AI and Competitive Advantage

In July 1985 Berkshire Hathaway closed the textile mills it had owned for 21 years. In his next letter to shareholders, Warren Buffett explained why. Year after year his managers had brought him proposals for new, more efficient equipment, and each one looked sound. Each promised lower costs and a good return. The trouble was that competitors at home and abroad were buying the same machines. Once enough of them had, Buffett wrote, “their reduced costs became the baseline for reduced prices industrywide.” Each decision, viewed alone, was rational. Viewed together, the decisions “neutralized each other”¹.

The looms worked. The savings were real. They simply did not stay with the companies that paid for them. They flowed through lower prices to the customers.

Many AI investments now on executive agendas look like those looms. A tool that drafts proposals, screens invoices or forecasts demand is offered on the same terms to every competitor. It will probably pay back. The strategic question is who keeps the payback once everyone has it.

Advantage is a gap rivals cannot close by buying

Richard Rumelt gives a practical definition of competitive advantage. A business has one when it can produce at a lower cost than its rivals, deliver more value as buyers perceive it, or some mix of the two. A gap like that lasts only if something stops competitors from closing it; Rumelt calls that something an isolating mechanism². An advantage therefore needs two parts: a gap, and a reason the gap stays open.

AI can produce the first part, at least as a capability: a pilot that cuts a cost shows a gap is possible, not yet that it has reached the accounts. AI is poor, on its own, at producing the second. As The AI Competitive Advantage showed, a licensed model is not rare on the day the contract is signed, and what rivals find hard to copy is what an organization builds around it. That chapter explained why those layers resist imitation. This one asks the strategist’s follow-up questions: which AI investments should be treated as advantage at all, what keeps a lead open, and how a lead can grow instead of decaying.

When everyone adopts, customers keep the gains

Buffett’s mechanism is worth spelling out, because it applies to any AI productivity tool that rivals can buy on the same terms. Suppose, as an illustration, that a firm cuts its cost per order by ten units with an AI tool. For a while the saving is profit. Then rivals license the same tool or something close to it. Their costs fall too, and in a competitive market lower costs become lower prices. The firm keeps whatever share of the saving its rivals have not yet matched, and that share shrinks with every rival who catches up.

None of this means the firm should skip the tool. The firm that refuses the new loom does not keep its margin; it loses it faster, because rivals now undercut it. Michael Porter’s distinction between operational effectiveness and strategy, met in Start with Business Strategy, describes the same trap: doing the same things better than rivals is necessary, and it is not a position.

The practical lesson is about labeling and spending. An AI investment that every competitor can make is parity. It protects the business; it does not set it apart. Parity work deserves to be done quickly, cheaply and with shared platforms, and to be judged on speed and unit cost. It should not be presented to the board as a source of advantage, and it should not absorb the money and senior attention that the genuine candidates need.

Four competitive roles for an AI initiative

Porter’s older work gives the other half of the test. In Competitive Advantage he argued that there are two basic types of advantage, lower cost and differentiation, and that both are built in the specific activities a firm performs, from inbound logistics and operations to marketing and service³. An AI initiative matters strategically only if it changes one of those activities enough to move relative cost or the value buyers perceive. So two questions sort any portfolio. Does this clearly move cost or customer value? And how quickly could a well-funded rival reproduce the result?

In our judgment, the top right is where much of today’s AI spending sits, and that is fine. The bottom left is the quiet failure: an initiative that is genuinely hard to copy because it is technically ambitious, but that no customer would pay more for and no cost line notices. The top left is where strategy lives. It is usually a short list, in our experience often two or three items even in a large company, and leaders should expect it to be short.

What keeps a gap open

A gap in the top left still needs a guard. Rumelt lists the classic isolating mechanisms: patents, reputation, commercial and social relationships, network effects, dramatic economies of scale, and the tacit knowledge and skill that come from experience. He adds a practical point that matters a great deal for AI: a firm that keeps improving is harder to copy, because it presents a moving target².

Two features of the list stand out. First, almost none of the mechanisms is technical. Contracts, relationships, reputation and know-how are built by commercial and operating decisions, which is why advantage is a leadership question before it is an engineering one. Second, the last mechanism is different in kind. The others defend a position; the moving target keeps enlarging it. That is the mechanism the rest of this chapter builds toward. Proprietary data is a further guard, and the most debated one; the next chapter tests when it actually holds.

First is not the same as ahead

Many AI strategies rest on speed: move first and the lead will take care of itself. The research on first movers is more sobering. Marvin Lieberman and David Montgomery reviewed the evidence in 1988 and found three real sources of first-mover advantage: technological leadership through learning and research, preemption of scarce assets, and switching costs that make buyers reluctant to change supplier. They found equally real disadvantages. Followers can free-ride on the pioneer’s investment, wait for uncertainty about the technology and the market to resolve, benefit from shifts in technology or customer needs, and exploit the pioneer’s inertia once it has committed⁴.

Applied to AI, in our judgment, the balance tilts toward the follower. A pioneer that builds on this year’s models pays more for less capable technology than the firm that starts next year, and it may be tied to a design that ages quickly. Being early buys time, nothing more. The pioneer’s lead is worth exactly what it converts that time into: learning that stays inside the firm, customers whose work now runs through its service, or a contract that rivals cannot easily match. A lead that is not converted is simply the follower’s research budget, paid by someone else.

The flywheel: advantage that compounds

The strongest conversion is a loop. Marco Iansiti and Karim Lakhani describe the firms that compete well with AI as running a virtuous cycle: use of the service generates data, the data improves the algorithms, better predictions improve the service, and a better service draws more use. They argue that learning now joins scale and scope as a basic driver of a firm’s economics⁵.

A flywheel is the right picture for it. A flywheel is a heavy wheel that stores energy as it turns. Getting it moving takes real effort, and the first pushes barely show. But each push adds to the speed it already has, and once it is turning, each further push costs less than the one before. A competitor can buy an identical wheel the same afternoon. It cannot buy the speed yours has built up.

The picture also shows when a loop is worth building, and leaders should hold it to three conditions. The loop must run through something the firm owns: its customers, its operations or its contracts, not a vendor’s platform that every rival shares. Each turn must improve something customers pay for, which places the wheel in the top left of the portfolio and not in the science-project corner. And someone must be paid, or held to account, for keeping the wheel turning, because loops left to good intentions slow down. Whether the data in the loop is itself a durable asset, and whether it gets more valuable as more customers join, is a separate and harder question; the next chapter takes it up.

Story: the engine maker that sold hours

Rolls-Royce’s civil aerospace business shows the whole argument inside one organization, over a long period. Decades ago, by the company’s own account, its service model was built around repairing and fixing engines that had broken⁶. Under that model the work came when something went wrong. The maker had every reason to build good engines, but its service business earned its living from failures.

The change began with a contract, not a technology. In 1962 the company that later became part of Rolls-Royce offered “Power-by-the-Hour” for the Viper business-jet engine: a complete engine and accessory replacement service for a fixed cost per flying hour, which aligned the interests of maker and operator⁷. Its modern successor for airlines, TotalCare, works the same way. It is charged per flying hour and moves the risk of maintenance from the airline to Rolls-Royce. In the company’s words, “we are only rewarded for engines that perform”⁸.

That contract changed what data was worth. Once the company was paid only while engines flew, every sensor reading that could predict a problem became money. Engine Health Monitoring fed the service from onboard sensors⁷, and by the end of 2018 the company expected to receive more than 70 trillion data points a year from its in-service fleet⁶. The analytics let it plan maintenance ahead of failure; it reports extending the intervals between overhauls by around 25 percent, a company figure rather than an independent one,⁸. By 2012 services provided more than half of its revenue, which had reached 11.3 billion pounds in 2011⁷.

Now notice which parts a rival could buy. In December 2018 Rolls-Royce licensed commercially available industrial AI software to improve the availability of its Trent engine fleet⁹. The announcement states the aim; it does not report a measured result for the software on its own. Any engine maker could have signed the same license. What a rival could not buy was the flywheel the license plugged into: decades of contracts that pay for availability, a fleet of engines reporting from the wing, and the service network and know-how to act on what they report. The AI made the wheel spin faster. The contract was why it turned at all.

The story has an honest second half. A business paid by the flying hour earns little when aircraft do not fly. In 2020, as the pandemic grounded fleets, Rolls-Royce’s large-engine flying hours under its long-term service agreements fell to 43 percent of their 2019 level¹⁰. The guard that made the advantage durable also concentrated a risk. Strategy chooses which risks to carry; it does not remove them.

What this means for leaders

The first discipline is to give every AI initiative a competitive role and to fund, staff and measure it by that role. The four roles from the matrix each get their own funding rule. Parity and advantage candidates are both legitimate; an advantage candidate may win by differentiating or by compounding, and both are judged on customers and on rivals. A science project gets a small budget and a deadline. Noise gets nothing.

The second discipline is to look for the guard before celebrating the gap. A pilot that cuts cost is evidence of a gap. It says nothing yet about whether the gap will survive the year in which rivals buy the same tool. The guards that hold are mostly commercial and organizational: how the firm is paid, which customer workflows run through it, what its people know. Those are decisions for the executive team, not for the AI team.

The third is to treat speed as a means, not an end. Being early is worth what the firm converts it into. Each early initiative should name what it is meant to leave behind once rivals catch up: a learning loop, a switching cost, a contract. If it names nothing, it is parity done early, which is still worthwhile, but it should be priced that way.

Check yourself

1. If an AI tool cuts our unit cost by 15 percent, we have gained a 15 percent cost advantage.MythFact
2. Parity investments in AI are a waste because they create no advantage.MythFact
3. A competitive advantage needs both a gap in cost or value and a mechanism that stops rivals closing it.MythFact
4. Being first with an AI capability reliably produces a lasting lead.MythFact
5. A flywheel creates advantage only when the loop runs through something rivals cannot buy.MythFact
6. Rolls-Royce’s edge in engine services rested on AI software no rival could obtain.MythFact

Reflection

What comes next

Many of the guards in this chapter are commercial: contracts, relationships, reputation. One is claimed more often than any other and examined less: proprietary data. The next chapter, Proprietary Data, AI Moats and Differentiation, tests when data really keeps a gap open, when it grows more valuable as customers join, and when it is an attractive asset that a rival can work around.