Insights · Strategy

AI competitive advantage: what is left when everyone has the same tools

Every board pack now has an AI slide, and nearly all of them make the same claim: this will be our advantage. The evidence points the other way. The gains land where the work was most replicable, they reach every competitor inside the same eighteen months, and most of them leave again through price. What survives is what AI needs and cannot manufacture.

A competitive advantage is the ability to earn returns above the industry average for longer than it takes rivals to copy the move that produced them. That second clause is the whole of this article. The AI investment case that reaches most boards is built on a productivity number, and the number is usually real. Missing is the question that decides whether it is worth anything: what happens to it once every competitor has the same one. Operational gains become advantages only when rivals cannot obtain them, and the defining property of a frontier model is that anyone can obtain it this quarter with a corporate card.

The size of your gain measures how copyable the work was

The two cleanest field experiments agree, and what they agree on is uncomfortable. Brynjolfsson, Li and Raymond followed 5,179 customer support agents through a staggered rollout and found productivity up 14 per cent on average: a 34 per cent improvement for novice and low-skilled workers and, in their words, minimal impact on experienced and highly skilled workers. Their explanation is the interesting part. The model disseminates the best practices of the more able workers. Dell'Acqua and colleagues found the same shape across 758 consultants at Harvard and BCG, where below-average performers improved 43 per cent against 17 per cent for the above-average group.

Read those as statements about your business, not about the technology. The tool takes what your best people know and hands it to everyone else. Inside one firm that is excellent and worth paying for. Across a market it means the same thing is happening to your competitor's weakest people on the same timetable. The distribution of capability narrows, and a narrower distribution is the opposite of an advantage.

The corollary belongs in front of a board. Where AI lifts a process 40 per cent, that process was codifiable, which means it was always copyable and you were being paid for something a machine could learn from transcripts. Where it lifts nothing, the work resisted codification. That is where your advantage lived the whole time.

Why the gain stops before the P&L

The aggregate data says this is already happening. McKinsey's State of AI survey, published on 25 August 2026 across 1,719 respondents in 97 nations, found nearly nine in ten organisations using AI regularly in at least one function and 44 per cent scaling it across the enterprise. In the same survey 37 per cent attributed any EBIT impact at all to AI, a share essentially unchanged on the year, and only 6 per cent attributed 5 per cent or more. McKinsey's own summary is blunt: conviction is growing faster than the returns.

The standard reading of that gap is an execution problem. That is partly fair and mostly insufficient, because if the issue were implementation maturity the share would climb as deployment widens, and it has not. A cost reduction that reaches your whole category at once is what economists call a common cost shock, and those behave differently from ones that hit a single firm. The cost pass-through report RBB Economics prepared for the Office of Fair Trading in 2014 puts it plainly: the extent of firm-specific cost pass-through is typically less than industry-wide cost pass-through. A firm that alone gets cheaper is disciplined by rivals whose costs did not move, so it holds price and keeps the margin. When the saving is universal, nothing holds the price up and competition hands it to the customer.

The arithmetic, on your own numbers

Take a plausible structure: revenue 100, gross margin 40, SG&A 30, EBIT 10. A serious programme removes 15 per cent of the SG&A base, which is 4.5 points of revenue, and costs 1.2 points a year once licences, integration, data engineering, evaluation and governance are counted. Viewed in isolation that is 3.3 points of EBIT and an easy approval.

Now assume what is actually true, that your three main competitors run the same programme in the same period. Apply a 70 per cent pass-through on the common shock, well inside the range the pass-through literature reports. Price falls 3.15 points. You keep 1.35 of the 4.5, spend 1.2 on the stack, and 0.15 points reach EBIT. Ten becomes 10.15: removing 15 per cent of your overhead moved operating profit by one and a half per cent. Run the same programme as the only firm that has it, where pass-through on a firm-specific saving is far lower, say 20 per cent, and 2.4 points reach EBIT instead. The operation is identical. The outcome differs by a factor of sixteen, and the only variable is whether your competitors could run it too.

The third case is the one that makes this a decision rather than a debate. Sit it out while the category does not, and the price still falls 3.15 points, because price is set by the market and not by your cost base, except that you took no cost out to meet it. EBIT goes from 10 to 6.85, a fall of roughly a third, earned by spending nothing. The programme is therefore mandatory and, on its own, unprofitable. Barney's resource-based view has a term for that condition: a resource that is valuable but not rare produces competitive parity, not advantage. Parity has to be bought, because the alternative is disadvantage. Nobody should be told at the investment committee that buying it is a strategy.

A capability your competitor can buy on Tuesday is a cost line, not a moat.

Strategy already had a name for this

None of this is new, which is the most useful thing about it. Porter set out the mechanism in 1996 and the sentences have aged well: competition on operational effectiveness shifts the productivity frontier outward, raising the bar for everyone, but although it produces absolute improvement, it leads to relative improvement for no one. He added the line that describes most AI programmes as currently scoped, that the more benchmarking companies do, the more they look alike. Nicholas Carr made the same argument about enterprise IT in 2003 and was told he had missed the point. He had not. This wave differs in speed, not in structure, and speed matters because it shortens the window in which an early mover keeps anything.

What AI needs and cannot manufacture

Put the model subscription through the four tests of the resource-based view. Valuable: clearly. Rare: not remotely. Costly to imitate: the opposite, imitation is a procurement decision. Organised to capture value: possibly, and that is the only one of the four you control. Three failures out of four is the definition of parity, and it points at where to look instead: at the inputs the model needs and cannot generate, and at the assets that decide whether its output ever reaches a buyer.

Three things get more valuable, not less. First, proprietary data with a feedback loop: the outcomes of your own decisions, which no competitor can reconstruct and no vendor can sell them. In Helmer's terms that is a cornered resource, and the one power a model amplifies rather than erodes. Second, distribution and memory. When the answer arrives as one model response rather than ten links, being the brand that gets named is the whole competition, which is the argument in how brands get cited by ChatGPT and Perplexity and, on the asset side, in distinctive beats different. Third, switching cost and process power: an AI embedded in a workflow holding your data and your organisation's accumulated decisions is a different object from the same model behind a login.

A fourth survives untouched, and smaller firms should note it. Counter-positioning never depended on capability, so better tools for the incumbent change nothing: the constraint was always their own profit pool. We set out the test in challenger go-to-market.

It narrows people and widens firms

The claim that AI levels the playing field for smaller companies is true of individual workers and false of firms, and the two get confused constantly. Everything above the worker level is fixed cost: data engineering, evaluation harnesses, security and legal review, governance, retraining, the cost of being wrong in public. Fixed costs favour whoever has the volume to spread them, which is the plain definition of a scale economy. The adoption data shows the split: the US Census Bureau's Business Trends and Outlook Survey, published on 26 May 2026, put overall business AI use at 19.8 per cent, against 37 per cent for firms of 250 or more employees and under 20 per cent for those below 20, where the number has not moved. Within a firm the technology compresses the gap between best and worst. Between firms it widens the gap between those who can amortise the overhead and those who cannot, and only one of those two facts is in the vendor deck.

How we run it

Three questions, in order, and none is about which model to use. Brand X-Ray inventories what you actually own and sorts it against the four tests, separating the line items that buy parity from the handful that are genuinely rare. Brand Wargame then runs the symmetric case, the one nobody models: not what we do with AI, but what this category's P&L looks like in two years when all four of us have it, who keeps the saving, and which of today's advantages survives a competitor who is suddenly as fast as we are. Cashstream puts the pass-through assumption on the page and splits the budget into the part defending parity and the part buying something rare. For the wider frame, our strategic advisory page sets out how we work.