The Business That Runs Itself Will Answer to No One — Least of All You

There is a new thing being sold to companies, and it is not what you think it is.

It is not software that helps your staff work faster. It is not some jobs automated while the rest of the business carries on as before. Those things are real, but they are last year’s argument. This is different in kind.

The difference is in the relationship between the company and the machine. Until now, that relationship has been episodic: a person asks, the system answers, and the person remains the one who judges and decides. What is arriving is continuous: the system perceives, holds context, acts in real time, takes on longer tasks, and stays present across the whole flow of activity. Place that inside a company — wired to its workflows, its customers, its money and its authority — and the question is no longer whether the tool saves time. It is what human capability you are removing, what authority you are granting in its place, and who can still take control when the system is wrong.

What is being sold now is a system that runs the business itself. Not a tool that does a task inside your company but a system that holds the goals, allocates the resources, coordinates the functions, and makes the decisions that your executives and your board make today. It does not assist the people who run the company. It replaces the running of the company.

And it does not stop at one function. There are business cases being developed and presented right now — by the major consulting firms and systems integrators — that would take you, piece by piece, one rung at a time, moving inexorably toward one destination: a business run entirely by artificial intelligence. First the call centre. Then the back office. Then procurement. Then operations. Each step a sensible-looking efficiency, each one approved on its own merits — until, at the end, the business runs itself. Every function. Every decision. No people. Because at no point does anyone hand you the paper that says: this is the part we keep human.

And here is why it will not be stopped along the way. Each step is a sensible-looking efficiency, approved on its own merits. No single approval is ever large enough to force the board to confront the destination. If that direction is not identified and bounded at the outset, the board will only ever be asked to approve the next reasonable step — never the sum of them.

This is not the future. It is being built now, with real money, by serious people.

It would be comfortable to file this under science fiction. That comfort is not available.

The spending behind this is not a bet on a maybe. Worldwide AI spending is forecast at more than two and a half trillion dollars this year — larger than the economy of most nations.

Look at what the most serious institutions are actually building. OpenAI raised four billion dollars for a venture whose name states its purpose plainly: the OpenAI Deployment Company — a vehicle to put its systems into production inside your business. Anthropic is building a company for the same purpose, and its founding partners are not technologists. They are Blackstone, Goldman Sachs and Hellman & Friedman. When the biggest names in global finance put their capital into companies whose product is putting AI into the running of your business, the question of whether this is real has been answered.

The consultants are already inside the door, laying the foundation. The largest strategy firms are no longer selling advice and leaving. One of them now sells what it calls an intelligence layer — a system that ingests how your company thinks, so that its agents inherit your company’s own judgement. Read that carefully. The product is the extraction of how your business reasons, encoded into a system they build and hold.

And at the smaller end, it is already installable. A British company will build a fully autonomous platform and an agentic team that runs your operation. Live in months, not years.

Capability. Capital. Product. All of it present tense.

“But automation already works” — yes, where the world holds still

The sharpest objection comes fast, and it deserves a real answer: automation already runs whole operations. A modern factory produces cars with almost no one on the floor. If that works, why would a business that runs itself be any different?

Because a factory and a business are not the same kind of problem.

A factory works because it is a cage built to remove surprise. The car body arrives in the same place every time. The weld is in the same spot. The inputs are controlled, the tolerances fixed, the environment engineered so that nothing unexpected can happen. The robot is not exercising judgement. It is repeating a motion with great precision inside a world that has been stripped of variety. And crucially, the factory decides nothing. It does not set the price, read the contract, weigh the risk, judge the customer, or respond to the thing that has never happened before. Humans did all of that upstream, and the machines execute it.

A business is the opposite on every axis. Its world does not hold still. It faces customers, markets, rivals, regulators and events, none of which arrive in fixed positions, and its central work is the very thing the factory robot never does: judgement under uncertainty — deciding, weighing, choosing the exception, handling the case the rules did not anticipate. The factory automated the hands, inside a solved problem. The autonomous business proposes to automate the mind, inside an unsolvable one. Automation works where the world is fixed. A business is exposure to a world that will not stay fixed — and the thing being removed, human judgement, was the only part that could cope when something new arrived.

What you are actually buying, and what it costs you

Here is the part that should hold a board’s attention, because it has nothing to do with sentiment and everything to do with control.

An autonomous business cannot run on the software alone. The agents are generic. What makes your business yours is institutional knowledge — how your pricing really works, which customers get latitude and why, where the risks are buried, what the rules do not say but everyone who has been there ten years simply knows. None of that is in the system when it arrives. The system is empty.

So before it can run your business, that knowledge has to be taken out of your people and your processes and put into the machine. That is not a side effect of the build. That is the build. It is what you are paying the consultant to do: not to install software, but to extract the operating intelligence of your company and encode it into agents. The installation is, underneath, a knowledge-extraction exercise. Your company’s mind is the raw material.

And once it has been extracted, it is lost to you. The knowledge is not simply copied into the system. It is operationally transferred. The records may still exist on your side of the table. The living capability to interpret, challenge and renew them does not. That capability now lives only inside the system — a system built, held, and understood by the vendor, not by you.

You did not buy a capability. You paid to have the accumulated intelligence of your own company extracted, poured into a model, and carried out of the building — and then you dispersed the people who understood how to interpret it, challenge it and rebuild it. You now rent the inside of your own business back from the firm that hollowed it out, permanently, because you can no longer operate without them and can no longer rebuild what they hold.

When it fails, it will fail in a way you cannot stop

Think of how automated trading systems can interact in a financial market — each acting sensibly on its own terms, together producing a collapse faster than any human can halt, destroying enormous value in minutes, with no villain and no single broken part.

Now place that same dynamic inside your own company — its money, its customers, its supply chain, its commitments — with no human left in the loop to catch it. A business run by people fails in small, uncorrelated ways: one person errs, the next does not. A business run by a fleet of agents built on the same few models fails in the same way everywhere at once — the same blind spot, the same bad assumption, executed simultaneously across the whole operation.

Human fallibility is local and slow. Machine fallibility is systematic and instant. The friction you paid to remove — the slowness of people — was the interval in which someone could notice and stop it.

A business optimised to need no humans has not been made robust. It has been made perfectly brittle: exquisitely tuned for the expected case, defenceless against the one that isn’t. And it is always the one that isn’t that kills a company.

The people advising you are paid if you say yes

There is one more thing a board should see before it signs. The firm walking you up this ladder is increasingly paid on the outcome and partnered with the platform underneath — the conflict set out in The Board Signed the Contract. Here it has a further edge. The firm will not be there when the thing fails, because the entire promise was that no one needs to be. You are counselled toward the decision by the party that profits when it is made, and left to carry the liability by the party that will have moved on.

The staircase ends at your own chair

The logic that names the human “a bottleneck to be removed” has no natural stopping point on the org chart.

It takes the administrative class first — the clerks, the processors, the people who did the work. Then it climbs, because each removal justifies the next. The supervisors who managed them are redundant once there is no one to manage. Then the middle managers who coordinated the supervisors. Then the function heads. The management layer only ever existed to run the people below it, and as those people are removed, the management is pulled down behind them, tier by tier.

And the staircase has a top stair.

Follow it up and the last human layer left standing is the board. You are approving the removal of every person who understands your business — and the ingestion of the institutional knowledge that made it valuable into a system whose workings no one, not even the people who built it, can fully read — in an order that leaves you, the board, as the last humans in the building. Fully accountable. Entirely liable. And the only people left who cannot see inside the thing you are answerable for.

The board does not get automated away. It gets stranded — kept in place, holding every duty, above a business that has quietly become unknowable to it. The administrative class understood the work, and they are gone. The managers understood the coordination, and they are gone. What remains is a board, accountable for everything, comprehending nothing, because everyone who comprehended it was removed to build the thing it now cannot govern.

The account your shareholders never agreed to

So here is the decision, in the only terms it was ever justified in.

You were told this was an efficiency — a better margin, a stronger dividend, applause. What you are actually approving is a distribution of two outcomes from a single signature. On most days, the numbers look better, which is exactly why a board says yes. On the day it fails, the business does not have a bad quarter. It collapses — faster than you can act, in a way no one inside can stop, and afterwards no one can explain, because the understanding of how your own company worked no longer exists on your side of the table.

Will your shareholders welcome the rising dividend, or the uncontrollable collapse? They are the same decision. And you gave away the one thing — knowing how your own business actually runs — that would have let you tell, before you signed, which of the two you had bought.

You were not asked to approve a saving.

You were asked to approve a business that no one can see, no one can stop, and no one can answer for — climbing toward it one reasonable-looking rung at a time, sold by the people paid for every step.

The questions a board cannot evade

None of this requires a board to become expert in the technology. It requires two questions, asked before the next approval rather than after the failure.

The first: if the vendor disappeared tomorrow, could we still run this function safely within ninety days? The second: which people inside this company could explain, challenge, override and rebuild this operation if the system got it badly wrong?

If the honest answer to either is no one, then the board has not bought efficiency. It has accepted dependency — and it should approve that dependency knowingly, with its eyes open and its name against it, or not at all. The one thing it must not do is back into it, a reasonable step at a time, and discover only afterward that it can no longer take its own business back.