Delaware wants companies run by AI agents. Who will answer for their decisions?

Delaware is preparing rules that would allow an AI agent to run a company on its own and would largely relieve its owner of liability. The House Oversight Committee is already demanding documents, and companies deploying agents are facing the same question on a smaller scale.

Delaware, the state where roughly two-thirds of Fortune 500 companies (the 500 largest U.S. enterprises) are incorporated, is working on a new type of company. Its day-to-day management would be handled by an AI agent, meaning software that independently plans and carries out tasks, rather than a human. The idea may look like a curiosity of American corporate law. In practice, it raises a question that managers deploying agents are already asking today: who is responsible when the machine makes a mistake?

A company managed by software

The draft bill provides for the creation of an Artificial Intelligence Company (AIC). Such a company could own assets, incur liabilities, sue and be sued in its own name. It could have any lawful purpose, except banking activity. The bill's lead author, corporate lawyer John Mark Zeberkiewicz of Richards, Layton and Finger, told Spotlight Delaware that an AIC could provide programming services, sign contracts, and even conduct litigation. In his view, “anything a company can do, an AIC can do as a company”.

Behind the company stands one “member”: a person or company that funds it, takes the profits, and is responsible for its adequate capitalization. An AIC would be required to keep a record of its actions. The whole model would first operate in a regulatory sandbox, meaning a test mode under state supervision, for 30 months. The draft will likely go to the state legislature next year.

The idea did not come out of nowhere. In July 2025, Governor Matt Meyer signed a resolution on a sandbox for agentic AI. A subcommittee of the state AI commission is working on the sandbox project. Meanwhile, the framework for cooperation with business is being developed in a public-private partnership led by Norm Ai, which describes itself as building AI agents for legal tasks. Secretary of State Charuni Patibanda-Sanchez also did not hide the fiscal motive: in her view, the new company form would be “a huge benefit to state revenues”.

A shield for the owner: order or loophole

Supporters present the proposal as a way to bring order to something that is happening anyway. In a July article in “Fortune,” Patibanda-Sanchez and Norm Ai chief John Nay argued that banning autonomous agents will not work, so it is better to give them a legal form. In their view, a company with legal identity, a record of actions, and assigned capital becomes a target against which liability and damages can be pursued.

The devil is in the clause about the company member. The draft states that the member is not liable for the actions or omissions of the AIC. Exceptions include fraud, willful violation of the law, undercapitalization of the company, and situations in which the member knew or should have known that the company was operating outside its declared purpose. Andrew Baker, a law professor at UC Berkeley, told the “Financial Times” that this is an overly generous liability shield that nobody has properly justified. Martin Petrin of York University in Toronto told Bloomberg that other states will propose similar solutions and warned of a regulatory race to the bottom.

House committee demands documents

On October 26, James Comer, chair of the House Oversight Committee, sent Governor Meyer a letter demanding documents about how the proposal was developed and how the risks related to AI safety were assessed. The committee fears that AICs will be used to shield influential AI companies and their investors from liability for reckless and harmful actions. Comer also suggested that the findings of the investigation could be used for a federal bill limiting states’ regulatory initiatives in AI. The governor’s office confirmed that it is reviewing the letter, while the secretary of state’s office declined to comment and referred questions to the governor. The matter is therefore ongoing.

For comparison: in January 2026, Singapore announced a governance framework for agentic AI, described as the first of its kind in the world. It emphasizes real human accountability, and applying it is voluntary. At the same time, it stresses that organizations remain legally responsible for the actions of their agents. That is the opposite direction from Delaware.

This is a question companies already have on their desks

Most companies will not hand an entire company over to an agent, but they are increasingly handing over specific tasks. Gartner forecasts that more than 40 percent of agentic AI projects will be canceled by the end of 2027. Among the reasons, analysts cite rising costs, unclear business value, and insufficient risk control. McKinsey’s State of AI 2026 survey shows that only 6 percent of companies see a significant impact of AI on EBIT (operating profit). This group was nearly three times more likely than the rest to redesign its work processes.

What a gap in oversight looks like can be seen in e-commerce. Shopping agents answer customers based on the store’s knowledge base, which Shopify has just moved into the seller dashboard. Bill Hunt reviewed stores on this platform in Search Engine Journal. Of the 88 where he tested the tool for questions about store policies, only two gave a useful answer to a question about order cancellation. The platform does not have a separate category for order cancellation policies, so the seller has to add the missing information manually — and the store is responsible for what the agent tells the customer.

Rick Folgmann, COO of insurance company Get Covered, described in an op-ed for Crunchbase News how his company manages 16 AI agents. Each has a named person assigned to review its errors, decide the scope of its autonomy, and take responsibility for it. Folgmann notes that automated credit decisions have always had to be explained to the customer or the regulator, and the answer “the model did it” was never accepted. This is one company’s account, not a researched market standard, but it illustrates the problem well.

What this means for a team deploying agents

The dispute around Delaware shows that even the most liberal proposal cannot do without a human: it requires capital, a record of actions, and someone responsible for the company’s purpose. In a company that entrusts an agent with real work, we believe it is worth starting with a few principles:

• Assign the agent tasks with rules set out in writing, for example in a contract or regulation.
• Assign each agent a named owner.
• Set autonomy separately for each task, not once for the entire tool.
• Keep a record of the agent’s decisions and review it regularly.
• Test the agent with the questions a real customer would ask.

When setting autonomy, a question proposed by Folgmann helps: if this activity goes wrong, how quickly will we realize it and how much will it cost? Fast detection and low cost make it possible to give the agent more freedom. The dispute over Delaware will be decided in the state legislature and in Congress. Responsibility for an agent in a company already lies with the company today, and for contracts and legal risk, it is worth seeking a lawyer’s opinion.