The AI Competitive Advantage

Suppose your organization signs with an AI vendor this week, after careful selection. Somewhere else, your strongest competitor signs with the same vendor, on the same terms. In 1985 the economist Edwin Mansfield found that even technology firms built themselves reached their rivals within about a year¹. A licensed model has nothing to leak at all. It is offered to everyone. As AI Leaders vs AI Followers showed, the best models are now within a few points of one another, so a rival can match your model in weeks, by signing a contract.

That changes the strategic question. It is no longer “do we have AI?” It is “what have we built with AI that a competitor could not reproduce by buying the same thing?”

The model may be shared; the context is yours

The answer starts with a distinction that sounds simple and is often ignored. A foundation model supplies general capability: language, analysis, code and generation. It is necessary, and choosing it well still matters, because a weak model can spoil a good process. But it is shared. What is not shared is everything the model works inside: what your organization knows, how its work is arranged, the people who run that work and the trust customers place in it.

That distinction gives leaders a working test, which the rest of this chapter applies again and again. Call it the copy test: if our strongest competitor got exactly our AI model tomorrow, what would still be hard for them to copy? If the honest answer is “nothing”, the organization does not have an AI advantage. It has AI. AI Is Changing Everything showed how many organizations are in that position; the copy test explains why.

Commodity technology, uncommon use

The argument that a widely available technology cannot be an advantage is not new. In 2003 Nicholas Carr published “IT Doesn’t Matter” in Harvard Business Review; the magazine’s staff voted it the best article it published that year. Carr compared information technology with railroads and electric power. Each, he argued, gave early movers a brief advantage, then became a commodity input as it spread and its cost fell: essential to competing, but useless for standing out. His advice followed: spend less, and follow rather than lead².

Carr was right about the technology and, the evidence suggests, wrong about what companies could do with it. Five years later Andrew McAfee and Erik Brynjolfsson looked at what had happened to competition in American industries since the mid-1990s. If IT had become a commodity, it should have leveled the field. Instead, the gap between leaders and laggards widened, markets became more concentrated and the churn among top firms rose, most of all in the industries that spent most on IT. Their explanation was that enterprise systems let a firm take a process innovation it had invented and copy it across its whole organization with high fidelity. Rivals could buy the same software. They could not buy the leader’s process³.

Put the two together and you have the position this chapter takes on AI. The model is becoming a commodity input, and no executive should build a strategy on having one. But a commodity input, used in an uncommon way and spread reliably through a business, can widen the gap between firms rather than close it. The advantage moves from the technology to its use.

Why some things are hard to copy

Strategy researchers have a precise vocabulary for “hard to copy”. In 1991 Jay Barney argued that a firm can sustain an advantage only with resources that are valuable, rare, imperfectly imitable and not easily substituted. A licensed AI model fails the second test on the day it is signed, because it is not rare. The interesting question is the third: what makes a resource imperfectly imitable? Barney named three sources⁴.

A hotel makes the idea concrete. When it renovates its lobby, the hotel across the street can copy the look within a season: whatever a guest can photograph, a rival can copy. What it cannot copy is the history of years of guest relationships, the causal ambiguity of not knowing whether guests return for the concierge or the breakfast, and the social complexity of staff who hand over without a word. Few guests choose a hotel twice for its lobby alone.

In AI terms, the model is the lobby. The copy test is Barney’s question in executive form: point to the parts of an AI initiative that took years to build, that a rival could not reverse-engineer and that live in how people work together. Those are the candidates for advantage.

The AI advantage stack

It helps to see where those candidates sit. Think of AI advantage as a building with six floors. The model is the ground floor. Each floor above it is harder to copy than the one below.

Proprietary data sits just above the model. Many boards assume it is their moat. Sometimes it is; often it is incomplete, hard to use or not theirs to use for AI. Whether data is a real moat is a harder question than it looks, and Proprietary Data, AI Moats and Differentiation in Module 04 takes it up properly. Here it is enough to say that data a competitor could also obtain, or that no process ever uses, fails the copy test.

Domain knowledge is the next floor, and it is often underrated. A general model may know what a mortgage is. A lender knows its own credit exceptions, its risk appetite and what its supervisors expect. An engineering firm knows the failure modes that never reached the manual. That knowledge is mostly unwritten, which is exactly why it is hard to copy, and also why it is fragile. If it lives in a few senior heads, a retirement is a strategic risk. Pouring every document into a system does not solve the problem: volume is not expertise.

The upper floors, workflow, capability and the customer, are where the copy test is most often passed or failed, so they get sections of their own.

Inside the workflow, not beside it

Executives tend to underestimate workflow integration because it sounds like IT. It is in fact how the business runs. Picture the same model used in two ways: in a chat tab that an employee opens when they remember to, or wired into the systems where the work happens, such as the case record, the review queue and the approval step.

In the tab, the value depends on each person’s initiative and disappears when they stop. In the workflow, the value is part of the process and spreads to everyone who touches it, which is the mechanism McAfee and Brynjolfsson found in the earlier wave of enterprise IT. The copy test is decisive here. A competitor can buy the same model and the same chat product within a week. It cannot easily copy the redesign of your hand-offs, your exception paths and the points where a person still decides. If the AI cannot see the work, and the work cannot see the AI, there is no integration, only a side conversation. For AI, this is still an argument carried over from the evidence on earlier enterprise IT. It is plausible, but it has not yet been shown in measured results for AI itself.

Why the layers compound

The stack would be a useful checklist even if each layer stood alone. Its real power is that the layers multiply. Michael Porter made the point about strategy in general in 1996. Improvements in operational effectiveness, doing the same things better, are copied quickly, because best practices spread. What is hard to copy is a set of activities that fit together and reinforce one another. Porter’s arithmetic: if a rival has a 90 percent chance of matching any single activity, its chance of matching two is 81 percent, and four, 66 percent⁵.

The numbers are illustrative; nobody can measure the true odds of copying a layer, and the layers are not independent. The direction is the point. AI productivity tools used alone are operational effectiveness: worth having, and copied within a year, as Mansfield’s executives would have predicted. AI woven through knowledge, workflow, capability and the customer offer is a system, and systems are slow to copy. A rival that imitates one layer gains little unless it also matches the rest.

Capability and the customer

Organizational capability is the layer that makes the others last. It is not a training day. It is a loop that turns every week: people use AI on real work, someone measures the quality of the results, managers coach the exceptions and judgment calls, and what was learned is taught to the next person. Each turn leaves something behind. Marco Iansiti and Karim Lakhani, who studied firms that compete on AI, argue that the winners rebuild their operating model around it rather than adding it to an old one⁶. That kind of capability is socially complex in Barney’s sense, so it is slow to copy. It also decays: it leaves with people if nothing is written down, and it withers when it lives only in a central team the line business never touches. How leaders build these habits was the subject of AI Leaders vs AI Followers.

Customer experience is the top floor, and distance to the customer is the simplest way to judge it.

AI that drafts and summarizes inside the company is productivity: often worth doing, and rarely an advantage on its own. At the service moment, customers begin to notice, for example when a request is resolved in one exchange because the answer already knows their history. In the offer itself, AI can change what customers buy or why they stay. That is the most strategic rung and the one that needs the most governance, because mistakes land on customers. It is also the least proven: evidence that AI in the offer creates a lasting advantage is still thin, so treat it as a possibility to test, not a demonstrated result. Not every organization should rebuild its offer this year. The useful question is simpler: is AI only making yesterday’s offer cheaper, or is it beginning to change the offer?

Story: the best model, and a market that barely moved

On 7 February 2023 Microsoft relaunched its Bing search engine around a chat assistant. It ran on what Microsoft called a next-generation OpenAI model, “more powerful than ChatGPT and customized specifically for search”⁷. For a few weeks Microsoft had something its far larger rival did not: early use of the strongest model on the market. Its chief executive, Satya Nadella, told The Verge that he hoped Google would come out and show it could dance, and that he wanted people to know Microsoft had made it do so⁸.

The lead on the model was short. On 14 March Microsoft confirmed that the new Bing ran on GPT-4⁹. The same day OpenAI made GPT-4 available to paying ChatGPT subscribers and opened a waitlist for any developer who wanted to build on it¹⁰. A week later Google opened its own chatbot, Bard, to users in the United States and the United Kingdom¹¹, and in May it began testing generative AI answers inside Google Search itself¹².

By the end of the year the market had barely moved. StatCounter’s figures, as reported by The Register, put Bing’s global share of search across all devices at 2.81 percent in February 2023 and 3.37 percent in December, while Google’s slipped from 93.37 to 91.62 percent¹³. Bloomberg, using the same source, reported that Bing ended 2023 with 3.4 percent of the global market, up less than one point¹⁴. Bing did gain something. On desktop computers its share rose from about 8 to about 10.5 percent, and Bloomberg reported that people were spending more time on it. But on phones Google kept more than 95 percent.

Neither outlet put the outcome down to the model. Both pointed to layers above it. The Register noted that the Bing chatbot was at first limited to Microsoft’s own Edge browser, which had about 5 percent of the browser market¹³. Bloomberg pointed to Google’s default status on Apple devices and to how entrenched people’s search habits are¹⁴. The court record shows what that default position costs. In United States v. Google, a federal judge found in August 2024 that Google paid its browser, device and wireless partners a total of 26.3 billion dollars in revenue share in 2021 under the contracts that made it their default search engine, almost four times all its other search-related costs combined¹⁵.

Read the case against the stack. Microsoft’s lead was on the ground floor, and that is the floor a rival could match, by buying, as any developer could from March, or by building, as Google did. The layers that decide where people search sat higher. Where the search box sits on a phone or in a browser is workflow integration in consumer form. The habits and trust of the people using it are the customer layer. A better model did not move either of them in a year. This is not a verdict on Microsoft’s wider AI business, which the case does not assess. It is the copy test seen from the challenger’s side: the asset that was quickest to obtain was also the quickest for a rival to match.

What this means for leaders

Three lessons follow. First, treat the model as infrastructure and choose it on fit, cost and reliability, but never present it to a board as the advantage. Second, invest where the copy test points: usually domain knowledge and workflow first, because they are scarce, and capability alongside, because it is what makes the other layers last. Third, judge an AI portfolio as a system. A dozen separate productivity tools, each easy to copy, add up to less than a few layers that reinforce one another around one process customers care about. How to turn this into a full strategy, with tests for durability, is the work of AI and Competitive Advantage in Module 04.

Check yourself

1. Our AI vendor gives us a lasting competitive advantage.MythFact
2. When a technology becomes a commodity, it can no longer widen the gap between firms.MythFact
3. A resource is hard to copy when it was built over time, rivals cannot see what drives it, and it lives in relationships.MythFact
4. The most advanced model always wins.MythFact
5. A rival can buy your model much faster than it can copy your workflow.MythFact
6. Several AI layers that reinforce one another are harder to copy than any one layer.MythFact

Reflection: what would survive the copy test?

What comes next

The layers above the model are slow to copy because they are slow to build, and that cuts both ways. If advantage takes time to assemble, how much time is there? AI is spreading through organizations quickly, yet that speed is not the same as transformation. The next chapter, The Speed of AI Adoption, looks at how fast the clock is really running.