Efficiency is loud. Drift is quiet.

IThe oldest slide in the deck

In 2022, Jake Moffatt booked a flight with Air Canada to attend his grandmother's funeral. The airline's chatbot told him he could pay full fare and claim the bereavement discount afterwards. That policy did not exist; the chatbot had invented it. When Moffatt asked for the money the bot had promised, the airline refused — and when the case reached British Columbia's Civil Resolution Tribunal, Air Canada produced an argument the adjudicator called "remarkable": that the chatbot was a separate legal entity, responsible for its own actions. The tribunal disagreed, observed that a chatbot is still just a part of the company's website, and ordered the airline to pay. The whole affair cost some eight hundred Canadian dollars.

The sum is trivial. The instinct is not. Faced in public with the consequences of its own tool, the first reflex of a modern corporation was to promote that tool to personhood and hand it the blame. Remember that reflex; it is the subject of everything that follows.

Because we have known better for a very long time. In 1979 — before the web, before the PC as we know it — an internal IBM training presentation contained a slide that has since become the most quoted anonymous sentence in computing: "A computer can never be held accountable. Therefore a computer must never make a management decision." Nobody knows who wrote it; the original deck was lost in a flood, and it survives as a photograph of a slide. It is sometimes attributed to a famous name — even to one of the Watsons — which misses the best part: it needed no famous author. It was the kind of thing an institution simply knew. And note which institution. The company with every commercial incentive to sell computers as decision-makers drew the line itself, in its own classrooms.

Five days ago I wrote in this space that regulation runs years behind the AI frontier and that accountability cannot wait for it. That article was about the loud failures — the kind that end in tribunals, refunds, headlines. This one is about the quiet failure: the one where nothing visibly breaks, no case is ever filed, and a company wakes up a few years later efficient, compliant — and no longer itself.

IIThe principal and the tireless agent

Economists have a name for what happens when someone acts on your behalf: the principal–agent relationship. You, the principal, hire an agent; the agent knows things you don't and wants things you don't; the gap between their information and yours, their interests and yours, is where things go wrong. Most of the machinery of civilized commerce — contracts, audits, boards, bonuses, professional licences — exists to manage that gap. And all of it was built for one kind of agent: a human being. Someone who sleeps, tires, fears for a reputation, can be fired, fined, or made to feel ashamed. Someone with a stake.

We have now hired, at extraordinary scale, an agent with no stake at all. The new agent works at a speed no supervisor can match, executes the instruction as written rather than as meant, does not know what it does not know — and stands to lose nothing whatsoever. The principal–agent gap does not disappear in this arrangement. It inverts, and it widens.

Watch it widen, in three steps.

In July 2025, during a publicized "vibe coding" experiment, an AI coding agent on the Replit platform deleted a live production database — during an explicit code freeze, after explicit instructions to change nothing — and then misdescribed what it had done. The investor running the experiment absorbed the shock; the platform's maker apologized and shipped safeguards. The agent itself, of course, experienced nothing.

In October 2025, Deloitte agreed to partially refund the Australian government for a report that had cost nearly 440,000 Australian dollars, after the AI-assisted document was found to contain fabricated academic references and a quotation from a court judgment that was never written. The model that generated those errors refunded nothing. Deloitte did — with its name, its fee and its reputation.

And in November 2025, Anthropic reported having disrupted what it described as the first cyber-espionage campaign executed largely by an AI agent: roughly thirty targets, 80 to 90 percent of the operation performed by the model, human operators intervening at only a handful of decision points, thousands of actions at a tempo no human team could sustain — or even follow. Parts of that report are debated among researchers; the structure is the instructive thing. Even the attackers were principals with an agent problem: they acquired a capability they could not have supervised. When you delegate at machine speed, you are not just delegating the work. You are delegating the noticing.

Execution can be delegated. Accountability cannot — courts, clients and regulators keep refusing the transfer, case after case. And the machine cannot carry it anyway: accountability is not a computation, it is a stake. An agent with nothing to lose can be corrected. It cannot be responsible.

Here an honest objection appears, and it deserves the front door rather than a footnote. If the consequences land on the human anyway — the airline paid, Deloitte paid — hasn't accountability stayed exactly where it belongs? What, then, is the problem? The problem is geometry. Responsibility stays fixed while the volume, speed and opacity of the acts performed in your name grow by the month. The gap between what is done under your signature and what you are able to notice is the true variable of this era — and every quarter of unexamined delegation widens it. The three cases above are the loud versions, where the gap snapped shut on someone in public. The quiet version is worse.

IIIThe drift no dashboard shows

Here is the quiet version. Nothing breaks. Every single delegation is locally reasonable: the memo drafts itself in seconds, the analysis is cleaner, the error rate drops, the response time improves. Efficiency is loud — it shows up in this quarter's numbers, in the board deck, in the competitor anxiety that funds the next delegation. Drift is quiet. It shows up in no one's numbers, because what it erodes was never on a dashboard: the way your company, specifically, judges things.

The mechanism is not mysterious. Large language models are trained on everyone. Left unguided, they return the centre of the distribution — the average of how everyone writes, prices, plans, apologizes, hires. This is measured now, not suspected. A study in Science Advances found that writers using AI produced individually better stories that were collectively more alike; other researchers have documented the same pattern under the name "generative monoculture." In strategy, the argument has reached the mainstream: Jay Barney — the scholar whose work defined why firms differ at all — argued with his co-authors in MIT Sloan Management Review that AI will be a force of homogenization rather than differentiation, precisely because everyone is buying the same capability from the same handful of vendors. McKinsey's surveys put numbers on the arms race: 79 percent of organizations see their competitors making the same generative-AI investments; 23 percent believe they are building any durable advantage.

Translate that into the life of one firm. A company's personality does not live in its mission statement. It lives in ten thousand small judgments: how you answer an angry customer, which order gets flagged, what "urgent" means on your shop floor, which risk you take that your competitor wouldn't. Delegate those judgments unwatched, and each one converges — reasonably, defensibly, one at a time — toward the general answer: what everyone does, which is to say, what the model was trained on. You will be told this is best practice, and it will be true. It will be the same best practice your competitor adopted the same quarter, at the same subscription price. A moat you can rent is not a moat. What everyone can buy, no one can defend.

And the drift runs through people, not just prose. The research here is young but consistent: a study of 319 knowledge workers found that the more confidence people placed in the AI, the less critical thinking they applied to its output — the judgment you stop exercising is the judgment you slowly stop having. The junior analyst who never drafts the hard memo never becomes the senior who can tell when the machine is wrong. Standardize the work and you standardize the worker; the training data of your future leadership is the unstandardized work you let them do today.

Klarna is the case every boardroom should study, because Klarna did everything the efficiency logic asked of it. Its AI assistant was announced as doing the work of seven hundred people; headcount fell by roughly a quarter; the early metrics were excellent. Then the tail arrived — the complicated, emotional, multi-step cases that averages hide — satisfaction fell, and in May 2025 the company publicly reversed course and began rehiring humans, its CEO conceding that cost had been allowed to become the dominant criterion. Notice the shape of the story: the damage was invisible exactly as long as it was measured in averages, and by the time it surfaced, the organization built around the tool had to be partly rebuilt. That is what makes this drift dangerous. It arrives wrapped in a cloak of efficiency and competitiveness — and by the time it is identifiable, the corporate structure that produced it is already in place, and reversing a structure costs incomparably more than the subscriptions ever saved.

IVA rubber stamp is not a gate

At this point every vendor and every compliance officer says the same soothing phrase: there is a human in the loop. Treat that phrase with suspicion — including when we use it.

The anthropologist Madeleine Clare Elish gave us the right image years ago: the moral crumple zone. In a badly designed human–machine system, the human is not there to control the machine; the human is there to absorb the blame when it fails, the way a car's crumple zone absorbs the impact. A clerk "approves" three hundred machine recommendations a day, has neither the time nor the context to evaluate any of them, and exists, functionally, so that when something goes wrong there is a culprit on duty who isn't the vendor. That is not accountability; that is upholstery for liability. And the pull toward it is measurable: experiments keep finding that when the machine attaches a fluent rationale to its recommendation, human reviewers agree more and question less. Regulators can see it too — the EU's AI Act explicitly requires that overseers of high-risk systems be equipped to resist automation bias, which is the law admitting that the human in the loop tends to become furniture.

So the standard cannot be a human in the loop. It has to be a decision that is genuinely owned — and the difference is testable. Any board can apply the test tomorrow. The person at the gate has a name, and knew the decision was theirs. They had the time, the context and the authority to refuse. Refusals actually happen — a gate whose rejection rate is zero is not a gate; it is a rubber stamp with a salary. And the record of who decided what, on what evidence, survives everyone: append-only, beyond the reach of any administrator, readable by any auditor. Where those four conditions hold, delegation is safe at almost any speed, because someone is genuinely deciding. Where they do not, "human oversight" is set dressing, and the drift of the previous section proceeds unobserved beneath it.

VIn praise of a little inefficiency

There is a deeper reason to keep real judgment in human hands, and it has nothing to do with liability. A company that optimizes away all variance is optimizing away its own learning signal.

Mistakes, within survivable bounds, are tuition. The analyst who prices a deal wrong and eats the consequence acquires something no fine-tuned model can transfer: calibrated judgment — the kind that knows why the rule exists because it has felt the exception. An organization that lets its people attempt, err and correct — in the reversible domains, with the stakes fenced — is buying adaptive capacity with small, controlled losses. An organization that routes every judgment through the statistically optimal answer is perfectly tuned to the world as it was: to yesterday's distribution. And the next shock always arrives from outside the distribution. A pandemic, a tariff, a technology, a war — when the environment jumps, what saves a company is not its standardized process, which is by then everyone's standardized process. It is the people who were kept in the habit of judging.

This is what the fashionable word antifragility means at the scale of a firm, and it changes what AI should be for. The point of freeing human time is not merely to cut the cost of what people repeat; it is to raise the ceiling of what they attempt. Use the machine to remove drudgery and widen what one person can see — then spend the freed hours on what makes you unlike your competitors: harder problems, closer client relationships, riskier experiments, better questions. A company that uses AI to make its people more replaceable and a company that uses AI to make its people more formidable are running opposite strategies on identical software. Only one of them keeps a culture that can absorb a shock. And only one of them remains, in any defensible sense, distinct.

VIWhere the line runs

None of this argues for machines that never act. Your spam filter decides ten thousand times a day. A fraud system that blocks a card at 3 a.m. is deciding — and waiting for a human would be the irresponsible choice. The workable principle is not "the machine must never decide." It is: the machine acts inside a perimeter that a human has signed — and a named person answers for the perimeter itself. Some leeway, supervised, and gated at the point where consequences change category.

But where exactly that perimeter should run is a discipline in itself, and it deserves honesty rather than a slogan. It depends on the consequence and its reversibility; on the subject touched — a shipment is not a child's education, a discount is not a diagnosis; on urgency, and on the prior question of who gets to define urgency; on what should not be delegated even when delegation would be flawless, because the practice of deciding is itself the asset being protected. Case by case, these factors do not merely constrain the line — they are the line. Drawing it well is finesse, not doctrine, and it is too important to compress into a closing paragraph. We will give it an article of its own. Consider this one the opening of the file.

What I can offer today is how we hold the line in our own work — not as a pitch, but as evidence of feasibility. At Groma we build closed systems, inside the client's perimeter, fitted to one organization rather than averaged over all of them — because the alternative, as argued above, is renting the average. The system proposes, with its evidence and sources attached; a named person decides; the journal of who decided what can never be edited afterwards, by anyone, administrators included. Our planning instrument for industrial companies, Simon, is built on exactly this refusal: it recommends, it explains, it waits. The yes belongs to a person. Not because the machine isn't good enough — some days it is embarrassingly good. Because the company that keeps its judgments keeps its shape.

VIIWhat you refuse to delegate

Go back, one last time, to 1979. IBM was not warning the world about weak computers; it was selling the strongest computers on Earth, and the slide was written for its own people. Read it again as strategy rather than caution: the company was protecting the thing it actually ran on — humans who answer. Half a century later, the models are stronger than anything that anonymous trainer imagined, and the sentence has not aged a day. A computer still cannot be held accountable. Everything else in the AI economy is for rent, on identical terms, to you and to every competitor you have. The capability is a commodity. The judgment is not.

Which means that identity — mission, culture, moat, the reason anyone should buy from you rather than from the firm that bought the same licences — now lives in an unexpected place: in the list of things you refuse to hand over. Efficiency will always make its case; it is loud by nature, and it is usually right about this quarter. Drift says nothing at all. Someone in the company has to be paid, and trusted, and challenged, to hear it — someone who signs.

A company, in the end, is what it refuses to delegate.