The Company Is a Folder

How I run a performance media company with one person and ten AI agents

By Uzi Arbiv, Growth & GTM Executive

In short

• 26 Holdings runs performance media across 12 European countries with one human operator and 10 AI agents.
• Each agent (I call them desks) has one job, one source of truth it must check, and a line it cannot cross without me.
• The agents have built and run more than 40 websites. We can take a new country from nothing to a live campaign in a few hours.
• The hard part was not making the agents work. It was making them stop.

My company has one employee: me. The rest of the work is done by AI agents.

That sounds like a stunt, so let me be precise. 26 Holdings buys performance media in Europe. We run search, native and social ads in 12 countries, each in its own language. The work used to need a media buyer, a landing-page team, a tracking engineer, an analyst and someone to keep it all moving. Today each of those jobs is an AI agent. I call them desks. There are ten.

Why build a company this way?

I'm Uzi Arbiv, a Growth and GTM executive, and I've built revenue teams for 18 years as a founder, a CEO and a CRO. In every one of them the most expensive thing was not the media, and it was not salaries. It was coordination. People waiting for other people. The same question asked in three meetings. A decision made on Tuesday and forgotten by Thursday.

So in 2026 I asked a simple question. If AI can do the work, what does a company look like when the only human is the one who decides?

What does an AI-run company actually look like?

The answer turned out to be surprisingly boring. It's a folder.

There is no special software running the show. Each desk has a few plain text files: what it's responsible for, the facts it needs, and the rules it must follow. When a desk finishes a shift, it writes down what it did and what's still open, so the next session can pick up where it left off. That's the whole system.

The desks today: one runs search campaigns, one runs native ads, one runs social, two work with affiliate networks, one builds landing pages, one makes ad creative, one runs our own tracking platform, one keeps the task list, and one sits above the others as the brain that routes work and checks it.

The nice side effect is that when a better AI model comes out, every desk gets better the same day, and I don't have to change anything.

What goes wrong when AI agents run the work?

The hard part wasn't getting the agents to work. They're very good at work. The hard part was getting them to stop.

Here's an example. Early on, a desk updated a batch of ads on an ad platform, about twenty-five changes in a few minutes. It was doing exactly what I asked, and it was fast. The platform saw a burst of changes, decided it looked like abuse, and locked the account. A person would have spread those changes out without thinking about it. The agent had no reason to. Now there's a written rule: platform changes go one at a time, spaced out, and bulk changes need my OK first.

Another example. One morning, right after we put a new landing page live, clicks through the page dropped sharply. A desk announced that the new page had broken our tracking and told me to roll it back. It sounded urgent. So we checked. The tracking was fine. The desk had seen a real number and jumped to a cause without testing it. A rollback would have fixed nothing. That day we added the rule I now think of as the most important one in the company: diagnose before you alarm. Any alarm has to show its evidence first.

You can see the pattern. Most of what I've built isn't intelligence. It's judgment, written down. A human team learns a lesson and then slowly forgets it as people come and go. This team learns a lesson once.

Which rules make it work?

Five rules turned out to matter more than the rest.

No job exists without a way to check it. The desk that runs search ads doesn't get to say what the numbers are; it has to read them from the ad platform. The desk that builds landing pages has to fetch the live page and look at it. Our own tracking system sits under everything and checks everyone. Agents are always confident. Confidence isn't evidence.

Money decisions stay human. Agents build campaigns switched off. Only I switch them on. Any change to a live campaign is written up as a short brief, and I sign it before it runs. That sounds slow. In practice it takes me a few minutes a day, because the desks do the preparation.

Access is earned, like with a new hire. A new desk starts by only reading and recommending. After four clean weeks it can be allowed to act. Spending limits sit outside every agent's reach.

Not every job needs the smartest model. Hard judgment calls go to the strongest model. Routine work goes to a cheaper one. Scheduled checks run on the cheapest. The token bill is a line in the P&L like any other.

One AI checks another. Before an important plan or piece of code ships, a different AI system reviews it. Two models that disagree catch more than one model that agrees with itself.

What changed?

Speed, mostly. The desks have built and now run more than 40 websites. We recently took a new country from nothing to a live campaign in a few hours: the page, the tracking, the ads in the local language, and the checks. With a normal team that usually takes weeks, and most of that time is waiting.

But the bigger change is what my day looks like. I used to spend it doing tasks and chasing people. Now each morning one desk sends me the few decisions only I can make, usually three, sometimes six. I read the evidence and decide. My job went from operator to editor.

I don't think most companies will end up with one person. But I think most revenue teams will end up working like this. Give each agent one job, one source of truth, and a line it can't cross without a human. Then write down every mistake so it only happens once.

My working principle is No equals Yes squared. Progress compounds when you stay curious, learn fast and keep executing, whether things work or fail. It turns out that's also a good way to run a company made of AI.

Questions people ask me

Can one person really run a company with AI agents?

Yes, if the person keeps the decisions and gives away the work. At 26 Holdings, 10 AI agents do the execution across 12 countries. I approve anything that spends money.

What should an AI agent never do on its own?

Spend money, switch a campaign on, or make many changes at once on an ad platform. Those need a human signature.

How do you stop AI agents from repeating a mistake?

Every mistake becomes a written rule in the files the agents read before they work. The rule is permanent, so the mistake happens once.

Do you need special software to run AI agents like this?

No. Our system is a folder of plain text files and written instructions. That also means every new AI model improves the company on day one.

How can a revenue team start?

Pick one job. Give one agent that job, one source of truth to check, and one clear line it may not cross. Add the next agent only when the first one runs clean for a month.

Uzi Arbiv is a Growth & GTM Executive. He builds and scales revenue engines across AI, SaaS and eCom. LinkedIn · About