If every competitor has the same AI, where does the advantage come from?

From whatever a competitor cannot buy on the same terms you did. A model subscription is sold to every firm at list price, so under the oldest test in strategy it fails at the first hurdle: valuable, and not rare. What still varies is everything the subscription touches. The 2026 US Census AI supplement puts firm-level use at 18 per cent, and among firms that do use it, 57 per cent have it in three or fewer business functions. So the premise of the question is not yet true, and it will not be true evenly. The durable positions are proprietary data, a redesigned process, and the judgement to know when the output is wrong. All three are slow, and none of them arrives with the licence.

Nobody has the same AI yet, and the gap is measurable#

Bonney and colleagues at the Census Bureau's Center for Economic Studies, using the 2026 AI supplement to the US Census Bureau's Business Trends and Outlook Survey, measured diffusion at three layers: whether the firm uses AI, where in the business it uses it, and whether workers use it in their tasks. Over the reference period of November 2025 to January 2026, 18 per cent of firms used AI in a business function, rising to 32 per cent on an employment-weighted basis, with adoption expected to reach 22 per cent within six months.

The distribution is the part that answers the question. Use rates reach 50 to 60 per cent for very large firms in Information, Professional Services and Finance, and 60 to 70 per cent on an employment-weighted basis in those same cells. Among firms that have adopted, scope is narrow: 57 per cent integrate AI in three or fewer business functions, most often Sales and Marketing at 52 per cent, Strategy and Business Development at 45 per cent and IT at 41 per cent. Worker-task use runs at 23 per cent of firms, 41 per cent employment-weighted, and 65 per cent of firms limit task use to three or fewer tasks. Two-thirds of users, 66 per cent, use AI solely to augment tasks, and AI-related employment decreases appear in 2 per cent of firms.

Jeffrey Allen at the Federal Reserve Board, writing in April 2026, put three independent measures side by side and found them 60 points apart. The Census BTOS reports about 18 per cent of firms at the end of 2025. The Atlanta Fed's Survey of Business Uncertainty, which asks senior leaders, produces an employment-weighted rate of about 78 per cent. The Real-Time Population Survey, which asks individuals, gives about 41 per cent of the workforce using generative AI at work. Allen's explanation is not that one of them is wrong:

The biggest driver of variation in these estimates likely relates to differences in sampling distributions and units of analysis, but question framing, the materiality of reported usage, information asymmetries between different target respondents, and social desirability bias may play a role as well.

He also names the direction of the pressure on the highest number: senior leaders "may face pressure to report AI usage as an efficiency initiative", which puts upward pressure on estimates that target corporate leaders. Anyone reasoning about competitive parity from a headline adoption figure is reasoning from a number whose meaning depends entirely on who was asked. See how to measure AI adoption properly.

Barney's four tests, applied to a subscription#

The framework for this question predates the technology by thirty-five years. Jay Barney, in the Journal of Management in March 1991, set out four empirical indicators of whether a firm resource can generate sustained competitive advantage: value, rareness, imitability and substitutability. His stated assumptions were that strategic resources are heterogeneously distributed across firms and that those differences are stable over time.

Run a commercial model licence through the four. It is valuable, on any reasonable reading of the productivity evidence. It is not rare, because the vendor's business model depends on it not being rare. It is trivially imitable, since imitation consists of entering a card number. It is substitutable by three or four near-equivalent products. One of four is a cost floor rather than a position.

The same four tests treat the surrounding assets very differently. Proprietary operational data, accumulated over years and specific to your customers, is valuable, rare and not purchasable. A process redesigned around what the model is actually good at is imitable in principle and slow in practice, because copying it requires knowing which parts mattered. The capability to tell a plausible wrong answer from a right one is the least imitable of the three, because it is held by people and rebuilt only by practice. This research has a name for the organisational version of that asset: organisational capability, and for what happens when it is allowed to erode, capability debt.

Same model, opposite results#

The strongest experimental evidence that identical access produces non-identical outcomes comes from Dell'Acqua and colleagues, who gave 758 BCG consultants the same GPT-4 and two sets of tasks. Inside the model's competence, assisted consultants were substantially better and faster. On a task placed just outside it, they performed worse than consultants with no AI at all. The boundary is invisible from inside the conversation, which is the argument of the jagged frontier. Two firms buying the same licence and deploying it against different task mixes will get different signs, not merely different sizes.

Where the tool does move the average, it tends to compress rather than separate. Brynjolfsson, Li and Raymond, tracking a staggered rollout across 5,172 customer support agents, found resolutions per hour up about 15 per cent overall, with the lowest skill quintile gaining 36 per cent and the most skilled seeing no significant change. Dell'Acqua's later field experiment with 776 professionals at Procter and Gamble found something adjacent: without AI, research and development staff proposed technical solutions and commercial staff proposed commercial ones, while professionals using AI produced balanced solutions regardless of background. The functional signature of the person disappeared into the output.

Compression inside a firm is a good thing for that firm's floor. Compression across an industry is the thing that removes the differential. If the same tool lifts your weakest performers and leaves your strongest untouched, and it does the same for your competitor, the relative position is unchanged and the cost base is higher for both. That is a rational purchase and it is not a strategy. The same convergence shows up in output itself: see does AI make everyone think alike.

The complements are the slow part#

Brynjolfsson, Rock and Syverson, in the American Economic Journal: Macroeconomics in January 2021, modelled why general purpose technologies show up late in the productivity statistics. The technology requires large complementary investments that are intangible and badly captured in national accounts, so measured productivity is understated in the early years and overstated later, when the intangibles are harvested. Adjusting for intangibles tied to computer hardware and software, they put the US total factor productivity level 15.9 per cent higher than official measures by the end of 2017.

That is a measurement paper, and it carries a strategic reading. The intangible complements are the advantage. Retrained staff, rewritten processes, cleaned data, new controls and the tacit knowledge of which tasks to hand over: these are exactly the assets that pass Barney's rareness and imitability tests, and exactly the ones that do not appear on a licence invoice. The firm that buys the model and skips the complements has bought the part everyone can buy.

Humlum and Vestergaard's Danish administrative data supports the sequencing. Across roughly 25,000 workers in 7,000 workplaces, two years after ChatGPT, they found precise null effects on earnings and hours, ruling out effects larger than 2 per cent, alongside substantial task reorganisation and new tasks in AI oversight and integration. The work changed first. On current measurement, the money had not.

Bacon's maxim, and where it broke#

Rahim Hirji has been arguing a version of this since six weeks after ChatGPT launched. "Being Unique in the Face of ChatGPT" (2023) took the position that knowledge itself had been commoditised, so what differentiates a person moves to uniqueness, creativity and adaptability. "Knowledge Is No Longer Power" (2025) put the organisational form of it: Bacon's maxim breaks when knowledge becomes instantly and universally available, and advantage moves from holding knowledge to judgement about which questions are worth asking. "Why curiosity is the only moat left" (2025) named the mechanism, the reflex to accept the first plausible answer, and the dividing line between people who ask a second question and people who have outsourced questioning entirely.

"Rules Before Tools" (2025) is the one that speaks directly to a board. Its argument is that chasing each model release substitutes for strategy, and that the advantage sits in the unglamorous work: redesigned processes, a clean data backbone, named accountable owners, guardrails. Read against Brynjolfsson, Rock and Syverson, that essay is a restatement of the intangible complements argument in operational language, written before the AI capital expenditure debate reached its current volume.

Two attribution notes, because this estate keeps them straight. The resource-based view is Barney's and the four tests are his wording. "Algorithmic drift" is Hirji's coinage. Capability debt has no dated first publication under his name and no claim of first use is made for it here; independent prior use exists in Rohde (arXiv 2605.27399, 23 March 2026).

Four questions that separate a cost floor from a position#

• Could a competitor buy this tomorrow at list price? If yes, it is a cost floor. Budget it as one, and stop presenting it to the board as strategy. Usage theatre is what happens when the licence count becomes the metric.
• What does this tool touch that only we have? Proprietary data, a customer relationship, a regulated licence to operate, an installed base, a physical asset. The model is the multiplier and the multiplicand is yours or it is nobody's.
• Which of our processes did we redesign, and which did we merely accelerate? Acceleration is copyable within a quarter. Redesign requires knowing which steps existed for a reason, and that knowledge sits with people who have done the work.
• Who here can still tell a good answer from a plausible one, and how would we know if that stopped being true? The verification capability is the least imitable asset in the list and the one no procurement process tracks. A capability audit is how it gets measured.

A fifth question is worth asking privately. If your answer to all four is thin, the strategic move may be to spend less on the tool and more on the complements, which is the opposite of what the market currently rewards a chief executive for announcing.

Where the vendor relationship becomes the exposure#

A capability rented from a third party is also a dependency on that party's pricing, roadmap and continued existence. Firms that build their differentiation inside a vendor's abstraction discover at renewal that the switching cost is the advantage, and that it belongs to the vendor. Two pages here work through the practical form of that: preserving capability across vendors and deployment is not a ratchet.

There is a second-order version. If every firm in a sector routes its judgement through two or three foundation models, the sector's collective error becomes correlated. Acemoglu, Kong and Ozdaglar model the extreme case formally, a knowledge-collapse steady state in which general knowledge vanishes despite high-quality personalised advice, and they state plainly that it is a theoretical model with no empirical estimation. That makes it a reason to keep an independent read on your own market rather than a forecast to plan against.

What a cross-section of firms cannot tell you#

It does not claim that AI confers no advantage. The Census paper reports a positive correlation between firm commercial performance and the breadth of AI integration, holding across functional deployment, task-level use and operational investment. Correlation in a cross-section of firms cannot tell you which way the causation runs. Better-run firms adopt more, and adopting more may make firms better run; the data cannot separate those.

It does not claim the diffusion numbers are stable. Allen documents that the Census Bureau broadened the BTOS question in November 2025 from use "in producing goods or services" to use "in any of its business functions", which moved the level, and he notes the "do not know" rate was 10 to 11 per cent of respondents. Any figure on this page is a measurement of a moving thing taken with an instrument that changed.

It does not claim Barney's framework settles the case. The resource-based view is a theory of advantage with a large critical literature, and applying it to a technology that is three years into commercial diffusion is an argument rather than a finding. The empirical claims here are the Census diffusion figures, the two Dell'Acqua experiments, the Brynjolfsson support-centre study, the J-curve estimate and the Danish nulls. The strategy reading of them is mine.

Key sources

• Bonney, K. et al. (2026). The Microstructure of AI Diffusion: Evidence from Firms, Business Functions, and Worker Tasks. US Census Bureau, CES Working Paper 26-25.
• Allen, J. S. (2026). Monitoring AI Adoption in the U.S. Economy. FEDS Notes, Board of Governors of the Federal Reserve System, 3 April 2026.
• Barney, J. (1991). Firm Resources and Sustained Competitive Advantage. Journal of Management, 17(1), 99-120.
• Brynjolfsson, E., Rock, D. and Syverson, C. (2021). The Productivity J-Curve: How Intangibles Complement General Purpose Technologies. American Economic Journal: Macroeconomics, 13(1), 333-72.
• Dell'Acqua, F. et al. (2023). Navigating the Jagged Technological Frontier.
• Dell'Acqua, F. et al. (2025). The Cybernetic Teammate.
• Brynjolfsson, E., Li, D. and Raymond, L. (2023). Generative AI at Work.
• Humlum, A. and Vestergaard, E. (2025). Still Waters, Rapid Currents.
• Acemoglu, D., Kong, D. and Ozdaglar, A. (2026). AI, Human Cognition and Knowledge Collapse.

Related SuperSkills research#

On measurement, how to measure AI adoption properly, the AI readiness lie and the most quoted AI statistics, checked. On the distributional question, who captures the productivity gains. On the organisation, the shape of the organisation after AI and keeping expertise in an organisation. On the assets that are hard to copy, tacit knowledge and capability debt.

About this research#

Rahim Hirji is the author of SuperSkills (Kogan Page, 2026), keynote speaker on AI and human capability, and founder of The SuperSkills Intelligence Company. The Census working paper abstract, the Federal Reserve note, the Barney abstract and bibliographic record, and the J-curve abstract and citation were each read at source and every figure on this page was checked against them. No consultancy estimate of AI's contribution to enterprise value is used here, because none of the ones in circulation publishes a method that can be checked.

How this research works  ·  Reviewed quarterly  ·  Found an error? Tell me and it is corrected on the page.