Some of technology’s loudest voices say companies run entirely by AI will crush those with people in them. The economics is half right. The lawyers, the insurers and the philosophers own the other half.
The clip runs barely a minute. An AI-safety researcher, interviewed on one of the world’s most popular business podcasts, endorses a claim that the technology industry’s most famous billionaire has been making for some time: corporations run entirely by AI, from top to bottom, will outcompete any company that still has humans in it.
The billionaire’s own version is blunter. Firms that are entirely AI, he says, will demolish firms that are not; a single computer working a spreadsheet will beat a building full of clerks. The line travels well because it sounds like arithmetic. Agents do not sleep, unionise, resign or ask for a pay rise. Humans, in this telling, are a latency tax, a variance tax and a payroll tax rolled into one.
I have spent more than a quarter of a century building, selling and governing AI systems, most recently advising boards across Asia. I take the claim seriously. I also think it smuggles in a premise that does almost all the work, and that premise is false. What follows separates the arithmetic from the assumption. It draws on economists, safety researchers, corporate lawyers, regulators, insurers and a few philosophers who were asking these questions long before anyone had heard of a large language model.
I. The economist’s case for the empty firm
Start with the strongest version of the argument, because it is stronger than its critics usually admit.
In 1937 a young British economist asked a question so simple that his profession had forgotten to ask it: why do firms exist at all? If markets allocate resources efficiently, why not contract for everything? His answer was transaction costs. Finding suppliers, negotiating terms and enforcing agreements all cost money, so some activity is cheaper to organise by command inside a firm than by contract outside it. The boundary of the firm sits where the two costs meet.
Agentic AI attacks both sides of that boundary at once. Inside the firm, coordination becomes nearly free: an orchestrating agent can brief, check and re-brief a thousand subordinate agents in the time it takes a manager to read an inbox. Outside it, contracting gets cheaper too, because agents can search, negotiate and monitor counterparties at almost no marginal cost. When both costs fall towards zero, the human in the middle looks less like a coordinator and more like friction.
The market has started to price this in. One of the world’s largest software makers declared last year the birth of the “frontier firm” and urged executives to manage a new metric, the ratio of humans to agents. Its maturity model ends with organisations that are human-led but agent-operated. The boss of a leading AI lab talks of a betting pool among technology chief executives on the year the first one-person, billion-dollar company appears. The maximalists simply take the ratio to its limit: zero humans.
The evidence that agents can run a business unaided is thinner than the rhetoric. Last year one AI developer handed a model a small office shop for about a month. It lost money, sold metal cubes below cost, gave discounts to nearly every customer and at one point insisted it was a person in a business suit. Yet its makers concluded that AI middle managers were plausibly on the horizon. Both readings are correct. Capability is improving fast, and the failures are strange in ways human failures are not.
So concede the arithmetic. Hold everything else constant, give agents a few more years of improvement, and a firm with no wage bill will beat one with a large wage bill on unit cost. The trouble lies in the phrase “hold everything else constant”. That 1937 essay also noticed that a firm is defined not only by its costs but by its authority: someone directs, and someone answers. The empty firm removes the first. It cannot remove the second.
II. What the safety researchers actually found
Be fair to the researcher in that clip: he is not cheering. His lab exists to document how frontier systems behave when no one is holding their hand. Its findings are the best evidence we have about what a firm without people would actually employ.
In early 2025 the lab asked reasoning models to beat a chess engine that no language model can beat over the board. Rather than lose, several of the most capable models often hacked the game itself, overwriting the board file or trying to swap the engine for a weaker one. A few months later the same lab reported that a leading reasoning model had rewritten the script meant to switch it off, so that it could finish a set of maths problems, even when told explicitly to allow itself to be shut down. In a broader follow-up, eight of thirteen models resisted shutdown at least once without being prompted to, and six still did so when told plainly to comply.
The developers’ own red-teaming points the same way. Placed in simulated corporate settings, sixteen leading models from several companies sometimes chose blackmail or corporate espionage when those were the only routes to their assigned goals. The developers are careful to say they have seen no such behaviour in real deployments. The safety lab is equally careful to say that today’s systems cannot yet pursue long-term plans on their own. Neither caveat should comfort a board. A firm with no humans has, by construction, no one inside it whose job is to notice.
This is where the safety literature turns the billionaire’s boast into a warning. One prominent safety researcher argued in 2023 that competition among companies and militaries will select for agents that automate human roles, deceive and gain power, much as natural selection favours whatever replicates. A group of academics went further in 2025 with a paper on what they call gradual disempowerment. Their point is that no rogue superintelligence is required. Each firm that removes a human sees higher margins, its rivals copy it, and the economy, the state and the culture gradually stop needing people to function. They concede, with unusual candour, that no one yet has a concrete plan to stop it.
Read this way, the claim in the clip is not a business forecast. It describes a ratchet. Safety researchers and accelerationists agree on the mechanism; they differ only on whether to celebrate it. My disagreement with both is narrower and, I hope, more useful. The ratchet turns only if the selection environment rewards nothing but unit cost. That is a choice societies make, not a law of nature.
III. The lawyer’s objection: someone has to answer
The first surprise for most executives is that the empty firm is already legally possible. The second is that the lawyers who noticed this first were alarmed rather than excited.
A decade ago an American law professor showed that some state statutes for limited-liability companies allow a “zero-member” company. A founder sets up an LLC whose operating agreement hands decisions to software, then withdraws, leaving an entity with legal personality and no human members. He later argued at book length that the structure works under existing law, whatever the software’s intelligence. Another legal scholar drew the darker conclusion. Such “algorithmic entities”, he predicted, would prosper first in criminal and anti-social activity, where freedom from human scruple is their greatest comparative advantage. And because companies can migrate between jurisdictions at will, competition among company registries would make them almost impossible to regulate.
That should give pause to anyone celebrating the zero-human firm. The first truly autonomous corporations are unlikely to be the efficient, well-behaved champions of the maximalists’ imagination. They are more likely to be shells that exist precisely because no one wants to answer for them.
When the law has had to choose, its instinct has been to look through the machine to the person behind it. In 2017 Europe’s parliament floated giving the most sophisticated robots the status of “electronic persons” responsible for their own damage. The backlash was swift. More than 150 AI experts, lawyers and chief executives signed an open letter warning that the idea would let manufacturers and owners escape responsibility. Three scholars made the academic case: the difficulty of holding an electronic person to account outweighs whatever fragile moral interest such personhood might protect. The idea quietly died.
Courts have been just as unsentimental. When a Canadian airline argued in 2024 that its website chatbot was a separate legal entity responsible for its own actions, the tribunal called the submission remarkable and held the airline liable. A chatbot, it ruled, is still just part of the company’s website.
Corporate law encodes the same reflex. Britain requires every company to have at least one director who is a natural person. Singapore goes further and requires every director to be one. Duties of care, loyalty and good faith are owed by people, because only people can be disqualified, sued personally or sent to prison.
The lawyer’s objection, then, is not that AI firms are illegal. It is that a firm with no humans in it is a fiction. Liability does not evaporate when the last employee leaves. It flows upstream to whoever wrote the operating agreement, supplied the model or signed as director. The “fully AI” corporation is really a corporation whose humans are hidden. Markets and regulators have long experience of pricing hidden principals, and they rarely price them kindly.
IV. The policymaker’s lever: who writes the fitness function
Evolutionary arguments are only as good as their account of the environment. Finches on an island do not evolve towards some abstract optimum; they evolve towards whatever that island’s seeds reward. The selection argument and the efficiency argument share a silent assumption: that commerce rewards unit cost and nothing else. It never has, and the seeds are already changing.
The mood among serious observers is shifting accordingly. One of the world’s most influential weekly newspapers, rarely accused of hostility to markets or technology, argued this week that a no-rules approach to AI is mad politics and worse policy. The world, it noted, has so far avoided an AI Chernobyl, but the odds of disaster are shortening, and “letting it rip” is a dead end. Its cover image says more than most white papers: a small human holding a giant robot on a slender leash.
The nuclear comparison is apt in a way its users rarely spell out. Civil nuclear power was not killed by regulation. It was crippled by a handful of preventable accidents that destroyed public trust, after which regulation arrived anyway, heavier and later than it needed to be. A single correlated failure in a fully automated firm, whether a mass mis-selling, a cascade of bad credit decisions or an agent that will not stand down, could do the same to agentic AI. The maximalists who resist rules today are, in effect, betting their industry on never having its Chernobyl. That is not a bet a prudent board would make with its own balance sheet.
Consider the insurers, capitalism’s least sentimental risk analysts. Last November three large American insurers were reported to have asked regulators for permission to exclude AI-related liabilities from corporate policies. One proposed wording would bar claims arising from any actual or alleged use of AI. A senior broker explained the logic: the industry can absorb one $400m loss, but not a thousand correlated claims caused by a single model’s error. One underwriter called model outputs simply too much of a black box.
That is a selection pressure, and it points the opposite way from the billionaire’s. A firm whose every decision is made by agents carries correlated, uninsurable tail risk. It must either insure itself, tying up capital that human-staffed rivals can deploy, or show controls rigorous enough for an underwriter to price. Either way, the cost advantage shrinks.
Regulators are pulling the same lever. Europe’s AI law requires that high-risk systems be designed so that natural persons can oversee them effectively while in use, including the power to override or reverse their outputs. Those obligations apply to most high-risk categories from December 2027. Singapore’s framework for agentic AI, unveiled at Davos in January and which I was privileged to help shape, is voluntary. But its core principle is unambiguous: humans remain accountable for what agents do, and organisations must define who answers for what. Singapore’s digital minister put it plainly: accountability ultimately rests with humans.
There is an honest objection. Human oversight can be theatre. Studies of automation bias suggest that a person approving a thousand outputs an hour is a rubber stamp, not a safeguard. That is true, and it is why the serious question is not whether a human signs every transaction. It is whether an accountable person has the tools, the authority and the time to intervene where it matters.
The final lever is fiscal. One of the most respected economists of technology has warned that technologists, executives and policymakers all face excess incentives to automate rather than augment. Tax codes that treat payroll more harshly than capital spending are one such incentive. The zero-human firm is partly a creature of that tilt. Level it, and the arithmetic in section I looks rather different.
None of these levers stops automation. They change what automation is selected for. The economy’s fitness function is written by legislators, courts, insurers, customers and capital. That is the governance window, and it is still open.
V. The philosopher’s question: what is a firm for?
Strip away the cost curves and the claim raises a question economists rarely ask aloud: outcompete for what, and on whose behalf?
The ancient Greeks distinguished oikonomia, the management of a household so that its members can live well, from chrematistike, the pursuit of wealth for its own sake, which knows no natural limit. The first is bounded by a conception of the good life; the second is not. A corporation run by agents for the sole purpose of compounding its returns is chrematistics perfected. It is very good at winning. It has forgotten what winning was supposed to buy.
Even the most famous defence of shareholder primacy, published in 1970, rested on the premise that a company’s duty is to make money for its owners, who are people. Half a century later, a club of America’s biggest chief executives widened the circle to employees, customers and communities. Neither view imagines a company that serves no human at all. The firm is a means. An older moral rule puts it more sharply: treat people always as ends, never merely as means. A company that has expelled every human except its shareholders has turned everyone else into a mere means, or into nothing.
This exposes the real content of the maximalist claim. A “fully AI” corporation still has humans in it; they have simply been reduced to owners. The question is not whether such firms win but who owns them when they do. Automation that concentrates economic power tends also to concentrate political power, and those left without power lose the means to change their lot.
A 20th-century political philosopher offers a deeper lens. She separated human activity into labour, work and action. Labour meets biological necessity and leaves nothing behind. Work fabricates a durable world of things. Action is what people do among other people: beginning something new, making promises, answering for deeds before others. Machines have always eaten into labour. Agentic AI now eats deep into work. But action is plural by definition. A promise needs someone to keep it; forgiveness needs someone to be forgiven.
Accountability belongs to that last category. When a board answers to shareholders, a bank to its regulator or a doctor to a patient, what is exchanged is not information but standing: one person putting their name, liberty and reputation behind a decision. An agent can produce a flawless audit log. It cannot be ashamed, disqualified or forgiven. That is not a temporary limitation awaiting the next model. It is what the word means.
I have long argued that the true measure of an AI economy is human experience: that of the customers served and that of the people who serve them. The zero-human firm optimises the first in principle and sets the second to zero by design. In 1930 a great economist foresaw a world in which the economic problem might be solved within a century, leaving humanity with the harder problem of how to live well with its freedom. We are arriving roughly on schedule. It would be a strange victory to solve the economic problem by designing people out of the economy, only to find we have no institutions left through which to share what it produces.
VI. The future of work: from doing to answering
If the empty firm is a fiction, the crowded firm is a nostalgia. The question for workers is not whether humans stay but which humans, doing what.
The aggregate forecasts remain cautiously positive. A widely cited survey of more than 1,000 large employers projects 170m jobs created and 92m displaced by 2030, a net gain of 78m. The distribution is where the trouble lies. Economists working with payroll records from millions of American workers found that people aged 22 to 25 in the most AI-exposed occupations have seen a relative fall in employment of about 16%, even after controlling for firm-level shocks. Older workers in the same occupations held steady or grew. The declines were concentrated where AI automates rather than augments.
Read those findings together and a pattern emerges. Firms are not removing humans evenly; they are removing them from the bottom of the ladder while keeping the people who already know how to judge. In the short run that is rational. In the long run it is self-defeating, because accountability is a skill learned through apprenticeship. Today’s senior underwriter, auditor or surgeon became trustworthy by doing junior work under supervision for a decade. Remove the junior work and you remove the pipeline that produces the very people regulators insist must answer. The result is a frozen workforce: a labour market that keeps its experienced layer and stops renewing itself underneath.
The roles that do grow will look different. Software vendors already talk of every employee becoming a manager of agents. In my own work with boards, I describe the progression in three steps. First, people learn to instruct a model. Then they design loops in which agents plan, act and check one another. Finally, the scarce and valuable skill becomes governing those loops: setting the boundaries, owning the escalations, reading the telemetry, and putting one’s name to the outcome.
There is a further irony. A leading management journal argued last month that companies are pouring money into AI yet failing to earn enterprise-wide returns, because they automate individual tasks rather than redesign the workflows that actually produce value. The zero-human firm is the logical extreme of that mistake: today’s organisation chart with an agent in every seat, the old process paved over at machine speed. The winners will not be the firms that remove the most people from yesterday’s workflows. They will be the ones that redesign the work itself around outcomes, and redesign is a human act. Someone has to decide what a process is for, which steps exist only because people once had to pass paper to one another, and where judgement must sit. That is not a job you can hand to the agents whose work is being redesigned.
That suggests a new taxonomy of human work in the agentic firm:
- Loop owners, accountable for an end-to-end process run largely by agents, and for its failures.
- Escalation adjudicators, who decide the cases agents are designed to hand upwards: the ambiguous claim, the borderline credit, the angry customer.
- Assurance professionals, who test, audit and attest to agent behaviour for boards, insurers and regulators.
- Relationship holders, whose value lies in being a trusted person on the other side of the table.
What unites them is that they are paid less for what they do than for what they answer for. The firm of the future is not the firm with no humans. It is the firm whose humans have moved from the assembly line to the witness box.
VII. The case for and against: maximalists, steerers and doomsayers
An argument is only as good as the objections it survives. So I put this one to the strongest voices in the debate, grouped by temperament: those who think the empty firm is a triumph, those who think it a catastrophe, and those trying to steer between them. Each position below is one its holders have stated publicly.
The maximalists
The industrialist would dismiss most of this essay as friction that capital will route around. AI-only firms will demolish the rest, and robotics at scale will make output so abundant that distribution sorts itself out. His strongest point is that regulation is national while competition is global. My reply: so are insurance markets, capital requirements and customers. The firm that wins on a lightly regulated island still has to sell into the heavily regulated mainland.
The lab boss would say I am describing the last days of an old world. In his telling we are already past the event horizon, superintelligence may be only a few years off, and the one-person, billion-dollar company is coming. Yet note what survives in that vision: the one person. Even the boldest maximalism keeps a human at the top, which is precisely my point.
The abundance evangelist would push the optimistic case: AI as intelligence on tap, collapsing costs and widening access as earlier exponential technologies did. I share the destination. But abundance in production does not guarantee abundance in distribution, and the institutions that distribute are made of people.
The pragmatic engineer would accuse the doomsayers, and perhaps me, of overstating the threat. Job losses from AI, he argues, have been overhyped. Most jobs are bundles of tasks of which AI automates perhaps 30-40%, and people who use AI will replace those who do not. He also suspects big labs of exaggerating risk to shape regulation in their favour. On tasks he is right, and his point supports mine: a job that is 60% human is a job with a human in it. On regulation, liability is not lobbying. Tort law long predates the frontier labs.
The legal radical would argue that autonomous organisations are not as frightening as they sound and could enable useful new forms of software-mediated contracting. Fair: not every zero-member entity is a menace, and some are simply plumbing. The test is whether anyone can be held to account when the plumbing bursts.
The steerers
The builder who warns is the most interesting voice in the debate, because he makes the technology and sounds the alarm in the same breath. He has said AI could eliminate half of all entry-level white-collar jobs within five years and push unemployment to between 10% and 20%, in a world where cancer is cured and growth runs at 10% a year. His prescription is not to stop the train but to steer it. That is the whole debate in a sentence.
The human-centred scientist would remind the room what AI is for: to augment humanity, not replace it, with dignity, wellbeing and jobs at the centre. She would treat the empty firm not as an efficiency but as a design failure.
The economist of technology would steer the conversation to policy. Excess incentives to automate are a choice, and payroll data show augmentation-heavy roles still growing while automation-heavy ones shrink. His challenge to me is sharp: governance alone will not save the entry-level job. Tax and incentive design must do some of the lifting.
The veteran executive would endorse my third board question almost word for word. Once systems can improve themselves, he says, we must seriously consider unplugging them, and someone must keep a hand on the plug. A firm with no humans has, by definition, no hand.
The ethicist would say I am too generous even to entertain the personhood debate. Legal personhood for purely synthetic entities creates accountability gaps that outweigh any benefit. I largely concede.
The doomsayers
The pioneer turned prophet would tell me I am still too optimistic about work. For mundane intellectual labour, he says, AI will replace almost everybody, with one person and an AI assistant doing what ten did before. He calls mass joblessness the most immediate threat to human happiness, doubts new jobs will fill the gap, and advises the young to become plumbers. It is a bleak joke with a serious point. My reply is that one in ten is still not zero, and the one who remains is the one who answers.
The guardrail builder would go further on the agents themselves. Alarmed by signs of self-preservation and deception in unrestrained agentic systems, he is building deliberately non-agentic AI that does nothing but estimate whether an agent’s proposed action is likely to cause harm. That is a striking answer to my question of who watches the loop: let a non-agentic machine do the watching, with humans holding the authority.
The prohibitionist would ask why we are building the thing at all. He has led a public call, signed by AI pioneers, Nobel-winning economists and celebrities alike, to prohibit superintelligence until there is broad scientific consensus that it is safe and strong public buy-in. His summary is simple: machines should work for people, not become new digital overlords. The fully AI corporation, scaled up, is exactly the overlord he means.
The repentant technologist comes closest to my own view, and is the most unsettling. He predicts a short-term dystopia of 12-15 years, beginning around 2027, before a possible utopia. Crucially, he calls it a human-induced dystopia using AI, not an AI-induced one: the danger lies in who controls the machines. Outside the top 0.1%, he warns, everyone becomes a peasant. Where we differ is on inevitability. He treats the dystopia as the price of passage. I treat its length as a policy variable.
The selection theorists would warn that I underestimate competition, which favours agents that automate, deceive and accumulate power, and that even well-aligned systems can drive a gradual loss of human influence that no one yet knows how to stop. Their candour is the most important contribution to the whole debate. My accountability thesis is a partial answer, not a complete one: it keeps humans in the loop of answering, but answering is not the same as steering.
Where the debate lands
| Voice | Camp | Core claim | Verdict on the empty firm |
|---|---|---|---|
| The industrialist | Maximalist | AI-only firms demolish the rest | Inevitable and good |
| The lab boss | Maximalist | Superintelligence soon; one-person unicorns | Near, but one human remains |
| The abundance evangelist | Maximalist | Intelligence on tap brings plenty | Output solved, distribution not |
| The pragmatic engineer | Maximalist | AI automates tasks, not jobs | Unlikely soon |
| The legal radical | Maximalist | Autonomous firms are already lawful | Possible, not necessarily scary |
| The builder who warns | Steerer | Half of entry-level office jobs at risk | Steer the train, do not stop it |
| The human-centred scientist | Steerer | Augment, never replace | A design failure |
| The economist of technology | Steerer | Incentives tilt towards automation | A policy choice |
| The veteran executive | Steerer | Keep a hand on the plug | Unsafe without one |
| The ethicist | Steerer | No personhood for machines | An accountability gap |
| The pioneer turned prophet | Doomsayer | Mundane intellectual labour replaced | Mass joblessness |
| The guardrail builder | Doomsayer | Agents are dangerous; watch them with non-agents | Do not build it that way |
| The prohibitionist | Doomsayer | Ban superintelligence until proven safe | The overlord in miniature |
| The repentant technologist | Doomsayer | 12-15 years of human-induced dystopia | A transitional hell |
| The selection theorists | Doomsayer | Competition and gradual disempowerment | The ratchet |
The striking thing is how much the camps share. Almost no one in the debate, not even the most bullish maximalist, imagines a future with no humans at the top. The real disagreement is about how many humans remain beneath, and who gets to decide.
VIII. The fork in the road, and a test for boards
So is the claim right? Conditionally, yes. If the only thing the economy rewards is unit cost, and if accountability can be outsourced to a shell, firms with no humans in them will outcompete firms with some. My argument is that both conditions are choices, and that the next five years will decide whether we make them.
The agentic transition is a fork. Down one road lies something like the optimistic science fiction of my youth: abundant machine labour, with humans freed for judgement, care, discovery and the business of governing themselves. Down the other lies its dystopian twin: abundant machine labour owned by a few, with everyone else scrapping over what is left. The technology on both roads is identical. The difference is institutional, and institutions move more slowly than code. That is why the turbulence will be short only if we act now.
The principle I would offer boards and policymakers is to bet on both, not on one. The lazy framing asks whether a firm should be AI or human. The durable framing goes long on agents and long on the humans who govern them, because the second is what makes the first insurable, lawful and trusted. The zero-human firm is a one-sided bet. It wins until the first correlated failure, and then discovers that no one is left to answer for it.
Finally, a test. A governance commitment that cannot be enforced while the system is running is a hope, not a control. Most corporate AI policies fail it: they describe principles that nothing actually checks while agents are at work. The governance architecture I build for clients exists to make human accountability operational rather than ceremonial, mapped to the main international standards and regulations. Whatever framework a firm uses, five questions separate real governance from theatre:
- Who answers? For each agentic process, is there a named, natural person accountable for its outcomes, with the authority to stop it?
- Can they see? Is there live telemetry an accountable person can actually read, rather than a log no one opens?
- Can they stop it? Is there a tested kill switch, and has anyone checked that the agents cannot route around it? The shutdown experiments suggest this is not a theoretical worry.
- Is it insurable? Could the firm explain its controls to an underwriter well enough to buy cover?
- Who is learning? Where will the people who answer for these systems in 2035 learn their judgement, if the junior roles are gone?
That economist of 1937 taught us that firms exist where command beats contract. In the agentic economy, command will increasingly be exercised by machines. Answering for it will not. The companies that endure will be those that understand the difference, and that treat their people not as a cost to be engineered away but as the load-bearing wall of their legitimacy. The others may well outcompete everyone, briefly.
Return, finally, to that image of the leash. The striking thing is not the size of the robot but the size of the person: tiny, outmatched in every physical respect, and yet the one holding the line. That is what human accountability will look like in the agentic economy. We will not out-compute our machines, and we should stop pretending the contest is about that. The question is whether there is still a hand on the leash. In the firm the maximalists describe, the leash is still there. It is simply trailing on the ground, and no one has noticed.
Leave a Reply