AI will eat the administrative economy

For years the argument has sounded reassuringly human, and it badly misunderstands why so many people are talking to each other at work in the first place.

People ask other people for status updates because the information is fragmented. They schedule meetings because nobody has a shared understanding of what is happening. Managers collect reports because systems cannot interpret their own state. Project leaders chase people because workflows cannot coordinate themselves. Developers receive specifications because the person who understands the problem cannot directly produce the software.

We have built enormous organizations around the cost of moving information between human beings.

AI attacks that cost directly.

When people hear administrative work, they tend to imagine secretaries, assistants, payroll clerks and back-office staff. That definition is far too narrow.

Consider what many project managers actually spend their days doing:

Now consider many middle managers. The formal job description might say leadership, strategy and personnel development. The working week can look rather different: meetings, reporting, approvals, coordination, escalation, more meetings, and then a meeting discussing what happened in the previous meetings.

These people may be intelligent, capable and highly educated. That is beside the point. The question is not whether the person is capable. The question is whether the work requires a person.

A surprisingly large part of the white-collar economy consists of humans acting as routers between other humans. That is administrative work, whatever appears on the business card.

I used to run a development organization with programmers and team leads, and that structure required constant communication. I would explain what needed to be built. Someone would translate that into development work. Developers would interpret it. Questions would come back. Meetings would happen. Work would be reviewed. Misunderstandings would be corrected. New instructions would be written.

Every handoff introduced latency, and every handoff introduced information loss.

Today I work very differently. I operate as the architect and direct multiple Claude Code sessions and agent swarms at the same time. I am not writing specifications for developers, I am hardly writing code myself, and I do not spend hours explaining the same architectural intent through several organizational layers. I describe the outcome, inspect what comes back, correct direction where necessary and continue.

Twenty parallel AI sessions do not need a stand-up meeting. They do not need a project manager to ask whether the ticket is still on schedule. They do not forget what was decided yesterday because nobody bothered to write meeting notes. And they do not spend three days waiting for another department to answer a question that could have been resolved in thirty seconds.

This is not a speculative version of work in 2040. It exists now.

Previous generations of enterprise software were also supposed to transform organizations, and they did, eventually. But the implementation itself was often brutally expensive.

Replacing an organization's systems could mean a multi-year ERP project involving consultants, integration teams, process mapping, migration programs, training programs and enough PowerPoint decks to qualify as a secondary construction material. Large implementations consumed years before producing meaningful value, and that acted as an enormous brake on automation.

The economics are changing, because software itself is becoming dramatically cheaper to create and modify. An organization increasingly does not have to choose between accepting a terrible standard system and launching a five-year custom software program. Systems can be built around workflows instead of workflows being distorted around systems, and requirements can become software almost continuously.

Once software production becomes cheap enough, an enormous amount of previously tolerated organizational inefficiency becomes economically indefensible.

There will always be circumstances where human interaction has real value. Friendship has value. Trust has value. Negotiation has value. Leadership has value. A doctor explaining a frightening diagnosis is not equivalent to a checkout machine scanning milk.

But we should stop pretending that every human interaction in the economy exists because customers cherish human connection.

Bank branches are the useful lesson. People once did their routine banking through tellers because that was how banking worked. Then ATMs, internet banking and mobile banking made most of those interactions unnecessary. Customers did not revolt because they desperately missed discussing account transfers with strangers behind a counter. They adopted the more convenient system.

The same happened with travel booking, with e-commerce, with grocery checkout and with food delivery. Nobody orders a pizza through an app because they are hoping for a richer interpersonal relationship with the delivery driver. They do it because pressing three buttons and having food appear at the door is convenient.

People often describe consumer behaviour as if values operated independently from economics. In reality, behaviour reveals trade-offs.

Someone might genuinely prefer the charming neighbourhood shop where the owner knows every customer by name. But if the groceries there cost 40 percent more and the shopping takes twice as long, that preference suddenly has competition.

There will always be premium niches built around craftsmanship, personal attention and human experience. That is different from the mass market. Mass markets are ruthless optimizers: if two services produce roughly the same outcome, the cheaper and easier one has an enormous structural advantage.

So AI does not need to become more pleasant than humans. It needs to become sufficiently good, dramatically cheaper and instantly available. That is a much lower bar.

None of this implies that AI will simply take all jobs. Jobs are bundles of tasks, and those bundles will be dismantled unevenly.

The most vulnerable work has a recognisable shape:

That description covers the obvious administrative jobs. It also covers substantial portions of project management, middle management, finance, HR, procurement, consulting, legal operations and software development. The title may survive long after most of the original job has disappeared.

A department that once required fifty people might require twelve. Then six. Perhaps eventually two people supervising a large automated system.

That distinction matters, because AI does not need to automate 100 percent of a job to destroy most of the employment associated with it. If one AI-equipped person can produce the output that previously required ten, nine jobs have become economically questionable even though a human remains involved.

The organizational chart of the twentieth century was partly an information-processing architecture. Workers reported upward. Managers aggregated. Senior managers received compressed versions. Decisions moved back down through the same hierarchy. Each layer existed partly because humans have limited bandwidth.

AI changes that constraint. An executive or a domain owner can hold direct visibility into thousands of processes, because agents continuously collect state, detect exceptions, evaluate conditions and execute routine actions. Instead of information moving through five layers of humans, the system itself maintains the operational picture.

That makes a great deal of middle-layer coordination unnecessary.

Future companies may therefore look surprisingly small. Not because they accomplish less, but because each human controls vastly more productive capacity. A company employing twenty people might produce what previously required two hundred, and a company employing two hundred might compete with one that previously employed several thousand.

Headcount could stop being a symbol of corporate success and start looking like evidence of organizational inefficiency.

Predictions about technological displacement usually assume slow adoption, and historically that was reasonable. Organizations moved slowly because changing software was expensive, risky and painful.

But what happens when the technology causing the disruption also dramatically reduces the cost of implementing it? That is the unusual feature of this moment. AI writes the software. AI helps migrate the data. AI interprets the old systems. AI trains the users. AI monitors operations, and AI modifies the new system when the requirements change.

The mechanism that automates administrative work simultaneously accelerates the deployment of automation. That is a feedback loop, and it is why forecasts stretching this transformation over fifteen or twenty years may prove far too conservative. The technology does not need fifteen years. Much of it already works.

The remaining question is how quickly organizations discover what their competitors are doing, and markets have a reliable way of accelerating reluctant executives. If one company finds it can operate with half the administrative overhead of another, the second company does not get to preserve its organizational traditions indefinitely.

Its cost structure eventually makes the decision for it.

My expectation is that the administrative layer of the economy shrinks dramatically. Not to zero, but enough to change what a company fundamentally looks like.

Routine knowledge work becomes mostly machine work. Humans stay concentrated around ownership, judgment, relationships where relationships genuinely matter, physical activity, unusual situations, creative direction, risk and final accountability. Individuals increasingly operate fleets of agents rather than teams of human intermediaries.

Software becomes less like a collection of applications people manually operate and more like an operational system continuously acting on behalf of the organization. The interface itself may become secondary: instead of humans opening software to perform administrative work, the software performs the work and brings a human into the loop only when a genuine decision is required.

That is a much bigger change than adding an AI chatbot to Microsoft Office. It is the removal of an enormous amount of organizational machinery that exists because humans have historically been bad at coordinating information at scale.

The uncomfortable implication is obvious. Millions of intelligent people currently earn their living performing work that exists largely because information systems have been inadequate. The people are not useless. The machinery around them is becoming unnecessary.

And once companies can replace expensive, slow administrative coordination with systems that run continuously for almost nothing, capitalism is unlikely to preserve that machinery out of nostalgia.

The meetings will not disappear because somebody finally convinces corporations that meetings are annoying. They will disappear because nobody will be willing to keep paying for them.