AI Has No Moat. Your Organisation Might.
Edition 16 - Why boards should stop treating AI as a source of advantage and start asking what the organisation must own for any advantage to survive competitors buying the same tools.
A board can approve a confident AI strategy and still be funding the slow erosion of its own competitive position. The reason is awkward to say in a boardroom that has just signed off on an eight-figure programme, but it is the right place to start. The large language model your company runs is almost certainly licensed from one of a handful of vendors, trained on broadly the same public corpus as everyone else’s, served from rented cloud infrastructure shared with your competitors, and prompted using techniques that are documented openly and copied within weeks. None of that is a criticism of the technology. It is an observation about who can obtain it, which is everyone. A capability available to every competitor on similar terms cannot, by definition, be the thing that sets you apart.
The most resourced company in the field reached a version of this conclusion three years ago. In May 2023, an internal memo by a Google engineer, leaked and published under the title “We Have No Moat, And Neither Does OpenAI”, argued that neither the search giant nor the leading lab held a defensible advantage in models, because open alternatives were closing the gap faster than either could pull away. The memo was about the model-makers and their own arms race, but it implies a sharper question for everyone else: if the firms building the models have no moat, where does that leave the companies merely licensing them? That points away from the technology itself and towards everything an organisation can build around it, which is exactly where competitive strategy has been looking for the past 40 years.
That tradition has a settled test. Jay Barney’s resource-based view, which has shaped strategy teaching since the early 1990s, holds that a resource produces a durable advantage only when it is valuable, rare, hard to imitate, and embedded in how the organisation actually works. Off-the-shelf AI is plainly valuable. It is rapidly ceasing to be rare, and it is trivial to imitate by the simple act of signing the same vendor contract. We have watched this film before. In the 1990s, large enterprises spent fortunes installing the same ERP systems from the same vendors, and a decade of research found that the software itself conferred no advantage; the firms that pulled ahead were the ones that rebuilt their processes around it while their rivals merely switched the system on. The board’s real question, then, is not which tools to deploy but what the organisation must own, so that any advantage survives competitors doing precisely the same thing.
Proprietary data is a moat only when it is genuinely yours
Data is the asset most boards reach for when they want to believe their AI is defensible, and it is the one most often misunderstood. The instinct is to equate advantage with volume, on the assumption that whoever holds the most data wins. The economics rarely cooperate. Beyond a threshold, additional data delivers sharply diminishing returns, and most corporate data is neither unique nor especially relevant to the decisions that move the business. The resource-based view is unsentimental here: data confers advantage only when it is rare and hard to replicate, which means unique to your operations, continuously refreshed by activity that competitors cannot observe, and legally defensible.
John Deere is the clearest current illustration of what the real version looks like. Its Operations Centre now has over 370 million acres enrolled, and every time a connected machine plants, harvests, or applies inputs, that data flows back and sharpens the next recommendation, creating a self-reinforcing advantage that more acres only deepens. A rival can buy comparable hardware and license a comparable model. What it cannot quickly buy is decades of customer relationships, an installed base of connected machines, and an agronomic dataset built across hundreds of millions of acres. Even here, the moat is contestable rather than permanent; right-to-repair litigation and the spread of open models could force the ecosystem open, and a serious analyst will tell you the durability is an open question rather than a settled fact. That nuance is the point. A board should be able to say which of its data assets pass the Deere test and which are simply volume the company has mistaken for an asset, because the second kind buys parity at best.
The profit goes to whoever owns the assets AI plugs into
The most useful strategy lens for AI is also the most neglected in boardrooms, and it predates the technology by 40 years. In 1986, David Teece asked why the firms that pioneer an innovation so often fail to profit from it, and answered that the returns flow not to the inventor but to whoever controls the complementary assets the innovation needs to reach a market: distribution, brand, manufacturing scale, service networks, regulatory position, and installed base. His own example has aged well. EMI, the British music company, invented the CT scanner and its creators won a Nobel Prize for the underlying work, but lost the market within a decade to General Electric and others who already owned the hospital sales forces, service operations, and manufacturing depth required for medical imaging. The invention was EMI’s; the returns went to the firms that already owned the route into hospitals.
The same pattern is visible in the AI market today, and boards can read it directly off the capital flows. The foundation-model labs created the innovation, yet a growing body of analysis argues that durable value is migrating toward incumbents who already control distribution and proprietary data. As foundational models commoditise, value is expected to spread to companies that control market distribution and proprietary data, and many business applications do not need the most advanced model to work well. Application-layer companies that lack proprietary data, deep workflow integration, or genuine distribution are, in the market’s blunt phrase, just wrappers around someone else’s model. For an established business, this is the optimistic reading of commoditisation. The thing you spent decades building, the distribution and the customer relationships and the regulated position, is exactly the complementary asset that captures the value AI creates. The board’s task is to point AI investment at those assets rather than letting it compete on ground where the company owns nothing.
What competitors can see, they still cannot copy
The strongest moat is also the least visible, which is why it is so often underfunded. When AI is woven deep into redesigned processes, incentive structures, and the tacit know-how of the people who run them, a competitor cannot copy it even while watching your results, because it cannot see why the results occur. Strategists call this causal ambiguity, the condition Lippman and Rumelt identified in which the link between what a firm does and how well it performs is genuinely unclear to outsiders, and sometimes to insiders. It is the difference between a recipe a rival can photograph and a kitchen a rival cannot reproduce.
History gives this its sharpest illustration. When factories first electrified, productivity barely moved for decades. Paul David’s study of the electric dynamo showed that the gains arrived only once firms stopped bolting electric motors onto layouts designed for steam, with their central drive shafts and stacked floors, and rebuilt the factory around distributed power: single-storey plants, machines arranged by workflow, work reorganised entirely. The technology was available to all within a few years, but the reorganisation took a generation, and that is where the advantage lived. Erik Brynjolfsson has spent years arguing that information technology follows the same logic, that the value comes from large, slow, complementary investments in process and human capital that lag the technology by years.
This is what the failure data is actually telling boards, once the headline is read properly. MIT’s GenAI Divide study found that around 95% of generative AI pilots delivered no measurable profit-and-loss impact despite billions invested. The lesson is not that the models are weak. It is that buying the model is the cheap, copyable part, and almost no one makes the expensive, slow, hard-to-imitate organisational investment around it. The misallocation is visible in the spending itself. More than half of corporate AI budgets in 2025 went to high-visibility sales and marketing pilots with low returns, while the real gains sat in unglamorous back-office work that few boards scrutinised. A board that funds the tool and skips the embedding is paying for the photograph of the kitchen.
The only durable edge is the ability to keep re-winning
Even a well-embedded advantage decays, because the frontier does not sit still. A workflow redesigned around last year’s model can be overtaken by a competitor that reconfigures around this year’s model, and agentic systems may let rivals replicate established workflows faster than embedding can be once protected against. The implication is not that moats are pointless, but that the deepest moat is a capability rather than a position. Teece, Pisano and Shuen named it dynamic capability in 1997: the organisational ability to sense what is changing, seize the opportunities it creates, and reconfigure the business repeatedly as conditions shift. The advantage lies in being built to keep winning as the field moves, rather than in any single deployment that happens to be ahead today.
Kodak is the standing warning. It invented the digital camera in 1975 and held the patents, but it could not reconfigure a business model built on film and chemicals, and filed for bankruptcy protection in 2012 while the technology it had pioneered remade the industry around it. The patents were real; the capability to act on them was not. For a board, dynamic capability is harder to assess than a data set or a distribution network because it shows up as a pattern over time rather than a line on a balance sheet, but it is assessable: how quickly has the organisation moved from one model generation to the next, how much of last year’s AI work has it been willing to discard, how fast can it redeploy people and budget when the frontier moves again. A programme that produces one impressive deployment and then ossifies is not building this capability, whatever its current returns suggest.
Parity spend, and advantage spend
Most AI investment reaches the board wrapped in an efficiency case: it will make the company faster, leaner, and cheaper. The trouble is that efficiency that every competitor can also buy is parity, not advantage. It is the cost of staying in the game, necessary and unavoidable, and a board should fund it without illusions about what it is. The error is mistaking it for a source of advantage and pricing it as one. The four mechanisms above provide directors with a cleaner way to test every AI proposal that management brings forward. Split the spend into two categories. Parity spend keeps the company in step with rivals and builds no moat; it should be approved based on cost discipline and treated as table stakes. Advantage spend builds or feeds something a competitor cannot easily copy, whether unique data, a complementary asset the company already owns, deep workflow embedding, or the capability to reconfigure faster than the field, and it should be held to a different and higher standard because it is the only spend that can actually change the company’s position.
This distinction is becoming a financial reality, not merely an analytical convenience. Citi has identified a credit-spread penalty for companies the market classes as AI “adopters”, heavy spenders without demonstrable returns, relative to “enablers” who can show conversion, meaning the debt market is already charging a premium for spending without proof. When the cost of borrowing starts to reflect whether AI spend is converting, the parity-versus-advantage question moves from the strategy offsite to the audit and finance committees. A board that cannot tell which of its AI budget is buying parity and which is buying advantage is, increasingly, a board the capital markets will price accordingly.
Questions a board should be asking
The test is only useful if it changes what directors demand. For each material AI investment, a board should require management to answer which of the four mechanisms it is meant to strengthen, and treat “none of them, it is an efficiency play” as an honest answer that puts the spend firmly in the parity column. It should ask whether any data described as a competitive asset is owned by the company and can be legally defended, and whether a competitor could assemble something equivalent within a year. It should ask whether AI is being pointed at the complementary assets the company already controls or competing on ground it does not own. It should ask what the programme has discarded in the last twelve months, as a direct measure of reconfiguration rather than accumulation. And it should require that the parity-versus-advantage split appear in the AI investment papers themselves, so that management makes the distinction before the board has to tease it out.
What this leaves the board with
The model has no moat and never will, because the next competitor can license the same one tomorrow. The organisation might have a moat, but only where it owns the data, holds the complementary assets, has done the slow work of embedding, or has built the capability to keep reconfiguring as the frontier moves. Each of those is expensive, difficult, and specific to the company, which is exactly why they are defensible and exactly why off-the-shelf AI is not. The board’s responsibility is to fund the advantage and to stop dressing up parity as an edge. The companies that struggle through the next few years will not be the ones that spent too little on AI. They will be the ones who spent confidently on the part that everyone else could also buy, and called it a strategy.
If this is the kind of board-level analysis you want more of, subscribe to AI in the Boardroom. I write for directors, executives, and advisers who need to separate the AI spending that builds a durable advantage from the spending that merely keeps a company level with its competitors. Future editions will continue to return to the strategy, governance, and value-creation questions that boards cannot sensibly delegate.