Executive Summary
Nearly nine in ten large organizations now use artificial intelligence in at least one function. Almost none of them are winning with it. McKinsey’s 2025 global survey finds that while 88 percent of organizations regularly use AI, only about 6 percent qualify as high performers attributing more than five percent of earnings to it. Adoption has become universal; advantage has not. When everyone owns the same tool, owning it stops being a strategy.
This article argues that the basis of competition is shifting from the assets a firm accumulates to the speed and quality with which it learns. I call this capability Enterprise Intelligence: the organizational capability to continuously sense, reason, decide, execute and learn by combining people, AI, enterprise knowledge and governance.
Scale is not dead. But scale is becoming an amplifier rather than a moat. It multiplies whatever intelligence a company already has, for better or worse. The firms pulling away are not the ones with the most AI projects. They are the ones that have turned learning into an operating discipline, and let it compound.
Opening
Two firms, one technology, opposite outcomes
Consider two retailers of comparable size, buying the same forecasting software from the same vendor in the same quarter. A year later one has cut stockouts and freed working capital; the other has a dashboard nobody trusts and a pilot quietly winding down. The software was identical. What differed was everything around it: whether the forecast actually changed a buyer’s decision, whether the decision was executed without three layers of override, whether the outcome was measured and fed back so the next forecast was better. One firm bought a tool. The other built a loop.
This is the central puzzle of the AI era, and the data now states it plainly. The technology has been commoditized faster than almost any general-purpose technology in history. The cost of a unit of capable inference fell more than two-hundred-fold in roughly eighteen months, by Stanford’s account. Yet the spread in business outcomes has widened, not narrowed. Cheaper, more powerful, more available AI has not produced convergence. It has produced separation.
That separation is the subject of this article. It cannot be explained by who has the best model, because the best models are available to everyone. It can only be explained by something the firm itself possesses: a capability that turns identical inputs into divergent results. Naming that capability, and learning to manage it, is the work ahead of every board and chief executive.
“The greatest danger in times of turbulence is not the turbulence; it is to act with yesterday’s logic.” Peter Drucker
The Argument · I
A short history of competitive advantage
Strategy has always been the search for an advantage that competitors cannot easily copy. What counts as durable, however, has changed with each economic era, and understanding that progression is the precondition for seeing what is changing now.
The industrial era: advantage as accumulation
For most of the modern corporate age, advantage was something you accumulated. Financial capital, physical plant, distribution networks and scale economies were the moats. Bigger factories produced cheaper goods; wider distribution reached more customers; deeper balance sheets outlasted rivals. Michael Porter codified the logic: a firm wins by occupying a defensible position in its industry’s structure, through cost leadership, differentiation or focus. Advantage was a place you held, defended by assets that were expensive to replicate.
The digital era: advantage as position in a network
The internet shifted the locus of advantage from physical assets to network effects and data. Platforms grew more valuable as more people used them; the marginal cost of serving the next customer collapsed toward zero. Advantage became less about what you owned and more about the position you occupied in a system of users and information. Yet even here, the underlying logic remained one of accumulation: accumulate users, accumulate data, accumulate switching costs.
The AI era: advantage as rate of learning
The current era breaks the pattern. When a frontier capability is available to every competitor as a metered service, owning it confers no edge. What cannot be bought off a price list is the organizational ability to convert that capability into better decisions, faster, and to improve at doing so over time. Rita McGrath warned years ago that competitive advantages are becoming transient, that the sustainable-advantage assumption itself was eroding. The AI era resolves her paradox in an unexpected way: the one advantage that endures is the capacity to keep generating new advantages. The moat is no longer a position. It is a rate.
The Argument · II
Why scale alone is no longer enough
The instinct of every large incumbent is to meet a new technology with scale: more budget, more pilots, more headcount, more compute. The 2025 evidence suggests this instinct is now actively dangerous, because scale applied to an unimproved operating model simply industrializes mediocrity.
The pattern in the data is stark. Roughly two-thirds of organizations have not begun scaling AI across the enterprise; most remain in what practitioners now call pilot purgatory, perpetual experiments that never reach production. McKinsey’s analysis of where value actually originates is unambiguous: of all the factors tested, the redesign of workflows has the single largest effect on whether AI moves earnings. Not the model. Not the budget. The work itself.
This is why scale, on its own, has become a trap. A larger organization that has not redesigned how it senses, decides and learns will deploy AI faster into more processes, and thereby scale its existing dysfunctions. The firms pulling ahead invest at materially higher levels (more than a third of high performers commit over a fifth of their digital budget to AI), but the spend is downstream of a deeper choice: to treat AI as a reason to rebuild the operating model rather than to ornament it. Scale amplifies. The question every leader must answer is: amplify what?
The Framework
Enterprise Intelligence, defined
If advantage is now a rate of learning, the strategic question becomes: a rate of learning at what, and produced how? The answer is a capability that no single technology supplies and no vendor can sell whole. I define it as follows.
Definition
Enterprise Intelligence
The organizational capability to continuously sense, reason, decide, execute and learn, by combining people, AI, enterprise knowledge and governance.
It is not a model, a platform, or a department. It is a property of the whole organization, the way liquidity or culture is a property of the whole organization. And like those, it can be cultivated or neglected, measured or ignored.
Each word in the definition is load-bearing. People supply judgment, accountability and the questions worth asking. AI supplies pattern recognition, synthesis and speed at a scale no workforce can match. Enterprise knowledge, the proprietary data, context and institutional memory unique to the firm, is what makes a generic model specifically yours; it is the one input competitors genuinely cannot copy. And governance is not a brake on this system but its steering: the trust, controls and feedback that let an organization act on machine output with confidence. Remove any one element and the capability collapses into either ungoverned automation or expensively augmented inertia.
This reframes the AI conversation entirely. The board-level question is not “how many AI projects do we have?” but “how good is our enterprise at sensing, reasoning, deciding, executing and learning, and is it getting better?” The first question counts inputs. The second measures a capability.
The Framework · Mechanism
The Enterprise Intelligence Flywheel
Capabilities that compound do so because they are circular. The reason high performers pull away quarter over quarter, and the reason the gap widens rather than closes, is that Enterprise Intelligence is a flywheel, not a funnel. Each turn makes the next turn easier. The mechanism runs in six stages.
The flywheel also explains the most counter-intuitive finding in the 2025 evidence: that the leaders’ advantage is accelerating. In a linear model, latecomers eventually catch up. In a compounding one, a head start in learning rate widens the gap mechanically, because the leader’s wheel is turning faster and each rotation adds proprietary knowledge that rivals cannot purchase. This is the structural reason behind what observers have begun calling an emerging bifurcation: a small group pulling away while the majority adopt without transforming.
The Argument · III
Enterprise Intelligence in the wild
The framework is not abstract. It is visible, in different forms, in the firms that have outperformed across the digital and AI eras. What unites them is not their industry or their technology stack but the presence of a working loop.
Amazon: the decision-velocity machine
Amazon’s durable edge has never been a single algorithm. It is an operating model engineered to shorten the distance between signal and action: the two-pizza team, the famous distinction between reversible and irreversible decisions, the relentless instrumentation of everything. Its recommendation and logistics systems are the visible AI; the invisible advantage is an organization built to decide and execute at a cadence rivals cannot match. Amazon optimized the flywheel decades before the word became fashionable.
Microsoft and NVIDIA: knowledge and infrastructure as flywheels
Microsoft’s resurgence under Satya Nadella was, in his own framing, a shift from a culture of knowing to a culture of learning. NVIDIA’s position is even more instructive: it does not merely sell the picks and shovels of the AI economy; it has reconceived the data center itself. As Jensen Huang puts it, these facilities are no longer storage. “They are, in fact, AI factories. You apply energy to it, and it produces something incredibly valuable.” A factory that produces intelligence is the literal expression of an Enterprise Intelligence flywheel.
Netflix, JPMorgan, Toyota, Costco: four routes to the same property
The capability expresses itself differently by context. Netflix turned viewing data into a content-investment engine, sensing demand, reasoning about what to commission, executing, and learning from the result. JPMorgan has industrialized risk and fraud detection while building governance robust enough to act on machine output in a regulated setting, proof that governance is an accelerator, not a constraint. Toyota’s decades-old practice of kaizen is, in essence, a pre-digital learning loop; its challenge now is to instrument that loop with AI without losing its human discipline. Costco demonstrates the inverse lesson: a deliberately lean technology footprint paired with extraordinary operational learning, a reminder that Enterprise Intelligence is about the loop’s quality, not the tooling’s expense.
The common thread is the absence of the thing most firms have: a pile of disconnected pilots. In each case the intelligence is wired into how the company actually runs.
“AI is now infrastructure, and this infrastructure, just like the internet, just like electricity, needs factories.” Jensen Huang, NVIDIA, COMPUTEX 2025
For the Chief Executive
How CEOs should think differently
The chief executive’s instinct under technological pressure is to delegate: to appoint a head of AI, fund a program, and ask for a quarterly count of initiatives. The 2025 evidence is decisive against this reflex. High performers are roughly three times more likely than their peers to have senior leaders who personally own and visibly use AI. Enterprise Intelligence cannot be delegated downward, because it is a property of the operating model, and only the chief executive owns the operating model.
Practically, this means shifting the questions one asks. Below are ten I would put to any leadership team, not as a checklist, but as a diagnostic of whether the flywheel actually turns.
- For our most important recurring decision, how many hours pass between the signal arriving and the action being taken, and is that interval shrinking?
- When AI produces a recommendation, who is accountable for acting on it, and what stops them from overriding it out of habit?
- What proprietary knowledge do we hold that no competitor can buy, and is it reaching the point of decision, or trapped in silos?
- How much of what we learned last quarter has changed how we operate this quarter?
- Which of our workflows have we genuinely redesigned around AI, versus merely automated in place?
- Do our people trust the system enough to act on its output without a parallel manual check?
- Where is the flywheel’s slowest joint, whether data, reasoning, decision or execution, and who owns fixing it?
- Are we measuring the rate at which we learn, or only the number of things we have launched?
- Does our governance let us move faster with confidence, or does it exist mainly to slow us down?
- If a rival’s learning rate is double ours, how many quarters until that gap is unrecoverable?
The shift in posture
From sponsor to architect
The CEO of the digital era sponsored technology programs. The CEO of the AI era architects the loop, taking personal ownership of how the firm senses, decides, executes and learns, because no one else has the authority to redesign the operating model end to end.
For the Board
How boards should think differently
Boards are, by design, instruments of measurement and accountability. The difficulty is that they are currently measuring the wrong thing. A board that asks management “how many AI projects are underway?” is counting activity and mistaking it for progress, the governance equivalent of judging research by the number of experiments rather than the discoveries. McKinsey’s data shows AI governance is increasingly a board-level concern; the task now is to give it the right instruments.
A useful board dashboard abandons project counts in favor of seven indicators of capability. Each maps to a stage of the flywheel or to the trust that lets it turn.
Implications
What this changes for the business
If Enterprise Intelligence is the new locus of advantage, several established management assumptions require revision.
Strategy becomes the design of loops, not the selection of positions. Porter’s question, which defensible position to occupy, gives way to a prior one: which decisions matter most, and how do we make them better faster than anyone else? Competitive analysis shifts from market structure to relative learning rate.
Investment logic inverts. The temptation is to fund the most projects; the discipline is to fund the slowest joint in the most valuable loop. A dollar that unblocks a stalled flywheel is worth more than a dollar spread across ten new pilots, which is precisely why high performers concentrate spend and redesign workflows rather than sprinkling AI across processes.
Governance moves from the back office to the engine room. In a world where firms act on machine output, the trust to act is the advantage. Robust governance, meaning auditability, human-in-the-loop where it matters and clear accountability, is what lets a regulated firm move at the speed of an unregulated one. Stanford’s data on rising AI incidents underscores the point: as deployment scales, the governance moat becomes structural, not optional.
Talent strategy reorients around augmentation. The evidence suggests AI is changing what people do more than how many are employed; demand for engineers and data talent rose even as workflows were redesigned. The winning organizations design work so that machines handle what they do well and people focus on judgment, accountability and the questions worth asking, the human stages of the flywheel.
Looking Ahead
Predictions for the next decade
I offer these not as forecasts of technology but as consequences of the compounding logic described above.
The performance gap will widen before it narrows. Because learning rate compounds, the present bifurcation, a small group of high performers separating from a large majority, will intensify through the late 2020s. Late movers will find that catching up is mathematically harder each year, not easier.
Agentic systems will raise the stakes on the operating model, not lower them. As autonomous agents move from experiment to production, the binding constraint shifts decisively from model capability to the firm’s ability to govern, trust and integrate machine action. Organizations that have not built the loop will find agents amplify their dysfunction faster.
Boards will add Enterprise Intelligence to their standing agenda. Just as cybersecurity and ESG became board-level disciplines, the capability to learn will earn its own dashboard, its own committee attention, and eventually its own disclosure expectations.
Proprietary knowledge will be revalued as the scarce asset. When models are commodities, the durable input is the firm’s own context. Expect knowledge architecture, meaning how an organization captures, structures and routes what it uniquely knows, to become a recognized source of enterprise value.
The CEO job description will change. Within the decade, “owns the firm’s learning rate” will read as naturally in a chief-executive mandate as “owns the P&L” does today.
Conclusion
The advantage that writes tomorrow’s logic
For two hundred years, the companies that won were the ones that accumulated the most: capital, plant, distribution, users, data. That logic built the modern corporation, and it is not wrong so much as it is finished as a source of distinctive advantage. When the most powerful general-purpose technology in a generation is available to every competitor at a falling price, ownership confers no edge. What confers an edge is what a firm does with it: how quickly and how well it senses, reasons, decides, executes and learns.
That capability is what I have called Enterprise Intelligence. It cannot be bought, only built. It does not sit in a department, but suffuses the operating model. It is not measured by the number of AI projects underway, but by whether the organization is demonstrably getting smarter, quarter over quarter. And because it compounds, the firms that build it first will not merely lead. They will widen their lead, on terms that latecomers find increasingly impossible to match.
Drucker observed that the best way to predict the future is to create it. In the AI era, the firms that learn fastest do exactly that: they write tomorrow’s competitive logic before anyone else can read it. Scale still matters. But scale now answers to intelligence. The winners of this era will not be the largest companies. They will be the ones that compound.
Key Takeaways
- Adoption is universal; advantage is not. Roughly 88% of firms use AI, yet only about 6% capture material value. Owning the tool is no longer a strategy.
- The basis of competition has shifted from accumulation to learning rate. The durable moat is the capacity to keep generating new advantages, not any single position.
- Scale amplifies; it does not differentiate. Applied to an unimproved operating model, scale industrializes mediocrity. Redesigning the work, not the model, moves earnings.
- Enterprise Intelligence is the capability that matters: sensing, reasoning, deciding, executing and learning, through people, AI, knowledge and governance, together or not at all.
- It compounds because it is a flywheel. Learning re-enters as better data, so leaders accelerate away and the gap widens mechanically.
- CEOs must architect the loop, not sponsor projects; boards must measure capability, not count pilots. Proprietary knowledge and governance are the scarce, defensible assets.
A question for the boardroom
If a competitor’s learning rate were double yours, how many quarters would pass before the gap became impossible to close, and who in your organization owns the answer?
Selected Sources
- McKinsey & Company, The State of AI in 2025: Agents, Innovation, and Transformation, November 2025.
- McKinsey & Company, The State of AI: How Organizations Are Rewiring to Capture Value, 2025 (workflow-redesign and EBIT-impact analysis).
- Stanford Institute for Human-Centered AI (HAI), AI Index Report 2025, Chapter 4: Economy (inference-cost decline; corporate investment $252.3B; adoption).
- Stanford HAI, AI Index Report 2026 (organizational adoption 88%; investment trends; generative-AI diffusion).
- IBM Think, “Key Findings from Stanford’s 2025 AI Index Report,” 2025.
- NVIDIA, COMPUTEX 2025 keynote, Jensen Huang on AI factories and AI as infrastructure.
- Fortune, “Jensen Huang on AI factories and the future of manufacturing,” 2025.
- Peter F. Drucker, Managing in Turbulent Times and collected works.
- Michael E. Porter, Competitive Strategy (1980) and Competitive Advantage (1985).
- Rita Gunther McGrath, The End of Competitive Advantage (2013).
- Clayton M. Christensen, The Innovator’s Dilemma (1997).
- Andrew McAfee & Erik Brynjolfsson, Machine, Platform, Crowd (2017).
- Marco Iansiti & Karim R. Lakhani, Competing in the Age of AI, Harvard Business Review Press (2020).
- Satya Nadella, Hit Refresh (2017).
- Gartner, AI maturity and project-longevity research, 2025.
- Deloitte, State of Generative AI in the Enterprise, 2025.
- World Economic Forum, Future of Jobs Report, 2025.
- PwC, Global AI / CEO Survey, 2025.
- Menlo Ventures, State of Enterprise AI, 2025.
- Harvard Business Review and MIT Sloan Management Review, selected articles on AI operating models, 2024 to 2025.
Note: Statistics are drawn from the cited primary research as available at time of writing. Figures from rapidly updated annual reports (such as the Stanford AI Index and McKinsey State of AI) should be read against the most recent edition. Quotations are reproduced from public records of the speakers’ remarks and writings.