Five percent of companies capture AI value. The rest are fixing the wrong problem.
Only five percent of organizations are generating material returns from AI investments. The other ninety-five percent are spending without seeing results.
The difference is not deployment speed or technology choice. The five percent are redesigning how work gets done. The rest are inserting AI into broken operating models and expecting it to fix them.
BCG's 2025 study of more than 1,250 firms and McKinsey's global survey both show the same pattern: the strongest predictor of AI value is not investment, talent, or model selection. It is redesign of the work itself.
The gap between pilots and understanding
A CEO showed her board pack: five AI pilots running and a twelve percent margin target for 2026. When asked which pilots she had been inside, which workflows she had walked, and which decisions the AI was now making, she paused. She had not been inside any of them.
She was reporting on the system. She did not understand it.
This is common. Most executives treat AI as a technology deployment problem. They approve pilots, distribute them across functions, track progress and report to the board. This works for cyclical upgrades. It fails for structural change.
Why AI exposes what was already broken
When AI enters an operating model, it does one of two things: it compounds what works, or it accelerates what does not.
In hierarchical models, AI compresses analysis but not authority. Decisions still move through layers, so cycle time falls in one phase and rises in another. The bottleneck shifts rather than disappears.
In models built around manual execution, AI lifts productivity within tasks but leaves the queues between them untouched. Output improves incrementally, while expectations remain structural.
These are not technology failures. The model was already misfitted to its conditions. AI makes that misfit visible, and faster than leadership can ignore.
The five percent do something different
High performers are not faster at deployment. They are different in kind.
They do not insert AI into the existing model. They design a new one. The unit of design shifts from a human team supported by tools to a human-agent system, where decision rights, accountability, judgment and execution are distributed across both.
The data shows this works. The five percent already generate 1.7 times more revenue growth and 3.6 times greater shareholder return than companies around them. The gap compounds.
Three disciplines, in order
High performers follow three disciplines - and crucially, they sequence them. Most organizations do not.
RAZE comes first. This is strategic decommissioning of programs, structures and the underlying assumptions that sustain them - including the idea that all decisions default to humans. Without RAZE, AI is layered onto a system not designed to share work.
ENRICH follows. This is preparation of the substrate: governance, data and trust. Governance defines how decisions are distributed between humans and agents. Data must be clean, contextual and continuously available. Trust must be calibrated, with human oversight where the stakes require it. Without ENRICH, AI amplifies disorder.
GROW comes last. This is not iteration on the existing model but design of the next one - starting from the work itself, not the legacy structure that once delivered it.
The board's role shifts
A Controller board approves pilots and tracks margins. A Steward board asks different questions.
Before the next AI investment is approved, these four belong on the agenda:
- Which decisions does AI currently own in this organization, which does it augment, and which remain exclusively with humans - and who has verified that mapping?
- Which operating assumptions embedded in our current model were designed for conditions that have since moved?
- What does our CEO understand directly about the AI workflows we are governing - not reporting on, but working inside?
- How do we know whether our AI investment is compounding disorder or reducing it?
If those answers are not clear, the outcome is predictable.
The CEO shift mirrors this
The Controller CEO delegates AI and reports outcomes. The Steward CEO is inside the work - understanding which decisions AI makes, which it augments and which remain human.
She does not report the system. She understands it.
Survey data shows the gap. Directors report cohesion at the top, while executives do not. Boards assign ownership of AI strategy to the C-suite, yet the C-suite itself does not agree on who owns it.
Misalignment at the level of diagnosis creates misalignment everywhere else. The gap closes only when both parties are working from the same structural view of the system.
The structural decision comes first
Before the next AI investment: which decisions will AI own, which will humans own, and which will they share?
The structural decision comes first. The technology decision follows.
In that order, value appears. For more on how executives and boards should approach AI strategy, see AI for Executives & Strategy and AI for Management.
Your membership also unlocks:
