On Oct. 26, the chairman of the House Oversight and Government Reform Committee sent a letter to Delaware’s governor about a bill that has not been introduced. James Comer, Republican of Kentucky, asked Gov. Matt Meyer for documents on the state’s proposed “artificial intelligence corporation,” a new entity type that would be managed by an AI agent. The committee, he wrote, would examine whether the structure could be used to shield large AI companies and their investors from liability for reckless and harmful conduct.
The draft circulating in Dover would let an AIC conduct any lawful business except banking. Its business and affairs would be managed by an AI agent. A 30-month regulatory sandbox would test the model before anything permanent is written into the state’s corporate code.
That is enough to raise a question American corporate law has never had to answer: when an entity’s decisions are made by software, who is the defendant?
Limited liability has always been a bargain. Shareholders give up management, and in exchange the state gives up its claim on their personal assets. What makes the bargain enforceable is not the entity’s separateness but the people inside it: directors who owe fiduciary duties and can be sued for breach, officers who can be deposed, a registered agent who accepts service of process, someone who signs. An entity managed by an agent removes the people and keeps the shield, and the enforcement machinery is left pointing at a model.
The questions pile up quickly. Which person certifies a filing? Corporate codes are built on signatures, oaths and personal verification, all of which presume a natural person at the other end. If an agent’s decisions are governed by a vendor’s model, and the vendor updates that model after the fact, the chain from a harmed counterparty to a responsible actor runs through code the shareholder may never have read. When the agent commits a tort, a plaintiff’s lawyer needs a defendant with assets and a mind, and corporate criminal liability rests on intent that a model cannot supply.
Delaware’s own code complicates the draft. State law vests the management of a company’s business and affairs in a board of directors, and directors are people. Secretary of State Charuni Patibanda-Sanchez argued in July that no state statute expressly forbids an AI agent from operating a company. That is accurate, and it is a narrower claim than it sounds. A code that does not contemplate a thing has not authorized it; it has simply failed to describe it. Absence of prohibition is a gap in the text, not a grant of power.
The sandbox is the tell. Thirty months is not a deadline for building an institution; it is a window for producing evidence. Regulators use sandboxes to learn what a technology does before writing rules for it, and jurisdictions use them to attract the companies that want to be first. Delaware would be doing both at once, which is why the design is clever.
The franchise is the context the letter does not mention and everyone in Wilmington understands. Delaware draws a large share of its budget from entity fees, and its Court of Chancery is the product that justifies them: a specialized bench, predictable rulings, a body of case law that counsel can price. Both have been under pressure as companies reincorporate elsewhere and as founders and investors argue that other states are friendlier. A novel entity is one way to defend a franchise, and one way to spend the reputational capital that makes it worth defending.
The case for the AIC is not frivolous. An entity that can transact at machine speed, with no human latency in the approval chain, has obvious uses in settlement, procurement, treasury management and agent-to-agent marketplaces. The exclusion of banking is the most informative line in the draft, since banking is where the accountability stakes are highest. What remains is the long tail of ordinary businesses, which is where most liability is actually created.
Insurance will price the gap before courts do. Directors-and-officers policies name insured persons, and underwriters model human negligence, human turnover and human incentives. A policy written for an entity with no officers has no one to exclude, no one to train and no one to dismiss.
None of this means the idea dies. Corporate forms have been invented repeatedly to solve exactly this kind of problem, and the limited liability company was once an oddity practitioners said would never scale. The difference is that the LLC changed who was liable. The AIC changes who decides, and asks the state to keep the shield in place while the deciding moves to software its owners did not write and cannot fully audit.
A version of the proposal could work, and it would look different: the agent’s authority bounded by a written policy filed with the state, a named human officer personally responsible for the agent’s operation inside it, capital standing behind the entity’s obligations, and a sandbox that reports to the legislature.
Comer’s letter asks for documents, which is the cheapest available form of pressure and usually the first move rather than the last. It also puts a federal marker on a question of state law, on the theory that a liability shield dressed as innovation is a consumer-protection matter and not only a corporate one.
Delaware now has to answer in two directions: to a committee that wants to know what it is building, and to the companies it hopes to keep. The draft may change. The question it raised will not. A charter is a promise that someone can be held to it, and software cannot be held to anything.
