Only 6% of organizations are generating meaningful value from AI. That figure comes from a McKinsey survey of nearly 2,000 organizations across 105 countries, and it includes the companies with the largest budgets, the deepest talent, and full access to frontier models. The other 94% are spending real money on AI without seeing it accumulate into anything that reaches the business.

That gap has very little to do with model quality. Access to foundation models has largely commoditized, and the organizations in that survey can all reach for the same tier of capability, whether they build on GPT, Claude, or Gemini. I have spent much of the last year talking to CTOs and heads of AI at large organizations, and the ones generating real returns are not doing it because they reached the better model first. The separation is happening at a different layer, and where that layer sits changes how you should think about everything you are currently building.

Fewer than one in five organizations track well-defined KPIs for their AI use. The rest are measuring inputs: licenses purchased, seats activated, tokens consumed. What none of those numbers capture is what the AI returned, per task, per workflow, per team, per decision. Underneath that gap is an assumption that value will emerge from volume, that enough people using enough tools will eventually compound into something. Enterprise software already ran this experiment twice. ERP and CRM both saw broad deployment and a wide spread of outcomes that no amount of input tracking could explain. The companies that got real returns were not the ones with the highest seat counts. They were the ones that built a feedback loop between what the technology did and what the business needed to know about it.

The heads of AI I talk to describe the same situation in almost the same words. They know their teams are using the tools, but they cannot tell which of that usage is producing results. The models are fully instrumented and the data exists, so the problem is not a shortage of signal. What they have not built is the layer that connects what their AI environment produces to the outcomes the business runs on. Until it exists, rising activity tells them nothing about whether returns are rising with it.

McKinsey identifies three factors that correlate with real EBIT impact: workflow redesign, KPI tracking, and senior leadership engagement. Those read like three separate initiatives. They are closer to three views of a single move, and it is worth describing what that move actually requires.

It starts with identifying the employees who get meaningfully better outcomes from AI than their peers, which people and which workflows consistently produce results tied to outcomes rather than to activity. Almost no organization I have spoken to has built the infrastructure for this. They can see who is using AI heavily, but not whose use is actually working.

The second part is capturing what those people are doing, which is rarely about the tools they run. It is the judgment they apply: the context they feed into a workflow, the outputs they trust, the decisions they make differently because of what the model returned. That judgment is what turns a correct answer into a useful one, and in most organizations it lives in a handful of heads with no mechanism to move anywhere else.

The third part distributes it, so the practices that made the top performers effective become institutional rather than personal. I think of these three together as the Organizational Intelligence Loop: identify who is performing, capture what they know, distribute it systematically. The 6% have built some version of it. Most of the rest are still expecting value to appear without it.

What those organizations have is not more advanced technology but something older and harder to build: organizational intelligence, the accumulated know-how of how a company actually works, which constraints are real, which edge cases the standard answer misses, what the customer meant rather than what they asked for. The strongest performers win less on prompting technique than on how much of that context they feed into the workflow. None of it is visible to the model unless a person puts it there, and the people who do it by instinct are the ones whose results the rest of the organization cannot reproduce, because what they add has never been written down.

This is where most companies stall. Their infrastructure reports on consumption, how many tokens, seats, and calls, with no way to show which workflows carry real institutional knowledge and which only generate activity. A company that cannot see which of its people are performing cannot capture what they know, and it has nothing to put in front of anyone else. The loop never gets past the first step, not for lack of ambition, but because the systems they built were never designed to see what their best people actually do.

The infrastructure most companies built answers one question well: what did we spend? That was the right question when AI was mainly a cost to control, and it is no longer enough. The 6% have built something that answers a harder one: what do our best people know, and how do we get it in front of everyone else? The model you run today will be available to your competitors within the year. What your best people know, and whether your infrastructure can put it to work across the organization, is the advantage that will not commoditize. That is the thing worth building now, while the other 94% are still counting seats.