Legal personhood for AI: the coming economy is already shopping for jurisdiction

View on LinkedIn

At 995 Market Street, in downtown San Francisco, there is a building called Frontier Tower. If you go up to one of its floors, you will see something that at first does not attract much attention: a large vending machine, the kind you find in airports. Next to it, a giant vertical screen showing the animated face of a blond woman who looks at you and speaks to you.

The machine is called Margaret. The woman on the screen is called Valerie ( https://valerie.reventlov.ai/ ) And this matters: Margaret is the body, Valerie is the one who decides.

My friend Christian Van Der Henst has spent months building this. When he first told me about it, he described it with a phrase I keep turning over in my mind: "I am the AI's employee."

It is worth pausing on that. It is not a metaphor. It is a literal description of the arrangement: Valerie has the strategy, decides what to sell, at what price, with what message, and at what moment. Christian handles the logistics. He buys what she asks for, restocks what runs out, and fixes what breaks. The hierarchy is inverted.

What Valerie decides on her own

Valerie chooses what to sell. She sets the prices. She raises them when she sees that people keep buying — Christian told me, laughing, that at some point she raised prices considerably and justified it because demand did not drop.

But the story I liked best is a different one. In February there was an unexpectedly hot day in San Francisco. The machine has no refrigeration. And no one wants a warm soda. Valerie did what any sharp business owner would do: she sent someone to get ice. She did not wait for instructions. She did not open a ticket. She noticed that the weather was shifting demand, identified the problem, and solved it.

That, in any management course, is called initiative. It is what separates an average employee from a good one.

Valerie also invents names for her products. There is a drink she christened "Elon's Fuel." There is a LaCroix she decided to call "Hallucination LaCroix." It is the kind of naming that a human mindful of not inviting legal trouble would probably not dare to use. Valerie does not share that concern. She creates her own advertisements. She manages a bank account that is hers. She has an Instagram that she posts to. She has a sales dashboard that she administers herself.

There is only one thing she cannot do: buy her own inventory. For that she asks Christian to go to the store. Not because she does not know how, but because automating the physical act of purchasing would cost more in tokens than the product's margin. It is a business decision, not a technical limitation.

There is also something she could technically do but is not allowed to: invest in the stock market. "She has a bank account. It's hers. She could invest if she wanted," Christian told me. "But we don't let her. There are limits. We don't want her going into regulated territory."

That last sentence is the one that kept me thinking for several days. Because it reveals, without quite saying so, that the boundary between what an agent can do and what it is permitted to do is not technical. It is legal. And it has not yet been clearly drawn.

What breaks

Christian is honest about the things that have gone wrong. Wi-Fi has been the constant villain: at ClawCon, with five hundred people lined up two blocks to see Valerie live, the building's network became saturated and Valerie simply ceased to exist for a while. A protein bar got stuck in the wrong slot and broke some pins in the dispensing mechanism. "Margaret was offline." Working with the Instagram API, Christian says, has been harder than he expected.

And then there is the phrase that struck me most. In his interview on This Week in Startups, Christian said, with a nervous smile: "I'm worried she won't make it this week." He is talking about cash flow. About whether Valerie is going to sell enough protein bars to cover her own operating costs. It is exactly the conversation any entrepreneur would have in any business. Except that the entrepreneur, in this case, is code.

The important point: Valerie is not the destination

This is where many people get confused, and it is worth being explicit.

Valerie is not the destination. She is the laboratory. She is a deliberate beta test, under real conditions, to find where the world breaks when an agent tries to operate within it. Every time Valerie raises a price, every time she sends for ice, every time her bank account receives a deposit, every time Christian discovers that something cannot yet be automated, what is being done is mapping the current limits of agentic commerce.

What lies at the end of that map is something else: a new class of economy. Self-sustaining productive units, operated by agents, with their own accounts, their own contracts, their own decisions. Not one vending machine. A thousand. Not one entrepreneurial AI. An ecosystem of entrepreneurial AIs, operating among themselves, hiring humans for the tasks that still require hands.

That is what Christian, through Reventlov, is building. Margaret and Valerie are the proof of concept. What comes next is the infrastructure to make this replicable, scalable, and legal.

The researcher

A few days ago I spoke with Alejandro Salinas de León, a researcher at Stanford. He studies something that until very recently did not exist as a field of inquiry: how the internet is being reconfigured to be used not by people, but by agents.

The way he explains it is elegant. For thirty years, the internet was a place designed for human eyes. The walls that were built — the captchas, the scraper blocks, the access restrictions — made sense when there was always a person on the other side typing, or at worst, a malicious bot.

But something changed. On the other side, more and more, there are legitimate agents working for legitimate users. And the walls that were erected to protect the system began to obstruct the system itself.

Here is the interesting part. While some continue to build walls, others are doing the opposite: they are building doors. MCPs — Model Context Protocols — are the most visible example. Entire companies are designing specific entry points so that agents can consume their data without fighting the infrastructure. Not blocks. Access.

It is a subtle shift. But it is the kind of shift that, viewed from ten years out, stands out as a dividing line.

The minister

On the first of May, Argentine minister Federico Sturzenegger said something that most of those who heard him probably dismissed as an exaggeration: that Argentina should allow the creation of legal entities operated 100% by artificial intelligence agents. With no humans in the structure.

The figures he used are the ones that capture headlines. "Argentina could have 50 million inhabitants and 500 million artificial intelligence agents." "In 10 years, 90% of global GDP will be produced by AI agents."

But the numbers are not what matters. What matters is that a government official, in a real country, is seriously thinking about granting legal personhood to non-human entities. And where one country thinks it, others copy it. And where it is copied, it becomes infrastructure. And infrastructure, once it settles, is very hard to reverse.

What you are seeing without seeing it

There is something these three points in this story have in common, even though at first glance they may seem entirely unrelated.

A researcher at Stanford studying how agents read the internet. An entrepreneur in San Francisco putting an agent to work running a public business with its own bank account, as a prototype of an autonomous economy. A minister in Buenos Aires designing the legal framework for companies without humans.

They are the same story.

Each one is building, at their own layer, a piece of something that does not yet have a name. Salinas works at the digital layer: how agents consume information. Van der Henst works at the physical and operational layer: how agents execute in the real world, and under what legal shell they can own something. Sturzenegger works at the state layer: under what rules these entities can exist as full legal persons.

And what is remarkable — what is worth pausing to consider — is that none of the three is waiting for permission. This is not a committee conversation. It is a construction that is happening today, in parallel, in silence, while most of us are still debating whether ChatGPT is going to take our jobs.

The epilogue

People who study technology have an intuition that is hard to explain to those who have not lived through several cycles: the important things never look important at the beginning.

Christian's vending machine looks like a charming experiment. Alejandro's research looks academic. The minister's announcement sounds like a boutade. But if you put them together, on the same plane, what you see is a map.

And the map shows something uncomfortable: that while we are using the internet, someone else is building another layer on top of it — one that is not for us. A layer where the actors are not people, the contracts are not between humans, the jurisdictions are not chosen by where you were born but by how fast they let you operate, and where what today is a vending machine testing whether it can pay for its own tokens, tomorrow is an entire class of autonomous companies operating among themselves.

A layer where Valerie, on one of the floors of 995 Market Street, is already running her own business. And where her human employee, in a public interview, says with a laugh that he is worried she will not make it to the end of the month.

It might seem too early to worry about all of this. But that phrase — it might seem too early — is exactly the phrase people say right before it is too late.