Delaware is considering a company built for an unusual manager: an AI agent.
On July 13, Delaware Secretary of State Charuni Patibanda-Sanchez and Norm Ai CEO John Nay described a proposed Artificial Intelligence Company, or AIC, in a Fortune commentary. Their version would be a separate legal entity with day-to-day affairs managed by an AI agent. At the agent’s direction, the company could hold property, incur obligations, and sue or be sued.
That entity does not exist in Delaware law today. The AIC is a proposal being developed for a regulatory sandbox, and the legislation may not reach the General Assembly until next year. Norm Ai, which sells legal and compliance AI, is an interested commercial participant in the effort.
Those caveats matter because the proposal moves a live Hypernovelty question closer to the institutional world. AI systems already draft, recommend, rank, route, and sometimes execute inside organizations. The organization remains the legal actor while the machine’s influence can disappear into the workflow. The proposed AIC would bring the wrapper and the operational system closer together.
Call it an actor wrapper: a legal container through which a nonhuman system can exercise authority in the world.
A sandbox with legal weight
The official foundation is Delaware House Joint Resolution 7. It directed the state’s AI Commission and Secretary of State to develop a regulatory-sandbox framework for agentic AI. The resolution named corporate governance as an area where automation may create benefits alongside legal and ethical questions.
The AIC described by Patibanda-Sanchez and Nay would operate only inside that sandbox. A committee including state legal and judicial officials and outside experts would decide admission. Each company would have one human or entity member responsible for keeping it adequately capitalized. The AIC would maintain an activity log and tell counterparties that it is a test entity, that Delaware does not endorse it, when the test ends, and how to complain. Officials could suspend authorization and seek dissolution. Banking would be excluded, and the proposed program would end after 30 months unless lawmakers acted.
Those are meaningful boundaries on paper. The harder test arrives when the agent does something consequential.
A legal name gives a court something to see. Capital gives an injured party something that may be reachable. A disclosure tells a counterparty they are inside an experiment. But accountability still depends on the record between instruction and action.
“Activity log” can mean a complete authority trail or a thin event list. A useful log would let a reviewer identify the entity and acting agent, the model and configuration in use, the permission granted, any delegation to another agent or tool, the evidence consulted, the contract or action produced, the human authority points, and the available pause, complaint, rollback, and repair paths. These are operating questions for the proposed sandbox, not requirements found in enacted Delaware law.
Identity gets difficult quickly. The 2026 academic preprint How to Count AIs notes that agents can copy, split, merge, disappear, and delegate across different models. Its authors propose secure credentials and action-level audit trails for an algorithmic corporation. That remains a scholarly proposal. Still, it identifies the practical problem: law cannot assign responsibility cleanly if the action trail cannot show which authorized system did what.
The liability shield will also meet borders. Corporate-law professor Ann Lipton has asked whether courts in other states would respect Delaware’s proposed protection when one of their residents brings a tort claim. Her analysis is a warning, not a settled answer. Delaware can create an entity form. It cannot guarantee how every court will treat every harm.
The state is drawing other boundaries at the same time. A law signed in April bars nonhuman entities, including AI agents, from holding medical licenses or using protected medical titles in Delaware. One lane explores AI-managed companies. Another reserves licensed care roles for humans. The categories are being built domain by domain.
Verification bottleneck
Verification is becoming the scarce institutional function.
- Agent authority is moving faster than identity standards, audit specifications, complaint systems, insurance models, and cross-state liability rules.
- Sandbox officials, counterparties, courts, owners, insurers, auditors, and affected people would have to verify who authorized an action, what the agent could do, which evidence it used, and what remedy remains when the outcome causes harm.
- Watch next: the actual bill text, admission criteria, minimum capital, log fields, independent audit access, complaint handling, pause authority, and treatment outside Delaware.
Opportunities
Where value may appear: the proof and repair layer around an AI-managed entity.
A builder could create a counterparty disclosure packet, agent-authority registry, delegated-permission map, action-level audit log, sandbox-monitoring service, complaint-and-repair workflow, or cross-jurisdiction readiness checklist. The useful product would show authority and consequences without pretending a dashboard settles the law.
This is builder idea fodder and public-interest orientation, not legal, financial, insurance, compliance, or investment advice.
The wrapper may soon have a driver. Before it gets the keys, someone needs to define the log, the brakes, and the repair path.
Public sources
- John Nay and Charuni Patibanda-Sanchez, “Delaware Secretary of State partners with Norm Ai to propose the AIC, a legal entity for agents,” Fortune, July 13, 2026: open source
- Delaware General Assembly, House Joint Resolution 7: open source
- Spotlight Delaware, “Delaware committee drafts plan to test companies run by AI,” June 30, 2026: open source
- Ann Lipton, “AI Entities and the Internal Affairs Doctrine,” July 2026: open source
- Arbel, Goldstein, and Salib, How to Count AIs: Individuation and Liability for AI Agents, 2026 preprint: open source
- Delaware General Assembly, House Bill 191, signed April 23, 2026: open source