Globalization Without Geography

When intelligence becomes abundant, agency becomes scarce.

For most of human history, intelligence had an inconvenient property: it was attached to people.

If you wanted a thousand hours of research, analysis, writing, negotiation, accounting, customer support, or software development, somewhere there needed to be humans willing and capable of performing roughly a thousand hours of work.

Economies became extraordinarily good at finding ways around this constraint.

Industrialization substituted machines for human muscle.

Globalization moved production to wherever combinations of labor, capital, and infrastructure were cheapest.

The internet made information almost instantly global.

Cloud computing made computation available on demand.

But knowledge work remained stubbornly human.

A company in San Francisco could move a customer-support operation to Manila. It could hire developers in Bangalore or designers in Warsaw. Zoom and Slack could make those workers feel closer.

But the company was still buying human time.

Artificial intelligence changes that constraint.

And agents change it even more.

AI allows intelligence to be produced without proportional human labor. Agents turn that intelligence into work: researching, reasoning, communicating, coding, operating software, making decisions, and increasingly completing multi-step tasks.

The result may resemble another enormous wave of globalization.

Except this time, the new supply of labor isn't coming from another country.

It doesn't have a country at all.

Agents are globalization without geography.

I. The First Labor Arbitrage

One of globalization's great economic forces was arbitrage.

A piece of work worth $100 an hour in one labor market might be performed for $20 somewhere else.

Companies reorganized around that difference.

Manufacturing moved. Then call centers moved. Then accounting, software development, design, financial analysis, back-office operations and increasingly sophisticated knowledge work became internationally tradable.

Economists Richard Baldwin and Jonathan Dingel have described a related phenomenon as “telemigration”: workers in one country performing tasks in another without physically migrating.

Remote work reduced geography's importance further. A worker no longer necessarily needed to sit in New York to perform knowledge work for a New York company.

But there was still an important floor underneath the system.

Someone had to perform the work.

If an organization needed another 10,000 hours of analysis, somebody had to supply those 10,000 human-hours.

Globalization changed which human performed the work.

AI potentially changes whether a human needs to perform it at all.

That is a fundamentally different form of arbitrage.

The old equation was:

Expensive human labor → cheaper human labor

The emerging equation is:

Human labor → machine intelligence

And unlike workers, machine intelligence is not geographically constrained in the conventional sense.

It doesn't need to immigrate. It doesn't need to commute. It doesn't need a local labor market.

And, most importantly, software can be replicated.

Discovering an extraordinary analyst does not allow a company to instantiate 10,000 copies of that analyst.

Software changes the supply curve.

II. From Intelligence to Labor

There is an important distinction between generative AI and agents.

Generative AI makes cognition cheaper.

Agents make cognition economically deployable as labor.

A chatbot can answer:

“Which flight should I take?”

An agent can potentially research flights, inspect your calendar, understand your preferences, compare prices, choose an itinerary, log into the airline, use your loyalty account, request permission to spend $1,200, book the ticket, add the flight to your calendar, send the itinerary to whoever is traveling with you, and check you in tomorrow.

The first system produces information.

The second performs work.

That transition matters because the economic unit has changed.

OpenAI's 2026 research on agent use describes this as moving from individual interactions toward delegated, long-horizon tasks: work that can continue for minutes or hours while the human does something else.

The consequences are already visible in simpler forms of AI-assisted work.

In a study of 5,179 customer-support workers, Erik Brynjolfsson, Danielle Li, and Lindsey Raymond found that access to a generative AI assistant increased productivity by approximately 14 percent overall. The gains were much larger—34 percent—for novice and lower-skilled workers.

One interpretation is profound:

AI wasn't merely generating words faster.

It was distributing some of the practices of highly experienced workers across the rest of the organization.

Expertise was becoming cheaper.

That is still primarily augmentation.

Agents suggest something more radical: delegation.

Once software can accept a goal, reason about it, use tools, operate software and continue until the goal is accomplished, companies gain access to something resembling a new supply of knowledge labor.

III. The Price of Intelligence

The easiest prediction about AI is that cheaper intelligence means companies will spend less money performing the same amount of knowledge work.

That prediction may be wrong.

History contains a useful paradox.

When improvements in steam-engine efficiency reduced the amount of coal necessary to produce a unit of power, economist William Stanley Jevons observed that total coal consumption did not necessarily decline.

Cheaper useful energy made energy economical in more places. Demand expanded.

Something similar could happen with intelligence.

Consider how much analysis organizations don't perform today.

A company doesn't assign 100 analysts to evaluate every sales lead. A founder doesn't commission a market study before every minor product decision. A family doesn't employ an economist, travel agent, nutrition researcher, personal shopper, accountant and administrative assistant around the clock.

Not because those services would provide zero value.

Because the value doesn't justify the cost.

Lower the cost sufficiently and the calculation changes.

Imagine that a piece of research requiring five human hours costs $500. You perform it only when you expect the answer to be worth more than $500.

Now imagine an agent can perform useful equivalent work for $5.

Thousands of previously irrational tasks become rational.

This produces two curves moving in opposite directions:

Cost per unit of intelligence ↓

while

Demand for intelligence ↑

There are already hints of this phenomenon in agent economics. Gartner recently forecast that inference cost per agentic workflow could increase more than fivefold through 2028 even while underlying model economics improve. Its reasoning is straightforward: cheaper intelligence enables much more complicated workflows consuming dramatically more inference.

In other words:

Tokens can get cheaper while tasks get larger.

Imagine an AI task costs $0.05 today. Five years later, a substantially more capable agent might spend $2 completing a complex task autonomously.

The cost of the machine task has increased fortyfold. But suppose equivalent human work costs $200.

The agent is simultaneously 40× more expensive than its primitive predecessor and 99% cheaper than the human alternative.

The relevant question therefore isn't simply: “How cheap will AI become?”

It's: “How much intelligence will civilization consume once intelligence becomes cheap?”

The answer could be: an extraordinary amount.

IV. The Million-Worker Company

That leads to a strange possibility.

Companies may employ fewer people while simultaneously using vastly more labor.

Imagine a software company in 2020: 500 employees, collectively performing roughly one million working hours annually.

Now imagine a company in 2035: 50 human employees, each coordinating persistent AI agents.

Sales agents research every prospect. Engineering agents continuously test software. Security agents attack the company's own infrastructure. Finance agents model thousands of scenarios. Support agents investigate every complaint. Research agents monitor competitors. Procurement agents continuously negotiate prices.

And beneath those persistent agents sit millions of ephemeral workers instantiated for individual tasks.

The company might contain fifty humans and effectively consume the cognitive output of tens of thousands of workers.

The traditional organizational chart becomes misleading.

Human headcount falls.

Machine headcount explodes.

Revenue per human employee could become enormous.

This isn't merely replacing workers. It's changing what a company is.

The corporation begins looking less like a collection of employees and more like a small number of humans allocating enormous amounts of machine intelligence.

The scarce skill becomes orchestration.

V. The Scarcity Inversion

And this is where the argument becomes more interesting.

If intelligence becomes abundant, intelligence becomes less differentiating.

Imagine every company has access to extraordinarily capable models. Every company can generate copy, analyze documents, write software and reason.

Then having access to intelligence alone isn't much of a moat.

The scarce resources move elsewhere.

An intelligent system might know exactly how to purchase a house. That does not mean it can purchase one for you.

It needs to know: Who are you? What do you actually want? What can you afford? What accounts belong to you? What information may it disclose? What contracts may it sign? What money may it move? When must it ask permission? Which institutions recognize its authority? And who is responsible when something goes wrong?

These questions reveal the next layer of the agent economy.

Context

Intelligence without context is generic. The valuable agent understands your history, preferences, relationships, obligations, goals and current state.

Identity

An agent needs to know whom it represents. The internet was largely built around humans authenticating themselves to software. The agent economy requires software to act on behalf of authenticated humans and organizations.

Credentials

Knowing how to perform a task is useless if the agent cannot access the systems required to perform it.

Permission

Credentials answer can this system access something? Permission answers is it authorized to perform this particular action? Those are different questions.

Capital

An agent capable of economic action eventually encounters money. Buying. Selling. Paying. Depositing. Investing. Negotiating. Moving value becomes part of the execution layer.

Distribution

An extraordinarily intelligent agent trapped inside a chat window has limited economic power. It needs access to browsers, applications, APIs, marketplaces and eventually physical systems.

Trust

Humans must become comfortable delegating consequential decisions. Companies must accept machine-originated actions. Institutions need ways to distinguish legitimate delegation from fraud. Regulators need accountability. Agents need limits.

Which produces what I think is the central economic inversion of the agent era:

Intelligence becomes abundant. Agency remains scarce.

VI. Intelligence Is Not Authority

This distinction is easy to miss because intelligence and agency have historically been bundled together.

A capable employee usually comes with both.

Hire an accountant and you're buying accounting knowledge plus some authority to operate on your company's behalf. Hire an executive and you're purchasing judgment plus significant organizational authority.

Software separates these things.

A model can possess extraordinary knowledge while possessing zero authority.

It can understand precisely how to wire $50,000 without being allowed to move a dollar. It can know exactly which flight you would prefer without having your passport number. It can determine that a contract should be signed without possessing the legal authority to sign it.

So we arrive at a useful distinction:

AI commoditizes intelligence.

Agents commoditize portions of labor.

But neither automatically commoditizes authority.

Authority has to be granted. Scoped. Authenticated. Revoked. Audited. Trusted.

This may become one of the most valuable layers of the agent economy.

VII. The New Economic Stack

The internet economy was organized around information.

The agent economy may be organized around action.

Its stack could look something like this:

INTELLIGENCE — Models reason.

CONTEXT — Systems understand the person or organization.

INTENT — Systems determine what outcome is actually desired.

IDENTITY — They establish whom they represent.

PERMISSION — They establish what they're allowed to do.

EXECUTION — They interact with software and institutions.

CAPITAL — They deploy economic resources.

VERIFICATION — They prove what happened.

ACCOUNTABILITY — Humans and institutions can inspect, revoke and correct their actions.

Models sit near the beginning of this chain. Economic outcomes sit at the end.

As model capabilities converge and intelligence becomes cheaper, an increasing share of economic value may migrate toward the infrastructure connecting the two.

VIII. The Argument Against This

There are good reasons this future might arrive more slowly than its proponents expect.

First, productivity improvements at the task level do not automatically become productivity improvements at the company or economy level.

A 2025 field experiment across 66 firms and more than 7,000 knowledge workers found that workers given generative AI saved roughly two hours per week on email, but researchers did not observe major changes in the quantity or composition of their work.

And new 2026 NBER research characterizes current workplace generative-AI adoption as widespread but shallow: AI appears across many occupations and tasks, but within most tasks fewer than half of workers use it.

There is an enormous gap between AI can perform a task and organizations reorganize themselves around AI performing that task.

Second, agents make mistakes. A hallucinated paragraph is annoying. A hallucinated wire transfer is catastrophic. Reliability requirements rise with consequence.

Third, many jobs aren't bundles of easily separable tasks. They contain relationships, accountability, tacit knowledge, politics, taste and human trust.

Fourth, regulation may intentionally preserve human bottlenecks. Societies may decide that certain decisions—medical, legal, military, financial or governmental—require humans even when machines become technically capable.

Finally, the physical world remains stubbornly scarce. AI cannot reason its way into infinite energy, semiconductors, land, minerals, factories or housing. In fact, exploding demand for machine intelligence may make some of these inputs more valuable.

So this thesis does not require believing that humans disappear from the economy.

It requires something narrower:

The marginal supply of useful cognition becomes dramatically more elastic than it has ever been.

That alone would be transformative.

IX. What Would Prove This Thesis Right?

Over the next decade, several measurable things should happen if this argument is correct.

Revenue per human employee should increase substantially among AI-native companies.

The ratio of machine tasks to human tasks inside organizations should explode.

Agent usage should expand beyond tasks currently performed by employees into tasks companies previously considered uneconomical to perform at all.

Organizations should spend more—not necessarily less—on aggregate machine intelligence even as the cost of individual inference operations declines.

Model intelligence should become less differentiating as comparable capabilities diffuse across providers.

Meanwhile businesses surrounding identity, memory, permissions, payments, verification, security and agent infrastructure should become disproportionately important.

And perhaps most importantly, management itself should change.

The defining capability of an extraordinary worker may gradually move from How much can you personally produce? toward How effectively can you direct intelligence?

X. Globalization Without Geography

The industrial revolution dramatically reduced the price of physical power.

Globalization dramatically expanded the supply of economically accessible human labor.

The internet dramatically reduced the price of information distribution.

AI is beginning to reduce the price of cognition.

Agents may reduce the price of knowledge labor.

Each abundance creates another scarcity.

Cheap transportation made location more valuable. Cheap information made attention more valuable. Cheap software made distribution more valuable.

And cheap intelligence may make agency more valuable.

The first globalization asked:

Where can this work be performed most efficiently?

China? India? Mexico? Vietnam? Eastern Europe?

The next globalization asks a different question:

What can perform this work most efficiently?

A person? A model? An agent? A swarm of agents supervised by one person?

For the first time, the cheapest source of knowledge labor doesn't necessarily have a nationality.

It doesn't need to cross a border. It doesn't need to move anywhere.

It simply needs compute, context and permission.

The first globalization moved work around the world. The next may remove geography from work altogether.

And when intelligence is available everywhere, almost instantly, at increasingly negligible marginal cost, the defining economic question will no longer be who has access to intelligence.

It will be:

Who has the authority to turn intelligence into action?