AI can carry out a transaction. But if that transaction commits a company and causes harm, who is responsible? The project under consideration in Delaware puts this question at the heart of the debate. It envisages a form of company whose day-to-day management would be handled by an artificial intelligence agent, within a supervised experimental framework.
On 26 October 2026, James Comer, the Republican chair of the US House of Representatives Oversight Committee, announced an investigation into the initiative. His request concerns, in particular, the risks considered by the state and outside contributions to the preparation of the scheme. It is neither a judicial decision nor proof that a company has already committed a wrongdoing.
A proposal, not a general authorisation
The name Artificial Intelligence Company, or AIC, may give the impression that a new category of company is already available. Yet the material consulted describes a project that is still under discussion. In a report published on 30 October, Delaware Public Media quotes elected representative Krista Griffith as saying that the text had not yet been submitted to the governor or the legislature, with the state commission still having questions.
The framework described provides for a thirty-month trial for approved entities, with a regulator and an oversight committee. Such a “regulatory sandbox” is used to test a scheme within a defined scope. Its name alone does not guarantee safety or validate every possible activity.
The Delaware legislature’s Joint Resolution No. 7, whose text bears a drafting date of 25 June 2025, sheds light on the origins of the initiative: to build a framework for experimenting with agentic AI alongside the relevant commission and the Secretary of State’s office. It already distinguishes the prospects for efficiency from the legal and ethical questions associated with governance.
What James Comer is asking for
In its statement of 26 October, the federal committee sets out two concerns: the liability protections envisaged for the people behind these companies and the transparency of the project’s development.
Comer fears that the scheme could protect AI players and their investors from the consequences of harmful actions. He is also requesting information about risk assessment and the interests behind the proposal. These concerns must be attributed to their author: the statement sets out the committee chair’s position, not the findings of an independent assessment.
According to Delaware Public Media, the request for documents calls for a response by 9 November. The governor’s office had said it had received the letter and was reviewing it. At the time of the reported information, the process was therefore about gathering documents and explaining the project, not imposing a ban.
Responsibility does not disappear because software acts
The debate’s significance goes beyond the dramatic image of a “robot boss”. A company can delegate a task to a tool while retaining a human organisation that makes decisions, approves actions and answers for the consequences. The Delaware project takes the question further by asking which entity would carry out an agent’s operations.
The Delaware Public Media report describes a company that could sue and be sued in its own name. It also reports on protections being considered for its owners. These two points are not synonymous: making an entity legally identifiable does not, by itself, tell us which person would ultimately bear a loss.
Consider a purely hypothetical example: an agent orders a service that the company cannot pay for. The existence of a computer log would help explain how the order was placed. It would not be enough to determine who must pay the bill. That second question would depend on the contract, the applicable framework and the responsibilities actually assigned. The example illustrates a distinction, not a rule already adopted for AICs.
Announced safeguards that will need to be scrutinised
In the same report, Krista Griffith defends several provisions in the project: keeping activity logs reviewed by a lawyer, informing partners of the entity’s status and not dealing with the general public during the trial. These details describe safeguards advocated by one participant in the work. They do not yet demonstrate that those safeguards are effective.
Their practical value is nevertheless clear. Declaring the entity’s status would allow a partner to know what kind of organisation they were dealing with. Activity records could make a decision open to scrutiny. Limiting the scope would aim to contain the situations exposed during the trial.
An important distinction remains between keeping a record and having a control capable of acting in time. A report after an operation does not serve the same purpose as authorisation beforehand. This is one useful criterion for assessing any future text: what actions would be possible, who could stop them, and how could affected people assert their rights?
A test of governance as much as technology
The dossier does not present a result showing that an agent can manage a company over the long term. It concerns the framework in which that possibility would be tested. Technical capability, permission to act and the assignment of responsibility remain three different questions.
To assess what happens next, performance claims will therefore matter less than precise rules and documented results from any trial that may take place. A system can perform a task correctly without being reliable in every circumstance surrounding that task. Conversely, the possibility of errors does not, by itself, explain which framework would be most appropriate.
The next concrete development will be the response to the federal request for documents, followed by the progress of the project in Delaware. Until then, saying that companies have already been freed from all responsibility or that there is general authorisation for AI governance would go beyond the available evidence. The real question is more specific: how can operational autonomy be tested without making responsibility impossible to understand?
Additional source: report by Bente Bouthier for Delaware Public Media, 30 October 2026.
Illustrative photograph of a review of documents; it does not depict participants in the Delaware project.
