When good becomes cheap.

"Good is the enemy of great." So begins one of the best-selling management books ever written. In the late 1990s Jim Collins and a team of researchers sifted through 1,435 companies and found eleven that, after years of ordinary results, had gone on to beat the stock market by at least three to one for fifteen years.¹ The book that described them, Good to Great, appeared in 2001, sold some four million copies, and left behind a vocabulary that boardrooms still speak: Level 5 leaders, the hedgehog concept, the right people on the bus, the flywheel. Its opening line carried a theory of complacency. Most companies, Collins argued, become good, find good comfortable, and stop.

A quarter of a century later the line reads differently. A competent first draft of a contract, a working module of code, a tidy analysis of last quarter's customer losses, a courteous reply to an angry client: work that once took skilled people hours now takes a subscription and a few minutes of review. Good used to be an achievement, and its comfort was the danger. Increasingly, good is something you can buy, and it is on sale to your competitors at the same price.

My argument is that this changes the meaning of "great" more than it changes the road to it. When competent execution gets cheaper, competitive advantage moves toward the things cheap execution does not supply: choosing which problems deserve solving, judging whether an answer is right, and building an organization that turns capability into dependable results. Collins's emphasis on people, focus and discipline becomes more important, not less. But one of his best-loved ideas, the flywheel, carries a hazard he had little reason to dwell on in 2001. An accelerator speeds up whatever it is attached to, including a mechanism that is quietly losing its value.

One qualification before going further. AI does not make all execution cheap, and it makes nothing excellent on its own. The claim is narrower. The cost of particular capabilities is falling fast, and companies that mistake owning those capabilities for having an advantage are in for a surprise.

What Collins found, and how far to trust it

It helps to recall what the book actually claimed. Collins paired each of his eleven companies (among them Abbott, Kimberly-Clark, Nucor, Walgreens, Wells Fargo, Circuit City and Fannie Mae) with a comparison company in the same industry that had faced the same opportunity and failed to seize it. Then he looked for the differences. He found leaders who combined personal humility with ferocious will. He found that they chose people before strategy, confronted brutal facts, narrowed their focus to what they could be best in the world at, and built a culture of discipline that made heavy bureaucracy unnecessary. And none of them got there in a single stroke. The transformation, in his image, felt like pushing a giant, heavy flywheel. There was "no single defining action, no grand program, no one killer innovation," only turn upon turn until momentum took over.¹

One finding bears directly on this essay. Writing during the dot-com fever, Collins reported that 80 percent of the good-to-great executives his team interviewed did not mention technology among the top five factors in their transformation.² That was not because they ignored it. Several of the companies were sophisticated, even pioneering, users of technology. They simply treated it as an accelerator of momentum they had already built, never as the source of that momentum. They adopted what fit their focus and ignored the rest.

A fair reader should also know the book's weaknesses, because they matter for what follows. In 2007 the management scholar Phil Rosenzweig argued in The Halo Effect that studies like Collins's are compromised by the way they gather evidence.³ Choose companies because they performed well, then interview their executives and read their press clippings, and everything about them will look admirable, because success casts a glow over the culture, the leadership and the strategy alike. Outcomes get mistaken for causes. The record since publication lends the point some force. Circuit City went bankrupt in 2008, and Fannie Mae was placed in government conservatorship the same year.⁴ The economist Steven Levitt observed that summer that a portfolio of the eleven looked likely to have trailed the S&P 500 since the book came out.⁵

I don't think this makes the book worthless. It makes it a set of hypotheses rather than a set of laws: plausible mechanisms, gathered with care but without controls, that deserve to be tested against new conditions. AI is a new condition, and testing them against it is what the rest of this essay tries to do.

The floor comes up

The best evidence we have about generative AI at work points in a consistent direction. Erik Brynjolfsson, Danielle Li and Lindsey Raymond studied 5,179 customer-support agents at a software company as an AI assistant was rolled out among them. The number of issues resolved per hour rose 14 percent on average. For novice and less-skilled agents it rose 34 percent. For the most experienced and skilled agents the effect was minimal.⁶

The same shape appeared in a writing experiment. Shakked Noy and Whitney Zhang gave 453 college-educated professionals realistic writing tasks from their own occupations and gave half of them ChatGPT. Time spent fell 40 percent, graders rated the work 18 percent higher, and the gap between stronger and weaker writers narrowed.⁷ At Boston Consulting Group, a team led by Fabrizio Dell'Acqua gave 758 consultants access to GPT-4 for a set of realistic consulting tasks. They finished 12.2 percent more tasks, worked 25.1 percent faster and produced work rated about 40 percent higher in quality. Consultants who had scored below average improved by 43 percent; those above average, by 17 percent.⁸

Notice the pattern. The tools help nearly everyone, but they help the weakest most. Brynjolfsson and his colleagues offered an explanation: the system had learned from the conversations of the best agents, so it was in effect handing their habits to everyone else. The know-how that once made a top performer rare was being copied and distributed. In economic terms, AI compresses the spread of competence from below. The floor rises toward the ceiling. Figure 1 sketches what that does to a company that has been winning on execution alone.

The figure is a model, but the reasoning behind it is old. In 2003 Nicholas Carr published an essay in the Harvard Business Review under the deliberately provocative title "IT Doesn't Matter."⁹ His argument was that a resource confers advantage through scarcity, not ubiquity. Railroads and electricity transformed business, and the early companies that controlled them prospered, but once anyone could buy power from the grid, having electricity distinguished no one. Carr was criticized, fairly, for underestimating what companies like Amazon would build on commodity technology. Yet his critics' best rebuttal makes his underlying point. Amazon's advantage did not come from having the technology. It came from what the company chose to do with it, and that was scarce, because judgment always is.

Why output stops being a measure

When good becomes cheap, one of the oldest signs of a well-run company becomes treacherous: how much it produces. An organization that drafts twice as many proposals, ships twice as many features or publishes twice as many reports looks twice as productive. If its competitors can do the same with the same tools, it has merely kept pace. And if what it produces is something customers increasingly expect to get free, it may have doubled its output of a thing whose price is falling toward zero.

There is a second, subtler trap. Cheap capability is uneven capability. The BCG researchers also gave consultants a task chosen to sit just outside what the model handled well. There, consultants using AI were 19 percentage points less likely to reach the correct answer than consultants working without it.⁸ The authors called the boundary a jagged frontier: invisible, irregular, and easy to cross without noticing. A 2025 study by the research group METR found something related among experienced software developers working in large codebases they knew well. With AI tools they took 19 percent longer to finish their tasks. Beforehand they had predicted the tools would speed them up by 24 percent, and afterward they still believed they had been sped up by 20 percent.¹⁰

These results do not cancel the gains. They locate them. Knowing which side of the frontier a task falls on, and noticing when a fluent answer is wrong, is a form of judgment, and it is exactly the kind that does not come in the box. A company that has AI has the same capability as everyone else. A company that knows where its capability ends has something rarer.

The flywheel can spin the wrong way

Now to the complication. Collins's flywheel is a theory of compounding. Each disciplined action builds on the last, momentum accumulates, and eventually the wheel's own weight carries it forward. He gave technology a place in that picture as an accelerator, and AI is plausibly the most powerful accelerator a management team has ever been offered. Attach it to the flywheel and the wheel turns faster.

But an accelerator is indifferent to direction. It will speed a wheel that is carrying a company toward its customers' future, and it will speed one that is carrying the company toward their past. Picture a firm whose business is producing standard analyses, routine software integrations or first-pass document reviews. AI can make it extraordinarily efficient. Unit costs fall, utilization rises, and every number on the operating dashboard improves. Meanwhile the same tools let its clients do much of that work themselves and let new competitors offer it for a fraction of the price. The wheel turns faster each quarter as the ground beneath it gives way. Nothing in the flywheel reveals this, because a flywheel measures momentum, not destination.

Circuit City offers a sobering illustration, even though its fall came a decade before generative AI. In Good to Great it was a model of disciplined execution, an electronics retailer whose big-box stores were backed by early investment in point-of-sale and inventory systems. By 2007, losing ground to Best Buy, it tried to repair its economics by pushing harder in the direction it already knew: cost. That March it fired about 3,400 of its highest-paid store associates, roughly 8 percent of its store workforce, and set out to replace them with cheaper hires.¹¹ Analysts blamed the next month's weaker sales on less knowledgeable staff, and by the holidays the company was writing to some of the fired workers to invite them back. It filed for bankruptcy in November 2008. Collins later studied falls like this one. In How the Mighty Fall he described a sequence that begins with hubris born of success and runs through the undisciplined pursuit of more and the denial of risk and peril.¹² Each stage can feel, from the inside, like a flywheel spinning with conviction.

So the question that matters is no longer only how to persist. It is how to tell the discipline to persist from the courage to stop. Collins's own framework holds the start of an answer, in the principle he called confronting the brutal facts. Persistence is a virtue only while the facts say the wheel still creates value, and a few questions make those facts harder to avoid. If every competitor could produce this as cheaply as we now can, would customers still pay us for it? Does the unit we count (reports, tickets, lines of code) still track the outcome the customer wants? And if we were founding the company today, which turns of this wheel would we decline to start? Figure 2 shows two wheels with the same accelerator attached.

First who, when the what keeps changing

Of all Collins's principles, the one most likely to survive AI intact is the one he put first: get the right people on the bus before deciding where to drive it. What changes is how you recognize the right people.

Credentials, tenure and mastery of an established process have long served as reasonable stand-ins for capability. They become incomplete when the process itself is being rewritten. Recall the support study, in which the most experienced agents gained little because the assistant had in a sense already absorbed what they knew. That does not make them less valuable. It moves their value from doing the work to judging it, teaching it, and deciding what the system should learn next. An organization that measures them by throughput will conclude they have been overtaken. One that measures them by the errors they catch and the problems they frame will conclude the opposite. Circuit City, measuring its most experienced people by their wages, drew the first conclusion and paid for it.

The people who matter most now are those who can evaluate an unfamiliar output, who know where the tools' frontier runs in their own field, who keep learning as the work changes, and who accept responsibility for results a machine helped produce. That standard has an uncomfortable implication for leaders. The colleague who questions an AI initiative may be showing better judgment than the one who embraces it. Enthusiasm is not evidence, as the METR developers' confident and mistaken self-assessments show. A skeptic who asks where the tool fails is doing the part of the work the tool cannot do.

The discipline to leave things unbuilt

Collins found that the good-to-great companies kept "stop doing" lists and treated them as more important than their to-do lists.¹ In How the Mighty Fall he quoted David Packard's warning that a great company is more likely to die of indigestion from too much opportunity than of starvation from too little.¹² Cheap creation makes that warning newly urgent.

When a prototype costs a week and a dashboard costs an afternoon, the number of plausible initiatives grows far faster than an organization's capacity to absorb them. Each one looks inexpensive, because its cost of creation is visible and small. Its cost of keeping is invisible and continuing. Someone has to own it, connect it to the systems around it, maintain it, train people to use it and keep paying attention to it. I have argued elsewhere that technology is not a toaster, and the receipt keeps printing just the same for software an AI wrote in an afternoon. A company can now build almost anything. It still cannot meaningfully operate everything. The scarce resource has moved from the capacity to make things to the judgment about which things deserve to exist.

Objections, considered

"Collins's own companies fell, so why build on him?" Because the way they fell supports the argument rather than refuting it. Their mechanisms kept turning after the world those mechanisms served had changed. The halo-effect critique is a good reason to treat Collins's principles as hypotheses, and the later history of his eleven companies is, among other things, a test of the hypothesis that disciplined persistence is always a virtue. It isn't.

"AI will soon supply the judgment too." For some kinds of judgment it may. If so, the floor rises again, and the question this essay asks comes back unchanged: which capabilities still make a difference once everyone has the new ones? The answer can move; the question does not. And one thing will not be delegated however capable the tools become, which is accountability. A board cannot tell regulators, customers or patients that the model decided.

"This undersells the productivity gains." It isn't meant to. A rising floor is good news for customers, who get competent work for less, and for the less experienced workers who gain the most. The point concerns where advantage lives, not whether value is created. History suggests that the largest gains will go to companies that redesign the work rather than bolt a new tool onto an old process. When the economic historian Paul David studied the electrification of American factories, he found that productivity barely moved for decades while owners simply swapped a dynamo in for the steam engine and kept the old layout. It surged in the 1920s, after managers rebuilt their factories around what small electric motors made possible.¹³ Deciding how to rebuild was judgment, not capability.

AI raises the floor. Greatness depends on knowing which capabilities still make a difference.

Questions for the next board meeting

• Name what just got cheap. Which of our strengths could a competent rival now buy with a subscription? Those are no longer advantages. They are the price of entry.
• Audit the wheel's direction, not its speed. For every process AI is accelerating, ask whether we would start it today. Faster is a fact about the wheel, not about the destination.
• Count outcomes, not output. Drafts, tickets and lines of code are cheap. Problems solved, customers kept and errors avoided are not.
• Reward the judgment you can see. Promote the people who catch the tool's mistakes, frame the right problems and are willing to say no.
• Keep a stop-building list. Before a prototype is launched, give it an owner and an estimate of what it will cost to keep.

Good, revisited

Collins called good the enemy of great because good was comfortable. It let companies stop. Cheap good is a different kind of enemy. It does not tempt a company to stop; it tempts it to speed up, and to mistake the speed for progress. The companies that look great a decade from now will not be the ones that adopted AI first or produced the most with it. They will be the ones whose leaders kept asking what still made them different, and had the discipline to point the wheel there.

The floor is rising for everyone. That is the good news, and it is also the warning.