DIGITAL SUCCESS HUB | AI STRATEGY
A quiet correction is underway in how the most serious operators think about artificial intelligence. For three years, the boardroom question was which model to adopt. That question is now largely irrelevant, and the executives who continue to treat it as strategic are already behind.
The central competitive question of this decade is no longer which model a company licenses. It is whether that company builds a self-reinforcing system that turns its own daily work into a form of intelligence no competitor can rent. That distinction deserves far more attention than it currently gets, because it describes a shift in where corporate value actually sits, one that has not yet been priced into current AI strategy conversations, hiring plans, or capital allocation decisions.
This is not confined to any single company or industry. It is structural, and it applies with equal force to a regional bank, a logistics operator, an insurer, or a manufacturer, wherever in the world they compete.
The Price of Intelligence Has Collapsed
Start with what has actually happened to the underlying commodity. Frontier-class model performance that cost roughly $60 per million tokens in late 2021 is available today for a fraction of a cent for comparable capability, and independent tracking by Epoch AI puts the broader price decline for equivalent performance at somewhere between 9x and 900x per year, depending on the benchmark. Stanford’s AI Index has documented similar territory: GPT-3.5-equivalent quality fell from roughly $20 per million tokens in late 2022 to about $0.07 two years later, a nearly 280-fold drop.
Set the exact multiple aside. The direction is unambiguous, and it is the only direction this market has moved since 2023. Intelligence, in the narrow sense of raw model capability, is being priced like a utility. A CEO who spends a board cycle debating GPT versus Gemini versus Claude is optimizing a line item that is falling in cost by roughly an order of magnitude every year regardless of what they decide. That decision matters operationally. It does not confer advantage, because every competitor has access to the same falling price curve and, within a quarter or two, roughly the same capability tier.
This is the condition every company is now operating inside. When the tool becomes a commodity, the tool stops being where the value lives.
The Asset Nobody Is Measuring
There are two forms of capital at work here, and they get collapsed into one conversation far too often. The first is the model layer: the frontier system itself, increasingly interchangeable, increasingly cheap, and owned by a small number of labs regardless of who is using it. The second is the accumulated judgment layer: the corrections a company’s own people make to AI output, the evaluations that measure what “good” actually means inside that specific business, and the record of which decisions produced which outcomes.
The distinction that matters is that only the second layer grows in value in ways a competitor cannot access. A company can switch from one model provider to another in an afternoon. It cannot switch to a competitor’s six years of captured correction data, because that data does not exist anywhere it could be purchased. It exists only inside the operating history of the firm that generated it.
This is not a new insight dressed in new language. It restates something strategists have understood since well before generative AI: durable advantage has always come from what a firm does with a commodity input, not from possessing the input itself. What is new is the mechanism. For the first time, the accumulated judgment of a workforce can be captured as structured data, fed back into a working system, and made to improve that system’s output on the specific problems the firm actually faces. The learning is no longer trapped in the heads of senior staff who eventually retire. It can be encoded, and it can persist.
A Framework for Where Companies Actually SitIt is useful to separate two questions that get collapsed into one in most AI strategy conversations: does the company have access to strong models, and does the company have a structured way of capturing what its people learn from using them? Plotting these as two axes produces four positions.Rented Intelligence. Strong model access, no structured capture. This is the default position of the large majority of enterprises today. Output improves modestly. Nothing accumulates, because every session with the AI starts from zero institutional memory. This is where most AI spending currently sits, and it is why so much of that spending shows up as cost rather than advantage.Wasted Investment. Proprietary or fine-tuned models, no structured capture. Rarer, and arguably worse, because the sunk cost creates a false sense of differentiation. A custom model without a feedback discipline around it degrades in relative usefulness at the same rate the commodity market improves, since nothing is being fed back into it that a generic model does not already have.The Loop Builders. Commodity model access, disciplined capture of corrections, evaluations, and outcomes. This is the position available to any organization regardless of size, because it depends on process discipline rather than model ownership. A mid-sized regional firm can occupy this position as fully as a large multinational.Structural Advantage. Proprietary infrastructure combined with structured capture. This is rare, expensive, and largely the province of a handful of platform companies. It is not the realistic target for most operating businesses, and treating it as the goal is a common and costly strategic error.The useful move for almost every leadership team reading this is not to chase the fourth position. It is to move deliberately from the first to the third.
A Framework for Where Companies Actually Sit
It is useful to separate two questions that get collapsed into one in most AI strategy conversations: does the company have access to strong models, and does the company have a structured way of capturing what its people learn from using them? Plotting these as two axes produces four positions.
Rented Intelligence. Strong model access, no structured capture. This is the default position of the large majority of enterprises today. Output improves modestly. Nothing accumulates, because every session with the AI starts from zero institutional memory. This is where most AI spending currently sits, and it is why so much of that spending shows up as cost rather than advantage.
Wasted Investment. Proprietary or fine-tuned models, no structured capture. Rarer, and arguably worse, because the sunk cost creates a false sense of differentiation. A custom model without a feedback discipline around it degrades in relative usefulness at the same rate the commodity market improves, since nothing is being fed back into it that a generic model does not already have.
The Loop Builders. Commodity model access, disciplined capture of corrections, evaluations, and outcomes. This is the position available to any organization regardless of size, because it depends on process discipline rather than model ownership. A mid-sized regional firm can occupy this position as fully as a large multinational.
Structural Advantage. Proprietary infrastructure combined with structured capture. This is rare, expensive, and largely the province of a handful of platform companies. It is not the realistic target for most operating businesses, and treating it as the goal is a common and costly strategic error.
The useful move for almost every leadership team reading this is not to chase the fourth position. It is to move deliberately from the first to the third.
Why the Data Itself Is Not the Advantage
There is a common misreading of this argument worth correcting directly: the belief that simply accumulating more proprietary data is the moat. McKinsey’s research on AI-enabled competitive advantage is more precise than that. Its analysis of firms building durable AI moats identifies embeddedness, not data volume, as the operative mechanism: systems integrated directly into core workflows that learn from proprietary interaction data through structured feedback loops, becoming more valuable the longer they are used, while switching away from them becomes progressively more expensive for the organization.
Amazon’s advertising business, built on proprietary behavioral data from search activity, product views, and purchase history rather than on any single dataset, is the scale example McKinsey cites, and it reached roughly $68 billion in revenue in 2025 on the strength of that dynamic.
The mechanism, not the anecdote, is the transferable part. A loop only produces advantage when three conditions hold simultaneously: the system is embedded in a workflow that runs continuously, the corrections generated by that workflow are captured in a structured and retrievable form, and there is a discipline for feeding those corrections back in a way that measurably changes future output. Remove any one of the three and the loop collapses back into simple tool use.
The Loop in Practice
Consider a composite illustration, drawn from patterns common across mid-sized logistics operators rather than any single named company. A freight consolidation firm (call it Continental Freight Partners) deploys an AI system to draft customs documentation and route pricing quotes. In its first month, the system performs adequately but generically. It does not know that a particular port terminal in one of its markets routinely requires an extra clearance document for perishable goods, or that one long-standing client’s contract terms cap fuel surcharge pass-through at a level the standard pricing logic does not reflect.
Without a loop, an operations manager corrects these errors by hand, every time, indefinitely. The correction lives in that manager’s head and in that week’s email thread, and it evaporates the day she leaves the company.
With a loop, each correction is captured as a structured record: what the system produced, what a human changed, and why. Over eighteen months, this record becomes a private evaluation set that reflects Continental Freight Partners’ actual customers, actual regulatory environment, and actual commercial terms. It is not the generic logistics knowledge available to every competitor using the same underlying model. A competitor across town can subscribe to an identical AI platform tomorrow morning. It cannot subscribe to eighteen months of Continental Freight Partners’ captured operating judgment. That gap is the entire advantage, and it is precisely what is now separating leaders from laggards on every continent.
Why Six in Ten Companies Are Leaking This Value Today
This gap also explains a statistic that has troubled boards throughout 2025 and 2026. Boston Consulting Group’s research on the AI value gap found that roughly sixty percent of enterprises generate no material value from their AI investment despite meaningful spend, while only about five percent generate substantial value at scale.
McKinsey’s parallel survey work reports a similar pattern from a different angle: a majority of organizations see AI improve innovation and employee satisfaction in qualitative terms, but under forty percent see measurable impact on earnings, and where that impact appears, it is concentrated in a narrow set of functions rather than distributed across the business.
These figures are usually read as an adoption problem or an implementation problem. Read through the loop framework, they are neither. They are a capture problem. The sixty percent generating no material value are, almost by definition, running rented intelligence: real spend on real models, with no structured mechanism turning daily corrections into a private, improving asset.
The five percent generating substantial value are, on close inspection of the pattern BCG and McKinsey both describe, the firms whose AI use is embedded in a continuous workflow with a feedback discipline behind it, the third quadrant in the framework above rather than the fourth. They did not need proprietary infrastructure. They needed proprietary process.
What Building the Loop Actually Requires
Three disciplines separate firms that build a working loop from firms that do not, and none of them require a research budget.
The first is capture. Every correction a human makes to an AI-generated output (a rewritten clause, a rejected price, a reordered priority) needs to be logged in a structured format from day one, not retrofitted later once someone notices the value walking out the door. This record is discarded by default across much of the market today, because AI tools tend to get selected for output quality rather than for whether they expose an editable, logged interaction.
The second is evaluation. A private evaluation set, a running record of what “correct” looks like for this specific business built from real cases rather than generic benchmarks, becomes more valuable the longer it accumulates. Very few companies maintain one today, in part because it looks unglamorous next to a model demo, and in part because building it requires the operations team and the technology team to agree on what quality actually means in a specific workflow. That conversation is the strategic work. The model is not.
The third is governance over how corrections feed back into the system, so that the loop reflects the judgment of the firm’s best operators rather than its most vocal ones, and so that the record is auditable when regulators or clients ask how a decision was reached. This is the discipline most likely to be skipped under time pressure, and the one whose absence is hardest to correct retroactively.
None of this depends on the size of the model a company can afford. It depends on whether leadership treats the correction as disposable or as an asset.
The Position This Leaves Every Leadership Team In
The frontier model race will continue, and the companies building those models will continue to command outsized attention and outsized valuations. That race is not, for the overwhelming majority of businesses, the one that determines competitive outcome. The race that determines outcome is quieter, runs inside the firm rather than inside a research lab, and rewards operational discipline over compute budget.
A company can rent the same intelligence as every competitor in its market by close of business today. It can only own what it has learned. The firms that understand this distinction now, and build the discipline to act on it, will be difficult to catch by the time the rest of the market catches up to the argument. The firms that continue treating model selection as their AI strategy will discover, later than they should, that they spent several years renting a commodity while their competitors built an asset.
RELATED READING