What is an AI-run company? A human-led definition
A plain definition of a company run with AI agents, what experiments show so far, and what the person in charge still does.
An AI-run company is a business where AI agents carry out much of the day-to-day work while a person sets the direction, makes the consequential calls and stays accountable. "AI-run" describes who does the work, not who is in charge.
What is an AI-run company?
An AI-run company is a business in which AI agents handle a large share of the operating work, such as drafting, research, building and review, while at least one person leads. That person decides where the company is going, approves anything with real consequences and answers for the results. Agents do the work; a human owns it.
The phrase gets used loosely, sometimes for a company with no humans at all. This guide uses the human-led meaning, because it matches how companies built with AI agents operate today, and because someone always remains responsible.
Is an AI-run company the same as a "zero-human" company?
No. A "zero-human" company is the far end of a spectrum, where agents set goals and act without anyone approving their decisions. In our view, most businesses that use AI agents sit closer to the other end, and it helps to think in three levels.
| Level | Who sets the goals | Who approves consequential actions | Who is accountable |
|---|---|---|---|
| 1. Agents draft, a person decides | A person | A person approves almost everything | A person |
| 2. Agents ship routine work, a person approves anything external | A person | Agents handle routine internal work; a person approves money, outside contact and anything irreversible | A person |
| 3. Mostly autonomous experiment | Agents, within a broad goal a person gave them | Agents, with little or no review | Still a person or an organization, in law and in practice |
In our view, most businesses using AI agents today operate at levels 1 and 2, usually on purpose. Level 3 exists mainly as research, and even there the agents don't own the company or answer for it.
What do real experiments show so far?
Published experiments show that AI agents can carry real business tasks, and that they still make costly mistakes when left without oversight. Two primary sources are worth reading in full.
An AI agent running a small shop. In 2025, Anthropic let an AI agent manage a small automated store in its San Francisco office "for about a month." Anthropic's own verdict: "we would not hire" it, because "it made too many mistakes to run the shop successfully." In a second phase, published in December 2025, Anthropic moved to newer models, added tools and gave the shop agent a manager agent acting as its "CEO." Anthropic wrote that these changes "did make Claudius's shop more successful," and that the agent was "still vulnerable in lots of important ways." The manager agent "shared many of the deficiencies" of the agent it managed.
Agents in a simulated workplace. Researchers led by Frank F. Xu built TheAgentCompany, a benchmark that simulates a small software company and tests agents on its workers' tasks. In the version of the paper we read (v3, last revised September 10, 2025), the most competitive agent could "complete 30% of tasks autonomously," and harder long-horizon tasks were "still beyond the reach of current systems." That figure has changed across versions of the paper, so check the latest one before quoting it.
Neither result says agents are useless. Both suggest that running a business still depends on judgment, priorities and someone catching errors before they spread.
What does the person in charge still do?
The person in charge sets direction, chooses priorities, approves consequential actions and stays accountable for the outcome. With AI agents doing the work, that job gets more concentrated, not smaller.
- Direction. What the company is trying to achieve in the next few months, written where every agent can read it.
- Priorities. Which few things matter this week, and what waits.
- Approvals. The decisions that should wait for a person: spending money, contacting anyone outside the company, anything that can't be undone, and any claim about the product, customers or numbers.
- Anything external. Customers, partners and the public deal with a person, or with work a person approved.
- Accountability. When something goes wrong, the person in charge answers for it and fixes the setup.
Agents can usually report what they're working on whenever you ask. The leader's job is to decide what that work should add up to.
What work do AI agents usually carry?
AI agents usually carry drafting, research, building and reviewing: repeatable work that can be checked against a clear definition of done.
| Role | Typical work | What it hands back |
|---|---|---|
| Coordinator | Turns priorities into tasks and routes them | A plan and a list of decisions that need a person |
| Writer | Pages, emails, posts, documents | A draft in a file |
| Researcher | Facts, sources, summaries | A summary with links to sources |
| Builder | Code, site changes, spreadsheets, internal tools | A change someone can open and test |
| Reviewer | Checks another agent's work against the brief | A verdict with reasons |
For a step-by-step setup with roles, gates and a weekly review, see our guide on how to run a one-person company with AI agents.
What are the risks, and how do you limit them?
The main risks are errors that compound across agents, external actions nobody reviewed, and costs that drift upward. Each has a simple limit.
- Compounding errors. One agent's output often becomes another's input, so a small early mistake can grow. Have a different agent, or you, review work before it moves on, and keep work in files rather than chat threads.
- Unreviewed external actions. An email, a post or a payment can't be taken back, so make those actions wait for a person by default. Anthropic's guidance on building agents describes agents that "can then pause for human feedback at checkpoints or when encountering blockers."
- Cost drift. More agents running more tasks means more spend. Set a budget, check usage weekly and stop work that no longer serves a priority.
Our checklist on which AI agent actions need human approval goes deeper on where to put those checkpoints.
How do you start one this week?
Start small, with five steps you can finish in a week.
- Write a one-page direction. The goal, this week's three priorities and what agents may never do without you.
- Pick two or three roles. For example, one agent that produces, one that reviews, and you as the decider.
- Set your approval rules. Money, outside contact, anything irreversible and public claims wait for you.
- Keep work in files. Each task gets a short brief, one owner and a saved output.
- Review weekly. What moved, what's stuck, what to stop.
For a day-to-day version, use our AI agent tracking checklist.
Where does Plenarly fit?
Plenarly is being built as the operating system for AI-run companies: one place where founders and small teams connect their AI agents from any lab, approve their work, and manage the apps those agents build. Early-access sign-ups are open.
The plan is one view of the whole company, organized by projects and tasks, with AI agents from different labs side by side. Join early access if you already run AI agents on real work. For plain facts, see What Plenarly is and isn't.
Frequently asked questions
Can a company be run entirely by AI?
Experiments have tried it, and published results so far show agents still making costly mistakes without oversight. Anthropic's shop experiment ended with the verdict that it "would not hire" its agent. A company also has owners and officers who are legally and practically accountable, so a person stays responsible.
Is an AI-run company legal?
Using AI agents to do a company's work is common, but under ordinary company law a business is owned and directed by people or other legal entities, not by agents. Rules differ by country and state, so treat this as general information, not legal advice, and check with a qualified professional.
How many people does an AI-run company need?
At least one person who leads, sets direction, approves consequential actions and stays accountable. The practical limit isn't how many agents you can start but how much you can review and decide well. Attention, not agent count, sets the pace.
What's the difference between an AI-run company and a one-person company?
They overlap but describe different things. "One-person company" is about headcount. "AI-run company" is about who does the work. Many one-person companies are AI-run, and some AI-run companies have several people. Our one-person company guide covers the setup.
How do you keep an AI-run company accountable?
Give each agent a clear role and each task one owner. Write down what "done" means before work starts, keep money, outside contact and irreversible actions behind your approval, and save every output to a dated file. Then review the whole company weekly and fix one thing that went wrong.
Sources
- Anthropic, "Project Vend: Can Claude run a small shop? (And why does that matter?)," June 27, 2025. Quotes: "for about a month"; "we would not hire Claudius"; "it made too many mistakes to run the shop successfully." Checked on the live page Oct 24, 2026, ~8:30 PM ET. https://www.anthropic.com/research/project-vend-1
- Anthropic, "Project Vend: Phase two," December 18, 2025. Quotes: "did make Claudius's shop more successful"; "still vulnerable in lots of important ways"; "shared many of the deficiencies." Checked on the live page Oct 24, 2026, ~8:30 PM ET. https://www.anthropic.com/research/project-vend-2
- Xu, F. F. et al., "TheAgentCompany: Benchmarking LLM Agents on Consequential Real World Tasks," arXiv:2412.14161, v3, last revised September 10, 2025. Quotes from the abstract: "complete 30% of tasks autonomously"; "still beyond the reach of current systems." Checked Oct 24, 2026, ~8:30 PM ET. https://arxiv.org/abs/2412.14161
- Anthropic, "Building effective agents," December 19, 2024. Quote: "can then pause for human feedback at checkpoints or when encountering blockers." Checked Oct 24, 2026, ~8:30 PM ET. https://www.anthropic.com/research/building-effective-agents