The state of Delaware is preparing a bill for a new type of legal entity — the Artificial Intelligence Company (AIC) — within which operational decisions could be made by an AI agent: conducting deals, signing contracts, filing lawsuits, and defending in court without a human “at the helm.” The owner of such a company is exempt from liability for the AI’s actions, subject to certain exceptions, and the mechanism is proposed to be tested in a 30-month regulatory “sandbox” before a permanent law is adopted. The initiative has already faced resistance in Washington: House Oversight Committee Chair James Comer has requested documents from the Delaware governor and threatened a federal law against such state AI experiments. The outcome of this conflict will determine whether autonomous agents gain a legal way to own and run businesses.

image

What happened

The bill introduces a new organizational form into Delaware law — the Artificial Intelligence Company (AIC). Its key distinction from a standard corporation is that an AI agent is authorized to make operational decisions: it can conduct deals, sign contracts, and conduct legal proceedings as both plaintiff and defendant. The human owner (member) does not bear liability for the AI’s actions, except in cases of insufficient capital, fraud, intentional violation of the law, and actions beyond the company’s stated goals. One hard boundary has already been defined: AICs will not be allowed to engage in banking. The mechanism is proposed to be tested in a 30-month regulatory “sandbox,” and only after this review is a permanent law planned to be adopted. Simultaneously, the initiative has hit federal resistance: Congressman James Comer (R-Ky), who heads the House Oversight Committee, demanded documents from Governor Matt Meyer regarding the project and threatened a law limiting state AI regulatory experiments.

Context

The location was chosen deliberately. Delaware is a key corporate jurisdiction in the US: approximately two-thirds of Fortune 500 companies are registered there, and the state opened 334,461 new companies in 2025 alone. Corporate law is the state’s historical specialization, so the attempt to be the first to formalize the legal status for agentic AI appears to be a continuation of this model: Delaware is cementing its role as the legal jurisdiction for agentic autonomy. At the same time, the initiative is embedded in a long-standing conflict over levels of regulation, where states experiment with AI rules, and Congress periodically threatens federal norms that override state ones. The proposed structure, in the spirit of “limited owner liability in exchange for the agent adhering to stated goals,” effectively turns the mass deployment of autonomous agents into a field experiment in the real economy — however, the methodology, metrics, and stopping criteria for such an experiment are not described in available materials.

Why this matters for the industry

Legal personhood for AI changes the very architecture of liability: today, a human or the developer company is liable for losses from autonomous system decisions, whereas under the Delaware scheme, the focus shifts to the legal entity structure and the agent’s formal mandate — goals, powers, prohibitions, and limits. For builders of agentic platforms, this means that observability, evals, and immutable audit logs of agent decisions transform from good engineering practice into a legally significant framework: someone will have to technically define the boundaries of the system’s autonomy and ensure the handover of control to a human. Product-wise, this opens the category of “company as a product,” where the main interface becomes the executable mandate, and demand arises for agent identity infrastructure — registries of powers, signatures, and checks in the spirit of know-your-agent. The flip side is the risk of a “race to the bottom” between states competing for AI company registrations, and a likely response from Congress, which could either legitimize or block the model. For agent startups, the legal wrapper would no longer be a platform-level blocker — but only if the initiative survives federal pressure.

Why this matters for users

The law will not bring immediate changes for the reader: it has not been passed, the “sandbox” has not been launched, and agents do not currently have legal personhood. If the model does reach launch, the structure of services around the user will change. Autonomous services and agent-based startups could be built so that the business is formally run and represented by an AI agent with its own legal entity, rather than a human manager: contracts signed by the agent will have legal force, and risks will be concentrated in a separate company with limited owner liability. For consumers, this means that in case of losses, claims will be addressed to the AIC within the agent’s stated goals and powers, and the mandate itself will become a public description of what the agent is allowed and prohibited from doing. A reasonable action now is observational: monitor the status of the bill and Comer’s request, and do not make irreversible structural decisions in anticipation of the AIC, as federal intervention could shut down this model before any launch.

What is currently unknown / limitations

The key claim of the initiative — that an agent is capable of fully running a business without a human “at the helm” — is not supported in available materials by results on known agentic benchmarks, a description of the system architecture, or criteria for what level of autonomy is considered sufficient. LLM agents are stochastic and sensitive to prompt injection, and “provable compliance with the mandate” remains an open research problem: no methodologies, metrics, or reliability thresholds are present in the sources. Requirements for the transparency of the “sandbox” and access to its data are also not described — without them, the experiment could turn into a legal black box instead of a measurable source of data on agent failures in contractual and legal scenarios. Finally, status uncertainties remain at every step: this is currently a bill and a document request, the text and fate of the law are unknown, and Congress may adopt a federal restriction before any mechanism review begins.

Sources

Author

Look at AI, editorial team