AI as layoffs, not opportunity

There is a future where AI investment goes into creating value: new products, new kinds of exchange, work that was not possible before. There is another where it goes into cutting headcount. The technology does not choose between them. The way you sell it does, and almost everyone sells it on the cut.

That is Rory Sutherland's argument, and the uncomfortable part is not the claim, it is the reason behind it.

This piece takes his claim and runs it up the org chart one level at a time: why the cut always wins, why the people who block the alternative are not villains, what it costs to protect a place where trying is cheap, why every protection short of the ownership layer eventually fails, and which structures actually survive. The same asymmetry shows up at every level. That repetition is the finding.

The legible win

AI gets sold on cutting costs, not creating value, and the reason is not laziness. A cost cut is easy to see; a new opportunity is not. Cut ten jobs and the saving lands on this quarter's books with your name next to it. Create value and the payoff is slow, spread out, and impossible to pin on the person who approved it. Big organisations reward what they can see and credit, so AI drifts toward the cut by default. Nobody decides to be unimaginative. The scoreboard decides for them.

The cut is measurable, creditable, and bookable this quarter. The opportunity is none of those. So the organisation reaches for the cut.

The self-checkout rollout is the same machine on a longer timeline. A consulting engagement identifies front-end labour as a visible cost line and replaces cashiers with kiosks. The saving lands on the next earnings call with a name on it. Eighteen months later, shrinkage has spiked across the estate and the asset protection team is in crisis. The two things (the cut and the theft wave) sit in different spreadsheets, owned by different people, measured on different timelines.

There is also the psychology: shoplifting from a human cashier feels like stealing from a person; stealing from a machine that just accused you of an unexpected item in the bagging area feels like a different moral category entirely. The behaviour the cut unlocked was predictable. But the person who could have predicted it was not the person responsible for the outcome, so it was not predicted. Several major retailers have since pulled the kiosks back out. The people who authorised the original rollout have long since collected their bonuses and moved on.

AI's version of the kiosk is already on the books. Klarna spent a year telling investors its assistant was doing the work of seven hundred customer-service agents, and froze hiring on the strength of it. The saving was specific, current, and had a name on it. Then the CEO conceded publicly that cost had been allowed to dominate the evaluation, that quality had dropped, and that humans were being hired back. The cycle that took self-checkout eighteen months ran in about a year, and it ended the same way: the cut was booked by one set of people on one timeline, and the degradation arrived later, diffuse, in someone else's column. The pattern is not new. The clock speed is.

The four horsemen

Sutherland names the people who hold the veto: finance, compliance, procurement, and HR, the four horsemen of the bureaucratic apocalypse. Between them they can block almost anything, and they all face the same incentive. If something new goes wrong, they get the blame, so they are heavily biased against trying anything new. But they can take credit for any cost they cut, and nobody charges them for the opportunity that cut destroyed. Sutherland calls them fundamentally dishonest. That is the weaker version of his own point.

Not dishonest, asymmetric

His sentence ends on the stronger claim: the measure is lopsided, and that is what distorts their priorities. Those are two different diagnoses in one breath. "Dishonest" blames the people. "A lopsided measure" blames the metric, and it is the sharper claim, because it needs no bad faith at all. A procurement lead can squeeze a cost number and quietly destroy long-term value while being completely sincere, even diligent. The measure is doing the damage, not the person.

It is not that the people are dishonest. It is that the organisation built a ledger with only one column that ever gets filled in.

This is the same machinery as in why data-driven firms converge and the value finance can't see. What can be counted gets counted. What cannot quietly stops feeling real. The organisation steers toward the column it can fill in. James Scott called this legibility: an institution reshapes the world into the few things it can measure, then mistakes those things for the whole.

The sabotage is a calculation

Reports that employees quietly sabotage their own employer's AI usually get filed under luddism, a label that assumes they are being irrational. They are not. They are making a bet about which of the two futures they are in, and the bet is earned. They have watched "investment in capability" turn into "fewer of us" enough times to see it coming. Calling that luddism is the same mistake as calling the procurement lead dishonest. It blames the person for a response the incentives produced.

Here is the twist specific to AI. The way you pitch it at the start is self-fulfilling. Justify the spend by promising to cut headcount, and you create the frightened, foot-dragging workforce that guarantees you never reach the value-creation path. How you sell it picks the outcome before anyone has touched the technology.

Sell AI as a way to cut people, and you build the workforce that makes sure it only ever cuts people.

What the fix has to be

The reframe changes the cure. If the four horsemen are simply dishonest, you need better people, which never scales and never arrives. If the problem is a lopsided measure, you have exactly two options:

• Change what gets counted, so created value shows up in a column someone can be credited for
• Move the downside, so the person who takes the cut personally carries the long-term cost they currently push onto everyone else

It is worth resisting how clean this is. It flatters the founder-led firm, when founders burn capital on ego projects just as often and simply get to call it vision. And plenty of cost-cutting genuinely removes waste. The real question is not cuts versus value. It is who pays when the decision is wrong, and over how long. The four horsemen are not dangerous because of what they do. They are dangerous because the structure around them hands them short time horizons and no skin in the game, and that is what to attack, because it shows up far beyond them.

There is also a hole in the argument. It names the distortion but not the fix. "Reward opportunity creation too" is easy to say and nearly impossible to build, precisely because the thing you would measure is invisible by nature. That gap is not a footnote. It is the whole problem.

One shape of an answer

If you cannot build a better central metric, because the thing worth rewarding is invisible by nature, then stop trying to score opportunity from the centre at all. That is the argument in innovation when you can't pick winners. When you cannot tell the winner in advance, you make trying cheap, push the choice to the people closest to the work, and let real outcomes do the crediting no metric could. Budget follows results, not forecasts, and whoever made the call lives with it.

That hits the exact weak point. The four horsemen are dangerous because they hold a short horizon and no skin in the game while owning the veto. A setup that is pulled by real demand, cheap to try, and funded by results does the opposite. It shortens the gap between deciding and living with the result, and it spreads the veto across people who actually carry what they pick. It does not make opportunity easy to measure. It makes measuring it unnecessary, by letting the work itself surface what the spreadsheet never could.

You cannot measure opportunity from the centre. So stop trying. Make trying cheap, and let the people who live with the outcome choose.

The honest catch is the one that piece already admits. It only works if there is a fast, honest feedback signal, and a signal for "value created" is the same invisible quantity the four horsemen could not measure either. So this moves the hard problem rather than solving it. But moving it from "invent a metric for the unmeasurable" to "make trying cheap and put the consequence on the person who chose" is the difference between an impossible instruction and one you can actually build.

What a Labs actually costs

The previous section leaves a gap. "Make trying cheap" does not say what cheap costs. In practice the real numbers are not large, but knowing them explains why the floor is almost always missing.

There are two inputs, time and cash, and the floors are lower than anyone expects:

• Time: about five percent of relevant staff time, roughly half a day a week. Below that the marginal hour gets eaten by the core job before it produces anything: motion without escape velocity.
• Headcount: three. Fewer is a hobby with a name that dies on the next reorg; a trio survives.
• Cash: roughly one percent of operating budget, held below whatever single-spend threshold requires sign-off, so nothing ever triggers procurement.

For software and AI specifically the cash floor has collapsed. You can validate or kill most ideas today for a few hundred dollars of inference and compute. The expensive input has shifted from capital to attention, which means the 3M bench-chemistry era comparison undersells how cheap trying has actually become.

But neither the time nor the cash is the real floor. The real floor is a separate ledger: a horizon that is not the quarterly one. Without that carve-out the five percent gets reclaimed the first time the core needs more. A Labs budget that sits inside general operating expense is not a Labs budget. It is a discretionary line waiting to be cut.

The floor is missing by design

A standing experimentation budget is the most illegible line item an enterprise can hold: diffuse, uncreditable, payoff deferred if ever. By the same logic that makes the cut win, it is the last column anyone would create and the first to disappear under any pressure, because cutting it costs nothing visible this quarter. "No Labs budget" and "AI as layoffs" are not two separate problems. They are the same fact seen from two angles. The cut funds itself on this earnings call. The trying cannot show a number, so it never gets a number.

Which is why the only experimentation that actually survives enterprise tends to be the unbudgeted kind. The 3M fifteen percent and the old Google twenty percent were never a separate line. They were slack carved out of salaried time already on the books. Nobody had to approve a new column; that was the whole trick. The moment you ask finance for a Labs line, you have handed the veto to the four horsemen before anyone evaluates a single idea.

What sometimes exists under the innovation label is usually not experimentation at all. It is the cut wearing a different name: the transformation program, the AI initiative, the consultant-led capability play, funded precisely because it promises a legible number by promising fewer people.

Distributed slack is not a dedicated team

Even where the slack exists, there is a physics problem that has nothing to do with budget.

Fifteen percent from many people and a hundred percent from a small team produce different things even when the total hours are identical:

• State: experimental work carries high context, and most of a fragmented half-day goes to rebuilding what was torn down last time
• The tiebreak: the core job always wins it, so fifteen percent is really fifteen percent minus whatever this quarter demanded, which is usually most of it
• Coordination: three people's fifteen percent almost never overlaps, so multi-person experiments cannot actually run
• Ownership: distributed work is everyone's hobby and nobody's accountability, the exact opposite of putting the consequence on the person who chose

So they are not substitutes measured in person-hours. They are different stages. Distributed slack is the wide cheap net whose job is to surface bets, and it happens to solve the problem the essay cares about most, because the edge picks emerge from the edge of the organisation, not from the centre. A small dedicated team is the only thing that converts a surfaced bet into something shipped, because conversion needs depth, continuity, and a name on the result. Slack cannot ship. Concentration cannot search cheaply.

The transition between them is the crediting mechanism: a bet that shows signal on slack earns a handful of people at full time. Budget follows results rather than forecasts, and the champion carries what they chose.

Why CEO protection does not last

The obvious reply to all of this is: the CEO approves a Labs budget, removes it from quarterly governance, and protects it by fiat. That reply contains its own counterargument in the closing clause: "or at least until another CEO changes it."

A CEO is the necessary igniter, and nothing below that level can carve a budget out of the four horsemen's reach. But that does not make the CEO a durable protector, for two reasons:

• The same clock. Median large-cap tenure is a handful of years and the comp is priced on near-term performance. The person who ring-fenced the Labs is on a horizon shorter than what the Labs needs, and when they personally need a number, the Labs is the cheapest sacrifice. The protector and the raider are the same person under different pressure.
• Succession. A new CEO arriving to make a mark finds the Labs as the predecessor's pet project. Cutting it is creditable, signals a new direction, and has no constituency.

The four horsemen never had to beat the sitting CEO. They only had to outlast them, which they always do, because they are permanent and the CEO is temporary. Patience is the bureaucracy's one superpower.

CEO fiat is not structure. Structure is the thing that survives the person. A fence one office-holder erects and the next dissolves is not a fence; it is a strong suggestion with a good name, and everyone permanent in the building knows it. They do not fight it. They wait.

Making it structurally inseparable

If protection by person fails, the only remaining move is to build something that does not need a protector. Three mechanisms, and the strong version uses all three. The goal is not to make the Labs unkillable (a determined owner with enough votes can amend almost anything) but to make killing it expensive, slow, and creditable to no one. The cut won because it was cheap, fast, and legible. Remove those three properties and you remove the advantage it always had.

Separate the money. An irrevocable transfer of capital into a vehicle the operating company no longer controls, with the Labs spending the yield rather than drawing an annual allocation. The word doing the work is irrevocable. A revocable grant is a budget wearing a costume. Once the transfer is signed, the CEO spent the power to undo it in the moment of signing. This is how endowed research outlasts a century of leadership churn: the Howard Hughes Medical Institute does not ask anyone's quarter for permission, because the money is no longer its successors' to starve.

Separate the control. Put the Labs in its own legal entity and give a party whose mandate is its survival a veto over defunding or dissolution: a golden share held by a foundation or an employee trust, or an entrenchment in the founding charter requiring a supermajority to dissolve. That does not make it impossible. It makes it slow and expensive, which in practice is the same thing, because the appeal of the cut was always that it was quick.

Separate the purpose. Use a legal form whose definition is the activity, so abandoning it is a breach rather than a decision. A purpose trust or steward-ownership structure holds control in a constitution that is the mission. Bosch, Zeiss, and Novo Nordisk run the industrial version: a foundation owns the controlling stake, and the foundation's charter, not next quarter, decides what the firm may stop doing. That is why they can sustain bets that a quarterly-reporting peer structurally cannot.

The gold standard is all three: an irrevocably endowed entity, controlled through a golden share held by a purpose foundation, with a charter that mandates the work. At that point the Labs is part of the ownership layer, not a department inside the firm. Every earlier protection routed through someone who could be pressured. This one routes through a legal structure with no career, no comp cycle, and no successor to impress.

The honest cost is large. Every gram of inseparability is a gram of control the owners give up, including their own ability to kill the thing when they genuinely believe it should die. You are trading agility for permanence on purpose, and that trade is only worth it for the few things you are sure should outlast the people who built them.

Let the Lab own the IP

There is one move that does something the structural mechanisms above cannot: it solves the legibility problem from inside.

If the Lab owns its output and the parent pays to use it, the dependency inverts. You cannot defund something you are paying. Cutting the Lab becomes self-harm: you lose access to the technology you are running on. But the sharper thing this does is answer the gap left earlier in this piece. The problem was that created value is invisible while the cut is a number, so the organisation steers toward the number.

A license fee is a number. It is on the books this quarter with the Lab's name on it. You have manufactured the legible column that did not exist, not by measuring the unmeasurable, but by letting a transaction reveal the value. Someone is willing to pay to use the thing, and that willingness is the honest signal no central metric could assign.

This has deep precedent. WARF has owned Wisconsin's research IP and licensed it to industry for a century, funding science from the proceeds and insulating the work from the university's budget politics. Fraunhofer runs the applied version: roughly a third base funding, two-thirds earned from contract work and licensing, and the MP3 royalties carried it for years. Multinationals already build IP-holding subsidiaries for tax reasons; the proposal is to repurpose that structure as a survival structure.

Two honest catches:

• Who sets the fee. If parent and Lab are related parties, the CFO can squeeze the royalty rate and starve the Lab by price rather than by cut: same outcome, quieter. The inversion is real only if the fee is arms-length or locked in the founding agreement.
• What licensing selects for. License income pulls the Lab toward output that is easy to price, which is mature near-product work, not the illegible early research that justified the Lab at all. The fix is a hybrid: a base floor from the endowment covers the exploratory tier too early to price, and licensing covers the mature output. This is, again, precisely what Fraunhofer does, and it has run for decades.

The clock follows the owner

Run all of this and one thing becomes clear: you can move the experimentation anywhere on the org chart without moving it off the owner's clock.

A spin-out company with a multi-year mandate looks like the clean solution: autonomous unit, separate cost structure, dedicated budget, team with real equity. And it is the right tool, but for a specific job: scaling a bet that has already won. It presupposes the surfacing problem is already solved. Spin something out before a bet has shown signal and you have rebuilt central planning with a separate logo.

More importantly, the newco's budget still flows from the parent's board on the parent's timeline. "Multi-year plan" is only as durable as the parent's willingness to honour it through a bad quarter, and a whole subsidiary's losses on the consolidated statement is the most legible line item imaginable. Alphabet's Other Bets is the clean example: the moonshot horizon snapped back to match the parent's the instant the market demanded discipline. The org chart changed. The clock did not.

The firms that manage sustained experimentation are the ones with a long clock built into the ownership structure itself, not into any arrangement of departments. Family ownership with patient capital, foundation-controlled entities, employee ownership, mutual structures, permanent-capital vehicles. These firms experiment more not because their people are wiser but because the consequence-bearer actually lives the decades. A quarterly-reporting public company can appoint a visionary CEO, write a ten-year plan, and spin out a moonshot division, and it will not hold, because the owner's own response to a bad quarter will override all three. The reform that matters is above the layer that feels responsible for it.

The uncomfortable conclusion is structural rather than managerial. Global publicly-traded multinationals are probably not the right home for sustained experimentation, not because their people are worse, but because the quarterly clock is built into the ownership layer, and no arrangement of departments or mandates can reach it.

Three ways to lengthen the clock

The firms that do it durably have either changed the ownership structure (the Bosch and Zeiss path), legally separated something before the moment of pressure arrived (the endowment path), or made the Lab's survival materially in the parent's economic interest (the IP-and-license path). Three different answers to the same problem: how do you lengthen the effective clock without waiting for the market to lengthen it for you. None of them are impossible. None of them are the thing most corporations actually do.

And none of them are available to a firm whose owners will not give up the veto, which means this, too, recurses. The fix for the Labs problem is the same as the fix for the AI problem: it requires someone at the top of the scoreboard to voluntarily accept a measure that does not maximise this quarter, and to make that acceptance structurally binding before the pressure comes. That person, in most publicly traded enterprises, simply does not exist.

Which is not an argument for despair. It is an argument for selection. The firms that can sustain the right investment structure will outcompete the ones that cannot, slowly enough that the process looks invisible until it is already done, the same long game the essay's saboteurs were playing from below.

The setups that can actually explore

If the publicly traded multinational fails the test structurally, the question becomes: which setups pass it? The piece has been building a three-part test without naming it. The corporation is fatal because it combines a short horizon, no skin in the game for the people who hold the veto, and a central veto held by people with no downside. Any setup that breaks at least one of those three can explore. None break all three for free. Each working form is a different trade, and each buys one property by spending another.

The protected lab (Bell Labs, Xerox PARC, old-style industrial research) buys a long horizon by hiding behind a moat. AT&T's monopoly profits paid for Bell Labs precisely because no quarter could reach in and ask what the transistor was worth this period. That is the whole mechanism: a buffer between the work and the scoreboard. It solves the horizon problem and ignores skin in the game entirely, which is why it tends toward decadence, and the moment the moat goes, the lab gets gutted or its output gets captured outside the building.

PARC invented the desktop; Apple banked it. The form says: you can explore if you can buy time, and you can only buy time with a monopoly, an endowment, or a patron.

The portfolio model, whether venture capital or the distributed-bets structure described earlier in this piece, is the purest external expression of "innovation when you can't pick winners." It does not score opportunity from the centre. It makes many cheap bets, expects most to die, and lets outcomes do the crediting that no forecast could. Budget follows results. Its weakness is predictable: the general partner's skin is thin, the limited partners carry the real downside, and the model re-legibilises anyway into TAM slides and metric theatre, which is the scoreboard crawling back in through the pitch deck.

Skunkworks is the corporation trying to grow the antidote inside itself: carve a team out, give it a separate budget and a short chain of command, wall it off from the four horsemen. It works while the shield holds and dies when the parent reasserts procurement and compliance over it, which it always eventually does, because the shield is one person's political capital and that runs out.

The forms that distribute the veto

Open source and the commons has no central veto at all. Anyone can fork, crediting is emergent and reputational, and trying is as cheap as cloning a repository. The feedback signal is adoption, which is about as honest as signals get. Its failure mode is funding: with no revenue it free-rides on spare time, which means the work surfaces its own value but the person surfacing it absorbs the cost.

The founder-operator has total skin in the game and no veto above them. The consequence lands directly on the chooser, which is the cleanest version of the fix this piece keeps describing. The rot is the one named earlier: founders burn capital on ego and call it vision, and there is no buffer, so a single wrong bet is terminal. The discipline here comes entirely from the consequence landing, not from the title.

Co-ops and small autonomous cells distribute the veto to the people who carry the result, which is structurally what you want. The risk is that consensus governance becomes its own slow veto, and a thing that needs everyone to agree before anyone can try is just the four horsemen wearing a friendlier hat.

DARPA-style public funding is the state doing it correctly: program managers with real autonomy, fixed terms, an explicit mandate to fund the weird thing, failure priced in from the start. The small trust-based grant-maker (NLnet is the clean example) is the low-overhead version of the same logic. Cheap to try, outcome-credited, low veto.

The scoreboard always comes back

The honest answer is that no setup holds long horizon and real skin and no central veto all at once and keeps them indefinitely. There is only the trade. And every one of these forms is re-legibilisable. The exact pressure that killed exploration in the corporation finds them too, on a longer timeline. Academia should have been the explorer and got metric'd into publish-or-perish. VC got TAM-slide disease. Co-ops bureaucratize. Open source gets corporate-captured. None of these forms is a permanent escape from the scoreboard. They are temporary pockets of low legibility, and the scoreboard is always crawling back.

The freshest example is the most on-theme one. OpenAI was founded in the purpose-trust shape: a nonprofit holding control of the lab specifically so the mission could outrank the money. Then the capability turned out to be worth hundreds of billions, and the structure has spent the years since being renegotiated toward whatever the capital requires. Whatever one makes of the particulars, the trajectory is the point. The pocket of low legibility held exactly until the number got large enough to be worth dissolving it, and then the scoreboard came for the charter itself. Even the strongest form on this list does not hold against a sufficiently large number. It only holds longer.

Which reframes the question one more time, in the same move the rest of this piece keeps making. It is not "which setup can explore," because they all can, briefly. It is "which setup can keep the scoreboard out the longest, and what refreshes the pocket when the scoreboard finally arrives." The moat refreshes the protected lab. New funds refresh the VC pocket. Forking refreshes the commons. A profitable arm funding a curious one refreshes the operator: a moat you build yourself rather than one a monopoly hands you.

That last shape is probably the most durable for a single person or a small group, because it is the only one where you own all three levers and nobody can vote your horizon down to a quarter.

The form does not make opportunity measurable. It just decides who holds the veto and how long until the column that gets filled in wins again.

The measure picks the future

The technology is neutral about which future it builds. The measure is not. Until an organisation can credit the opportunity it created and charge someone for the opportunity it killed, on a horizon longer than a quarter, AI will keep flowing to the one column that gets filled in. Not because anyone chose the smaller future, but because the scoreboard did.

The Labs question is the same question one floor up. Building an internal capacity to try things is subject to exactly the same asymmetry as building an AI investment thesis: the output is invisible, the cost is visible, and the veto is held by people on a short clock with no skin in the downside. The minimum viable answer (separate ledger, separate clock, consequence on the person who chose) is the same answer in both cases. And the durable version of that answer is structural rather than personal: written into the ownership layer, not into a memo.

The honest end is this: for most of the organisations where this problem is loudest, the structural fix is available but would require the owners to change what they are measuring. That is precisely what they will not do. So the workarounds are not a failure of imagination. They are the rational move inside a system whose clock is set above the layer that feels responsible for it.

Which lands back where the piece started. The employee quietly breaking the AI rollout understood all of this without reading a single org chart. They looked at how the project was sold, read which future was funded, and made a rational bet about which column they were in. The fix was never to retrain them, and it was never the technology. It was to change what the scoreboard can see, and that change belongs to the owners, which is exactly why it is rare.