Buying AI Is Not a Strategy

On the morning of April 15, 1894, in a four-storey brick building beside a canal in Columbia, South Carolina, a man named Arethas Blood walked to a switch and started a factory the way no factory had ever been started.¹

Blood was president of the Columbia Mills Company, and the switch connected his cotton mill to turbines at the end of the Columbia Canal. No boiler. No coal pile. No steam engine sweating in the basement. When the current came, seventeen electric motors woke together, and the first textile mill in the world to run entirely on electric power began to turn.

To feel what happened next, you need the picture of what a factory was — because the pattern this essay is about hides inside that picture. Every new technology begins by wearing the old one’s clothes: the first cars were built like horse carriages, coachman’s bench and all, because carriages were what everyone knew how to build. A factory in 1894 was built around its engine the same way. One giant steam engine sat at the heart of the building. From it, a long steel shaft — a rotating rod — ran under the ceiling of every floor, spinning all day. Leather belts hung down from the shaft to drive each machine below. Every machine in the building was, quite literally, chained to the ceiling. And the building stood four storeys tall because the physics demanded it: the shorter the shaft runs, the less power you lose, so you stacked your floors around the engine like a chimney.

Four storeys. Because the shafts said so.

Electricity very nearly didn’t make it into the building at all. When the mill was on the drawing board, the sensible choice was steam; every mill in the South ran on it. A General Electric salesman named Sidney B. Paine talked the directors into electricity with an argument that had nothing to do with productivity: electric motors of the kind he was selling don’t spark. In a building that breathes cotton lint, a spark is how a mill becomes a fire. Then he made a promise his own company had never kept. At the time, the largest motor of that type General Electric had ever built was ten horsepower. Paine took an order for seventeen of them at sixty-five horsepower each — the young company’s first major industrial installation, and about as large a bet as it had ever placed.²

Now the detail to hold on to. By the time electricity won the argument, the mill’s floor plan was already finished. There was no room on the floor for motors. So the engineers hung them from the ceiling.¹

Read that again, because it is the whole essay. The most advanced factory on earth — the first of its kind, the future arriving early — bolted its revolution to the ceiling of a building designed before the revolution existed. The seventeen motors turned the same overhead shafts a steam engine would have turned. The shafts turned the same leather belts. The belts dropped to machines standing exactly where the logic of shafts and belts had put them.

Nothing about the work changed that morning. The looms did not move. The workers did not move. The mill made heavy cotton cloth the way it had always been going to make it. What changed was invisible: the power now came out of a wire.

Keep the ceiling in mind. There is a reasonable chance you are working under it.

You were probably in the room when your company pulled its own switch. There was a town hall. There was a slide with a partner’s logo on it — OpenAI, Microsoft, Anthropic, Google. Someone senior said the word transformation, someone else said a copilot for every employee, and by the end of the quarter there was a licence with your name on it and, somewhere, a dashboard counting whether you had logged in.

And the looms did not move. The approval chain did not move. The weekly reporting pack, the seven sign-offs, the meeting in which eleven people watch one person share a screen — none of it moved. The work is the same work, in the same shape, done by the same people in the same sequence. What changed is invisible: some of the sentences now come out of a wire.

If that describes your organisation, nothing has gone wrong, and nobody has been foolish. You have simply joined the oldest tradition in the history of powerful tools.

That is the only claim this essay makes. Everything that follows is evidence: the two times this exact story has already run, a census of the present, the three purchases that pass for a strategy, the employee those purchases quietly depend on — and the companies, from fifty people to forty thousand, that are producing exceptional growth because they did the one thing the licence never asked of them. It will take you about half an hour.

Columbia Mills was not a mistake. It was the first move of a pattern that has now run twice at full length, in living memory both times: a powerful technology arrives, organisations buy it, bolt it onto everything they already are — and nothing happens. Both times, the growth arrived only when someone redesigned the whole system around the technology. And both times, that redesign was a strategy, not an upgrade.

First, electricity itself. For a generation after Gervais Street, electrification was mostly Columbia’s move repeated at scale: unbolt the steam engine, bolt motors in its place, and leave everything the steam engine had shaped — the shafts, the belts, the tall building, the layout — exactly where it stood. Through all those years of adoption, productivity barely moved. From 1899 to 1914, output per worker in American manufacturing grew about 1.5 per cent a year, the same drowsy pace as the steam decades. In the 1920s it jumped to over five per cent a year.⁴

Same motors. Same physics. Different factory.

What happened in between was not a better motor. It was a question. Engineers stopped asking where do we put the motor? and started asking a far more dangerous one: if power can come out of a wire anywhere, why does this building exist? Put a small motor on every machine, and the overhead shafts die. Kill the shafts, and the four-storey mill dies with them: a factory can sprawl across a single storey. And a single-storey factory can be arranged around a new master — the flow of the work itself. Machines lined up in the order of the tasks, instead of huddled near the power. In 1913, a plant in Highland Park, Michigan, arranged its machines in the sequence the work passed through them, and let the work move instead of the workers. You know it as the assembly line.⁵ The forty-year lag between the motor and the growth was never the technology maturing. It was buildings depreciating and men retiring.

Then it happened again, to a technology sitting on your desk. Computers spread through offices from the 1970s onward, and for two decades the growth statistics stayed flat — flat enough to become a famous joke. In 1987 the economist Robert Solow, who won the Nobel Prize that same year, wrote: “You can see the computer age everywhere but in the productivity statistics.”⁶

When the surge finally came, in the late 1990s, the McKinsey Global Institute went looking for its source — and found something almost embarrassing. Nearly a quarter of the entire American productivity acceleration traced to one unglamorous sector: retail. And within retail’s general-merchandise heart, the researchers concluded that the bulk of the gain, direct and indirect, was caused by a single company.⁷ Walmart had not bought better computers than anyone else. It had redesigned its whole operating system around the technology: goods flowing through its own distribution hubs, trucks unloading straight into trucks so stock barely touched a warehouse floor, checkout-scanner data flowing directly to suppliers, so that the shelves were restocked without anyone asking. The competitors it dragged along copied what they could and lifted the national statistics with it.

The same technology was on sale to the whole industry. Kmart bought plenty of it — and bolted it onto the company it already was. Its store scanners never fed purchasing data back to headquarters; a late, desperate technology programme ended with a nine-figure write-off of supply-chain software that never worked.⁸ In January 2002, Kmart filed what was then the largest retail bankruptcy in American history. The computers were available to everyone. The factory around them was not.

Economists eventually put a ratio on the invisible part. Studying exactly that computer era, Erik Brynjolfsson and Lorin Hitt found that inside the firms where technology paid, each dollar of it rode on roughly ten dollars of organisational change — redesigned processes, new decision rights, training, data.⁹ One part motor. Ten parts factory.

Business went through one loud, painful attempt to learn this lesson by name. In 1990, Michael Hammer told companies in the Harvard Business Review: “Don’t automate, obliterate” — layering technology onto an existing process wastes the technology; redesign the process or remove it. He was right, and a stampede followed. But in practice “re-engineering” collapsed into a euphemism for handing process design to outsiders with a mandate to cut payroll, the backlash was ferocious, and the word turned toxic for fifteen years — taking the correct idea down with it. Hammer’s own verdict on what went wrong deserves to be read slowly: he had been, he conceded, “insufficiently appreciative of the human dimension.”¹⁰

Hold on to his sentence. It is the hinge on which the second half of this essay turns.

Run the census of the AI era and its whole shape appears in three numbers.

Eighty-eight per cent of organisations now report regular AI use in at least one business function — McKinsey’s global survey, 1,993 organisations across 105 countries. Thirty-nine per cent report any effect of AI on their earnings, and most of those put it below five per cent of operating profit. The cohort that clears a real bar — meaningful, attributable earnings impact — is about six per cent.¹¹

Six per cent could be survey noise, except for this. BCG ran the question the other way around — outside-in, from audited financial statements, across more than six hundred listed US companies, no self-reporting involved — and its screen landed on the same six per cent.¹² A global questionnaire and a public-accounts screen, sharing no data, converging on the same thin sliver. The gap between the first number and the last is the entire subject of this essay.

The chief executives, asked directly, corroborate the gap from the inside: in PwC’s census of 4,454 of them across 95 countries, 56 per cent said their AI investments had so far returned nothing.¹³ Adoption is nearly universal. Growth is nearly absent. And the money is enormous — which is what makes the six per cent worth taking apart rather than admiring: they are walking proof that the technology is not the constraint.

Columbia Mills, at planetary scale: the motor installed, the building untouched, the meter never fitted. The rest of this essay is about the difference between the eighty-eight and the six — and it begins with what the money actually buys.

Walk into almost any large organisation and ask to see the AI programme, and you will be shown one of three purchases — usually all three. Each is expensive. Each is defensible in front of a board. None of them is a strategy.

The first play is seats for everyone. A licence rollout, a training calendar, a dashboard counting log-ins. The costs are concrete: Microsoft lists its enterprise Copilot at US$30 per user per month, so a 10,000-person company pays about US$3.6 million a year at list price before a single workflow changes.¹⁵ Then comes the programme around the seat — training, change management, a platform team — until the average large organisation now projects total AI spending in nine figures.¹⁶ What the seat money buys is access. What it assumes is that access converts itself.

The second play is a portfolio of pilots. The median AI portfolio sprawls across six-plus use cases at once, roughly double what the winning companies allow themselves.¹⁷ The attrition is brutal: nearly half of all proofs-of-concept are now scrapped, and the share of companies abandoning most of their AI initiatives more than doubled in a single year.¹⁸ Pilot purgatory has a simple function in organisational life: pilots are how an organisation feels motion without committing to redesign. A pilot can always be praised, paused, or quietly buried. A redesigned workflow cannot — which is exactly why it works and exactly why it is rarer.

The third play is the visible showcase. The Chief AI Officer, the innovation lab, the town hall, the mentions on the earnings call. And here it is worth pausing on why intelligent leaders buy all three plays, because the reason is not enthusiasm. Sixty-four per cent of chief executives told IBM’s surveyors — in writing — that they invested in AI ahead of understanding its value, for fear of being left behind; half say their tenure depends on AI delivering.¹⁹ Meanwhile, the same corporations praising AI on their earnings calls have been quietly adding it to the risk factors of their annual reports — 56 per cent of the Fortune 500 by 2024, up from nine the year before.²⁰ These are not the fingerprints of believers. They are the fingerprints of executives buying insurance against a narrative. That is not stupidity; under those incentives it is close to the only defensible behaviour. But one of the oldest marks of a bad strategy — we will meet the man who catalogued them later — is mistaking goals for strategy, and “adopt AI” is a goal wearing a strategy’s clothes.

There is one more thing to know about access without redesign: it can run backwards. In 2025 a research group called METR ran the field’s only controlled experiment — the same coin-flip method used in drug trials. Experienced software developers, using AI on their real work, believed it had sped them up by about 20 per cent. The stopwatch showed it had slowed them down by 19 per cent.²¹ They were not wrong about the size of the change; they were wrong about its direction. A second large experiment — 758 consultants at BCG, also assigned by coin flip — found AI lifted the quality of work by roughly a third on tasks inside its competence, and made the same professionals worse just outside it, where the boundary is invisible from the inside.²² A tool bolted onto an unchanged workflow does not merely under-deliver. It can quietly charge you, while feeling like speed.

Now step back and look at the three plays together, because they share a load-bearing wall. Seats for everyone, a portfolio of pilots, a showcase on top — every one of them rests on the same unstated belief: that, given the tool, your employees will work out how to transform your company from below.

That belief deserves a section of its own. Because it has been tested — by history, and by forty-eight thousand survey respondents.

The employee the three plays are counting on does exist. History keeps a handful of them — people who, handed a tool and no mandate, dragged an entire corporation into its own future. It is worth meeting the two most famous, and worth telling their stories properly, because the arithmetic at the end depends on feeling how rare they are.

In the mid-1980s, a young Sony engineer named Ken Kutaragi watched his daughter playing a Nintendo Famicom — and, against everything his employer believed, concluded that video games were about to become one of the most serious businesses on earth. Sony’s culture held that games were toys, and toys were beneath it. Kutaragi did not write a memo. He secretly designed a sound chip — the SPC700 — for Nintendo’s next console: unauthorised work for an outside company, and, in Sony’s eyes, for a mere toymaker.²³ When Sony’s executives discovered it, they were furious, and he was nearly fired. He kept his job for one reason: the chief executive, Norio Ohga, personally stood in the way.

Then Nintendo made it personal. Sony and Nintendo had agreed to bring compact discs to Nintendo’s console; Sony’s half of the partnership was a machine to be unveiled as the “Play Station” at the Chicago consumer-electronics show in June 1991. Sony announced it on stage on the first of June. The next morning, at the same show, Nintendo announced a partnership with Philips instead. Sony had been publicly jilted at its own wedding.²³ Kutaragi’s response was to raise the heresy: Sony, he argued, should stop building parts for other people’s consoles and build its own. Almost the entire company opposed him — “Most of the executives were fiercely opposed,” he recalled later; “everyone told us we would fail.”²⁴ At the decisive meeting in June 1992, he put the question to Ohga directly: “Are you going to sit back and accept what Nintendo did to us?” Ohga, by every account in a fury, declared there was no future in building for Nintendo’s machine — “Let’s chart our own course.” And because the electronics establishment wanted the project dead, Ohga sheltered it where the establishment couldn’t reach: inside Sony Music, a separate company in another district of Tokyo.²³ The PlayStation launched in December 1994. Within a few years the games division was supplying roughly a quarter of Sony’s operating profit — one of the most profitable things the company has ever built.²⁵

Three decades earlier, in America, an engineer at Hewlett-Packard named Chuck House was building a large-screen electronic display that everyone had turned down. The customer it was designed for rejected it. His boss killed it. Marketing forecast a market of about thirty units. Finally the founder himself, David Packard, toured the lab and issued an instruction with no room for interpretation: “When I come back next year, I don’t want to see that project in the lab.” House considered the sentence carefully and found the loophole he needed: if it couldn’t be in the lab, it would have to be in production. He took vacation days, loaded the prototype into his Volkswagen, drove it to customers, and came back with orders. When Packard returned the following year, the project was not in the lab. Seventeen thousand units eventually sold, and the display found its way into the monitors of the Apollo programme — screens that watched men walk on the Moon. In 1982, Packard hung a medal on him for it: the Medal of Defiance, for extraordinary contempt and defiance beyond the normal call of engineering duty.²⁶

Both stories are told at company off-sites, usually as encouragement. Told properly, they are a warning — because of what they cost, and how they were survived. Now the punchline both of them exist to set up. Hewlett-Packard awarded that medal exactly once in its history. The most celebrated engineer-initiative culture of the twentieth century produced, and honoured, one Chuck House. Sony nearly fired its one Kutaragi, and he survived only because the chief executive personally shielded him — twice. Two mavericks, two giant corporations, three decades apart.

Your AI business case quietly assumes ten thousand of them.

What does the typical employee actually do with the licence? Not nothing — something far more interesting. The largest survey ever run on the question — 48,000 workers across 47 countries, by KPMG and the University of Melbourne — found that 57 per cent of employees who use AI conceal it, or present its output as their own.²⁷ Concealment is not the exception. It is the majority behaviour. Slack’s workforce research explains what the hiding is for: nearly half of desk workers are uncomfortable telling their manager they use AI.²⁸ The reasons they give, in order: it feels like cheating; it makes them look less competent; it makes them look lazy. “Company policy” comes last on the list. The fear is not of the rulebook. It is of judgement.

And when the gains are visible, employees have learned precisely what happens to them. Denmark ran the natural experiment: 25,000 workers, high trust, strong unions, the best-protected labour market on earth. Of AI’s productivity gains, three to seven per cent passed through to workers’ pay — and more than 80 per cent of the time saved was silently reabsorbed into more work. One of the researchers compressed the finding into a sentence your employees already know by heart: “the reward for efficient workers is more work.”²⁹ Closer to home, the mirror is exact: 81 per cent of Singapore workers say AI has raised their productivity, and 46 per cent of employers can see it — a thirty-five-point hole in the ledger, which is what a workforce quietly banking its own gains looks like from the corner office.³⁰

A typical employee is not Kutaragi. A typical employee comes in, does the job well, protects themselves, and goes home — and given what revealing a productivity gain earns them, hiding it is not a character flaw. It is the rational move inside the incentives leadership designed. This is the human dimension Hammer confessed to under-appreciating, intact after thirty-five years: the three plays hand the tool to the workforce and wait for transformation from below, and the workforce, reading its incentives correctly, declines.

So growth cannot be summoned from below. What does producing it actually look like?

It looks nothing like a rollout. The companies converting AI into growth right now run from fifty people to forty thousand; they sell slide decks, language lessons, legal letters, furniture, banking. What they share is not a model or a vendor. Each of them decided what game it was playing before it bought a single seat — and then rebuilt the way it works, with AI woven through every part of the answer.

Start with the ones born this way. Gamma, a San Francisco company that sells AI-powered presentation software, is about fifty people. It has passed US$100 million in annual recurring revenue, serves seventy million users, has been profitable for more than two years, holds more cash than it has ever raised, and was valued at US$2.1 billion late last year. The New York Times profiled it when it was twenty-eight people serving nearly fifty million users.³¹ The interesting part is not the numbers; it is the deliberateness behind them. Founder Grant Lee runs an explicit strategy — stay small, reach profitability first, hire generalists — with refusals attached. And AI is threaded through every function rather than merely sold as the product: the growth lead built his own analytics system instead of hiring a data team; the marketing lead fed thousands of customer interactions into a model to build the personas that guide strategy. “I want to convey to the market,” Lee says, “that small teams can do amazing things.” It is the thesis of this essay in miniature: the organisation designed around the technology, not the technology bolted onto the organisation. And Gamma is no longer even the extreme case. Base44, an Israeli software firm, reached roughly US$189,000 of profit in a single month with eight employees and no outside funding, and was sold to Wix for US$80 million in cash six months after it was founded — its founder had spent, by his own account, a fifth of his time using AI to automate his own company.³²

The fair objection arrives immediately: those companies sell AI, and right now AI sells itself. True. So look at companies that sell ordinary things.

Duolingo sells language lessons. Its first hundred courses took about twelve years of human effort to build. Then it rebuilt the content operation itself — generative AI at the core of a redesigned production system, with shared content flowing across languages — and launched 148 new courses in roughly one year, the largest expansion in its history. Its chief executive, Luis von Ahn, put the before-and-after in one sentence: “Developing our first 100 courses took about 12 years, and now, in about a year, we’re able to create and launch nearly 150 new courses.”³³ The same push had a cost the company did not manage well: an “AI-first” memo announcing that contractors’ work would go to AI drew a very public backlash, and the CEO spent a month walking back the framing. A strategy includes owning its trade-offs out loud; Duolingo paid for learning that in public.³³

Garfield Law sells something even more ordinary: debt-recovery legal services for small businesses, from £2 for a polite letter to a late payer. It is a firm of roughly fifteen people, and in May 2025 it became the first AI-driven law firm ever authorised by England’s solicitors’ regulator — the AI is not a garnish on the practice; vetted and bounded by the regulator, it is the practice, with named human solicitors accountable for its work. In its first year it reports more than six hundred claims initiated and over £500,000 recovered for clients; in May 2026 it won its first contested trial, the client awarded £7,000, Garfield’s fees running to about £400 — against a defendant paying for a full human legal team.³⁴ The product is centuries old. The operating model underneath it is not.

Then there are the incumbents — the ones who had to do it the hard way, inside an existing organisation. Their stories matter more to a reader running one.

DBS is Singapore’s largest bank, roughly forty thousand people. In 2009 a new chief executive, Piyush Gupta, inherited an institution whose own customers joked that its initials stood for “Damn Bloody Slow” — and made a diagnosis that had nothing to do with other banks: the real competitors of the coming decade would be the technology platforms, and against them a slow bank was a dead bank. The direction he set was equally unbankerly: make banking joyful; run the place like a “27,000-person startup”; benchmark not against banking peers but against Google, Amazon and Netflix — inside the bank the acronym was Gandalf, and the D stood for DBS. The direction came with refusals attached: stop measuring against banks, dismantle the meeting culture, no innovation-lab theatre bolted to the side of an unchanged bank. Then came a decade of coordinated moves: a culture programme, mass reskilling, rebuilt data infrastructure, cross-functional platform teams. And one discipline that remains unique in this literature — AI value targets published in the annual report before being claimed, with outcomes measured against control groups, the way a trial measures a drug.³⁵ The results: World’s Best Bank from Euromoney, The Banker and Global Finance; a place on Harvard Business Review’s ten most transformative organisations of the decade; and roughly S$750 million of measured economic value from AI in 2024, rising to around S$1 billion in 2025, across more than 430 use cases. Asked late last year whether AI’s promise was still ahead, chief executive Tan Su Shan answered in five words: “It’s not hope. It’s now.”³⁵

IKEA’s story is quieter and, for most operating executives, more useful. In 2021 its largest franchisee switched on a customer-service bot named Billie — christened, inevitably, after the Billy bookcase. Billie worked: it now handles most inbound queries. On paper, that is the classic setup for a layoff announcement. Instead, someone read the bot’s own data and asked what the rest of the callers were asking for. The answer, buried under years of order-status calls, was help designing rooms — demand requiring taste and judgement, which no bot could serve and no busy phone queue had ever had capacity to notice. So the company retrained some 8,500 call-centre workers as remote interior-design advisers, and that channel now produces on the order of €1.3 billion a year in sales, with no layoffs attributed to AI.³⁶ The lesson generalises: the saving is not the prize. The capacity is — if someone decides in advance what the capacity buys.

Klarna shows what the discipline looks like when it wobbles. Its AI assistant was a real redesign: in its first month it did the measured work of seven hundred customer-service agents, with a projected US$40 million profit improvement. Then the company let cost dominate the design. Quality leaked, and Klarna corrected in public — humans rehired for the judgement tier of customer service.³⁷ Read properly, that sequence — redesign, measure, correct — is not the strategy failing. It is the strategy working, out loud.

Moderna, the Massachusetts drugmaker, answered a question most companies never ask: if AI changes what every role does, why are the people who design roles and the people who deploy technology in different departments? It decided its AI problem was a people problem, and merged human resources and information technology into a single function under one executive. Her mandate: sort every task in the company into automate, augment, or human, and rebuild the roles accordingly — on top of some three thousand internal AI assistants employees had already built.³⁸ Alfred Chandler compressed a century of business history into four words — structure follows strategy — and Moderna is the rare company that let it.³⁸

And the market, it turns out, has been grading all of this. Across BCG’s outside-in cohort, the AI leaders delivered three-year shareholder returns 9.3 percentage points above their industry-adjusted median. Decompose the premium and it is all delivered performance — revenue growth and margin — with approximately zero paid for the AI story itself. And the leaders grew headcount faster than the laggards while growing revenue per employee faster still: the winning cohort is not shrinking its way to the prize.¹²

Eight companies, from fifty people to forty thousand, on three continents. Not one of them has access to a model you cannot rent. So what, exactly, do they share?

Not a secret, and not a framework. What every winner in this essay did has had a name for a long time. The cleanest anatomy of it belongs to Richard Rumelt, whose Good Strategy Bad Strategy (2011) remains the best book ever written on the subject — and the man who catalogued the marks of bad strategy, including the one from Section IV: mistaking goals for it.³⁹ A real strategy, Rumelt showed, has three parts. A diagnosis: what is actually going on, and what is in the way. A guiding policy: the overall approach — which, like a guardrail, both directs and constrains, telling you what you will not do. And coherent actions: a coordinated set of moves that reinforce one another — not a wish-list. Three parts, in that order. Everything the six per cent share is these three moves, done with the seriousness their AI programmes usually skip — and with AI woven through each one.

The diagnosis comes first, and it is usually unfashionable. In 2003 the LEGO Group was dying — losses two years running, and an internal review that warned bluntly of possible debt default and break-up. Everyone knew the fashionable diagnosis: children had gone digital, and the plastic brick was obsolete. The company had spent a decade acting on that theory, diversifying into software, theme parks and television. When Jørgen Vig Knudstorp’s team went looking instead at what was actually going on, they found a much less romantic disease: runaway complexity. The number of unique components had more than doubled, to over fourteen thousand; thirty products carried most of the sales; the innovation sprawl was devouring cash. Kids, it turned out, still loved bricks — the company had simply lost control of itself. LEGO cut the sprawl, sold the theme parks, rebuilt around the brick, and within a decade was, by revenue and profit, the biggest toy company in the world.⁴⁰ The fashionable diagnosis was flattering — the world changed on us — and would have been fatal. The true one was embarrassing, and worth billions.

That is the move DBS made when it named its competitors as tech platforms rather than banks, and the move IKEA made when it read the bot’s own call data and found a design business hiding behind the phone queue — note that IKEA’s diagnosis came out of the AI deployment itself. Strategy and AI do not queue politely, one after the other; they interleave. And this is where AI now slots into the first move directly: the fastest route to an honest diagnosis of your own company runs through its operational data — the tickets, call logs and cycle times where the value is actually leaking. Fewer than one in five organisations baseline any of this before deploying,¹¹ which is a polite way of saying that most AI programmes begin without asking what is actually going on. You cannot diagnose what you never measured.

The guiding policy is a direction — and its refusals. In 1997 Apple was, by its returning founder’s own account, about ninety days from insolvency. Steve Jobs’s response was not a list of initiatives. In a product-review meeting he stopped the parade, drew a two-by-two grid on a whiteboard — consumer and professional, desktop and portable — and killed roughly seventy per cent of the product line to leave four squares. The refusals, not the additions, saved the company; as Jobs put it, deciding what not to do is as important as deciding what to do.⁴¹ Gamma’s version is refusing the headcount race and the growth-before-profit default. DBS’s version was refusing to benchmark against banks. The winners’ AI portfolios show the same signature in miniature: concentration on roughly three and a half use cases, against the market’s sprawling six-plus.¹⁷ The constraint that makes refusal necessary was named by the economist Edith Penrose back in 1959 — growth is rationed by management attention, not by budget. Every additional pilot is attention subtracted from the one redesign that would matter.

A guiding policy for the AI era carries two clauses the default plays always skip. It names AI’s role explicitly — which workflows the machine will own, which it will assist, which stay human; Moderna’s automate-augment-human sort of every task in the company is the working template.³⁸ And it decides, in advance and in writing, what the freed capacity buys. Ford’s cheap Model T bought ten million new customers; IKEA’s bot bought a €1.3 billion design channel. Where the written answer is “the same work with fewer people,” the record speaks for itself: 55 per cent of leaders who made AI-attributed layoffs now tell surveyors the decision was wrong.⁴² A layoff booked before the workflow is redesigned is usually not a strategy. It is evidence that a test was skipped.

The coherent actions redesign the ecosystem in coordination — and this is where almost everyone stops early. For decades, Toyota did something no secretive company would ever do: it opened its factory doors. Hundreds of thousands of executives toured its plants. They photographed the kanban cards — the little tickets that pull parts through the line — and the andon cords that let any worker stop production, went home, installed both, and waited for Toyota’s results. The results never came. When two Harvard researchers finally decoded why, in 1999, their explanation was almost insulting in its simplicity: the visitors had confused the tools they could see with the system that made the tools work. Toyota’s advantage was never the artifacts; it was the coordination of people, incentives, process and technology as one designed system.⁴³ The tool you can photograph is the smallest part — which is precisely the mistake of buying the AI tool alone. It is Highland Park’s lesson, and DBS’s platform teams, and Moderna’s merged function, all in one story.

In the AI data, the coordination shows up as the single most powerful variable anyone has measured: of twenty-five organisational attributes McKinsey tested, the redesign of workflows had the largest effect on AI’s earnings impact, and organisations that did it were 5.3 times likelier to report enterprise-level value. Roughly a quarter of companies practise it.⁴⁴ But the actions are not only mechanical, and here Section V comes back with its bill. One coherent action outranks every training programme on the calendar: making it safe — and profitable — for an employee to reveal what the tools really let them do. The question is not what training your people need. It is what protection they need before showing you where the hours actually went. IKEA’s no-layoff redeployment is what that protection looks like in practice; Denmark’s three-to-seven-per-cent pass-through is what its absence looks like at national scale. And threaded through all of it, the meter goes in before the motor: a baseline before each deployment, outcomes measured against control groups, targets published before they are claimed — the way DBS has done for a decade. AI belongs everywhere in these actions, from the diagnosis to the meters. Everywhere in the answer. Never the strategy itself.

Because what is being redesigned is not a toolchain. An organisation is a living system — people, incentives, workflows, structure, culture, technology — delivering inside a market that will not hold still, and its parts pay only when they move together. That is all the word strategy has ever meant here.

None of this requires a bigger model, a braver vendor, or a larger budget. The companies in this essay are not geniuses. They treated AI as what it is — the most powerful component ever handed to a strategist — and then they actually did the strategy. Everyone else bought the motor and kept the shafts.

The mill on Gervais Street ran for eighty-seven years. In 1981 it finally stopped. The building survived — it was too well made not to — and in 1988 it reopened as the South Carolina State Museum. Today, school groups file through the halls where the seventeen motors once hung, looking at exhibits about how people used to work.⁴⁵ The first building on earth to bolt the future to its ceiling is now, officially, a monument to the past.

Yours doesn’t have to be. The whole distance between the eighty-eight per cent and the six comes down to one question, carried out of this essay into Monday morning:

Thank you for the half hour. It was spent the way the six per cent spend their budgets: on the parts that move the outcome.