed to drive, and on what cadence does it run. These are the same questions that determine whether any data product survives, and AI does not exempt you from them. A pilot can ignore them. A system in production cannot. Most enterprise AI is stuck in pilot precisely because the operating model around it was never designed, only the demo was.
Then adoption. An AI capability that people do not trust enough to act on is a research project, not a transformation. Trust is earned the same slow way it always was: by being right, transparently, in cases people can check, until the organization is willing to change a decision based on it. There is no prompt for that. There is only the unglamorous work of building credibility one verifiable result at a time.
This is why I am skeptical of AI programs that lead with the model. The model is the most visible part and usually the least decisive. The decisive work is upstream and unfashionable: governed data, clear ownership, defined decisions, and an adoption path that respects how trust actually forms. Organizations that do this groundwork find that AI slots into a foundation that was already sound. Organizations that skip it find that AI makes their existing problems faster and more confident.
There is an uncomfortable implication here. A lot of what passes for AI strategy is an attempt to buy your way out of foundational work the organization did not want to do. The data was a mess, the ownership was vague, the operating model was implicit, and AI arrived looking like a way to skip all of it. It is not. It is a magnifier. It makes a credible foundation more valuable and a shaky one more dangerous.
So my advice to leaders chasing AI is consistent and slightly deflating. Before the model, fix the data foundation and decide who owns it. Before the pilot, design the operating model the system will run inside. Before scale, earn the adoption. The teams that do this look slower for a quarter and then pull decisively ahead, because they built something that holds. The teams that lead with prompts get an impressive demo and a system no one will stake a decision on.
AI does not remove the need for an operating model. It raises the price of not having one.