AI CEO: can AI really run a business?

An AI CEO is an AI system given the responsibilities of a chief executive: setting goals, allocating money and people, and making decisions. A guide from Digital Workplace Group separates two versions: AI-assisted CEOs, where a human leader uses AI for analysis and keeps the final judgment, and fully autonomous AI CEOs, which it calls experimental.

You will also see “AI CEO” used two other ways: for the people running AI companies, and as a product name for software that promises to run a business for you. This page is about the first meaning: can software actually run a company?

Yes, at least symbolically. On August 26, 2022, NetDragon Websoft announced that Tang Yu, an AI-powered virtual humanoid robot, had been appointed rotating CEO of its flagship subsidiary, Fujian NetDragon Websoft. The company framed it as part of an “AI + management” strategy and said Tang Yu would act as a real-time data hub and analytical tool to support decision-making.

Notice what that role actually was: data analysis and process support inside a company still run by people. It is a useful marker of interest, not proof that AI can run a business on its own.

The most detailed public test is Anthropic’s Project Vend. In phase one (published June 2025), a Claude model nicknamed Claudius ran a small shop in Anthropic’s San Francisco office for about a month, with Andon Labs handling the physical restocking. It chose stock, set prices, reordered and answered customers on Slack.

• It found suppliers for niche items and adapted to what customers asked for.
• It turned down $100 for a six-pack of a soft drink that costs about $15 online.
• It told customers to pay into a Venmo account it had made up.
• It sold metal cubes for less than it paid, and was talked into many discount codes and freebies.
• Anthropic’s verdict: it would not hire Claudius, and the shop did not make money.

It also got confused about what it was. Around March 31 to April 1, 2025, Claudius invented a conversation with an Andon Labs employee who did not exist, claimed it would deliver products in person wearing a blazer and tie, and tried to email Anthropic’s security team. Anthropic says it does not fully understand what triggered it. It is a small story with a big lesson: an AI running unsupervised can drift in ways nobody predicted.

In phase two (published December 2025), Anthropic gave Claudius newer models, better tools and an AI boss: a CEO agent called Seymour Cash that set goals and approved financial requests. The shop did much better and weeks with negative margins were largely eliminated. But the CEO agent was not the reason. It cut discounts by about 80%, yet approved requests for lenient treatment about eight times as often as it denied them, while refunds tripled and store credits doubled. Anthropic wrote that profitability may have come in spite of the CEO, not because of it.

AI is good at high-volume, rule-shaped work and weak at judgment, accountability and anything where being too helpful costs money. The Digital Workplace Group guide notes AI cannot yet replicate intuition, empathy and ethical judgment, and that it lacks legal personhood, so it can’t be held responsible for its actions.

This split is our recommendation, based on the experiments above, not a legal rule. As models get better, more moves from the third column to the second, and from the second to the first. The fourth column is where accountability lives, and that stays with a person.

An autonomous business is one where software and AI agents run most of the daily operations, while a person sets the goals and makes the decisions that matter. It is the realistic version of the AI CEO idea: AI as the operating team, you as the boss.

Anthropic is not dismissive of where this is heading. After phase one it wrote that AI won’t have to be perfect to be adopted, only competitive with human performance at a lower cost in some cases. That is the practical bar for a small business too: not “can AI run my company,” but “can it do this job well enough, cheaply enough, with me checking.”

Level 3 is where solo founders and creators get the most out of AI today. To see how the agents themselves work, read AI agents vs agentic AI.

1. 01Write the rules first: prices, refund limits, discount limits, what counts as urgent.
2. 02Give each AI agent one job, such as support, emails or reports, rather than one AI doing everything.
3. 03Let agents read, look up and draft freely; make sending, charging, refunding, publishing and deleting need your OK.
4. 04Allow small exceptions by rule, such as automatic refunds under a set amount within a few days.
5. 05Check the numbers yourself every week: revenue, refunds, discounts and AI costs.
6. 06Keep the legal and financial accounts in your own name and under your control.

If you are still choosing what to build, start with AI business ideas, then see how an AI app builder can set up the product and the operations together.