Why 94% of Companies Don't Profit from AI — and What the 6% Do Differently

The numbers are stark, and they haven't moved.

McKinsey's The State of AI in 2026 surveyed 1,719 organizations across 97 countries. Nine in ten use AI in at least one business function. Only 6% get more than 5% of their operating profit from it. And that 6% is the same 6% McKinsey found a year earlier, before the current generation of models shipped. Adoption rose over those twelve months. The share of organizations profiting from AI did not move by a point.

Read that again. A year of better models, wider deployment, and larger budgets — and the profit line didn't budge.

The same technology at the same price

Here's what makes the finding difficult to explain away: every organization in that survey has access to substantially the same models at substantially the same price. The 6% did not buy better technology. There is no secret model, no privileged API tier, no vendor relationship that explains the gap.

So what does?

Among the 6%, nearly three-quarters fundamentally redesigned how work flows through the organization. Among everyone else, one quarter did. That is the entire difference, and it isn't a technology variable at all.

Redesigning work is not a technical act

"Redesigning how work flows" sounds like an operations project. It isn't. It means deciding who no longer approves something, which handoff disappears, what a role is now for, and who absorbs work that used to justify a headcount.

Those are leadership decisions. They are also the decisions leaders are least equipped to make, because their own authority is defined by the design they would have to change. The approval step you're being asked to remove might be yours. The handoff that disappears might be the reason your span of control looks the way it does. The headcount absorbed might report to you.

This is why AI transformations fail at the leadership layer, not the technology layer. The tooling works. What's missing is a leader willing to dismantle the structure their own position rests on — and an organization able to absorb the result.

Why another platform won't fix it

The standard corporate response to the 6% problem is to buy something: a better platform, an AI transformation consultant, a maturity model with five stages and a roadmap. But the bottleneck was never the tooling. Nine in ten organizations already have the tooling. What they don't have is leaders who can make the redesign decisions, in organizations that can survive them.

This is worth saying plainly because the market keeps selling the opposite: that the separator is technological, and therefore purchasable. It isn't, and the McKinsey data is the receipt. Same models, same prices, wildly different outcomes. The variable is what leaders did to their own organizations.

What the 6% actually did

Strip away the mystique and the 6% did something unglamorous: they changed who decides what.

They removed approval layers that existed to demonstrate control rather than exercise it. They collapsed handoffs between teams that were passing work back and forth to protect turf. They redefined roles around outcomes instead of activities. None of this required a new model. All of it required someone with authority to spend that authority on changing the design — including the parts that benefited them.

That's a leadership capability, not a technology adoption curve. And it's measurable: you can assess whether an organization's leaders make these decisions, and whether the structure around them can absorb the result, before you spend another dollar on tooling.

The excuses that don't survive the data

Every organization in the 94% has a story about why the returns aren't there yet. The data rules most of them out.

"We need cleaner data." Every company in the survey faced the same data reality, and 6% profited anyway. Data quality is a real constraint on specific use cases. It is not the explanation for a 94% failure rate.

"The models aren't good enough yet." Adoption rose for twelve months while the profit share didn't move a point. The models got better. The outcomes didn't. Whatever the bottleneck is, it isn't model capability.

"Our people need more training." Training teaches people to use tools inside the existing design. The 6% didn't train harder — they redesigned the design. If your AI training programme ends and everyone goes back to the same approval chains and handoffs, you trained people to be faster inside a structure that still doesn't profit.

"Our industry is different." 97 countries, 1,719 organizations, every sector. The pattern held. Your industry is different the way every industry is different — and the separator was the same.

What's left, once the excuses are gone, is the leadership question. That's uncomfortable, which is exactly why the excuses persist.

What a serious response looks like

If the diagnosis is right, the prescription isn't another tool. It's three things, in order.

First, find out whether your leaders can make redesign decisions — and whether your organization can absorb them. That means measuring: the psychological resources of the people being asked to change, the real network work flows through, and whether the structure leaves any room for a new way of working to survive. Measurement before budget.

Second, pick one workflow and actually redesign it. Not a pilot with seventeen governance gates — a real redesign: who no longer approves, which handoff disappears, what the role is now for. One workflow, done properly, teaches the organization more than an enterprise rollout teaches anyone.

Third, develop the leaders, not just the users. The industry has spent three years training employees to prompt and zero years developing leaders to dismantle. That ratio is the 94%.

The uncomfortable implication

If you're leading an AI effort that isn't producing profit, the question probably isn't which model you're on or whether your data is clean enough. Those are comfortable questions — technical, delegable, solvable with budget.

The uncomfortable question is whether your organization's leaders have redesigned how work flows, or whether they've automated the existing design and called it transformation. Automating a bad design just produces bad outcomes faster. The 6% understood this. The 94% bought software.

The work, then, is developing leaders who can make redesign decisions — and measuring whether the organization around them can carry the result. That is a leadership problem, and it does not get solved by another platform.