Delaware is proposing a company run by an AI agent. Not advised by one, not assisted by one — run by one. Under the draft Artificial Intelligence Company, or AIC, the agent would hold assets, enter contracts, make business decisions, and sue or be sued in the company's name. The person or company funding it would generally not be liable for what the agent does, unless they left the company undercapitalized or used it for fraud or a willful violation of law.
The draft, 36 pages released in June 2026, would test the idea inside a 30-month regulatory sandbox, with capital and logging requirements and banking activity excluded. It traces to a July 2025 resolution directing the state's AI Commission to design the sandbox, and the bill is expected before the legislature next session. On October 26, House Oversight Committee chair James Comer announced an investigation, with Delaware's response due November 9.
The insurance question is not whether the technology works. It is who pays when it doesn't — and Delaware's answer, for this class of company, is: the company itself, up to its capital and its insurance, and largely nobody behind it.
What it means for carriers
First, this is the accountability question this series has been circling, now wearing a corporate charter. Every carrier's governance case rests on a named human: the underwriter who confirms, the officer who owns the decision, the ledger that keeps the tape. An AIC removes the human from the operating chair by design, replacing personal liability with entity capital, a log, and disclosure that the counterparty is dealing with a test entity. That is a governance theory — that legibility substitutes for liability.
Test it before you underwrite it, because visibility is not recourse. A log tells you what the agent did. It does not pay the claim. Capital pays, until it doesn't — and then the shield does its work, leaving the insurer and the counterparty holding a loss whose upside went to backers who are no longer reachable. There is no officer to depose, no director whose judgment can be second-guessed.
Second, Washington is moving the other direction. The FTC's chair said in September that he would not treat AI agents as independent actors and suggested the developers instructing them should be liable; the FTC has since opened an industry-wide inquiry into AI developers. Delaware would make the agent the actor and shield the people behind it, while Congress investigates Delaware. Three authorities, three answers, no settled law. Underwrite to the version you can defend, not the version you prefer.
Third, the exposure is already in the book, with or without Delaware. The Artificial Intelligence Underwriting Company estimated in July that more than 90% of insurers' AI agent exposure may sit inside conventional policies — cyber, D&O, CGL, and tech E&O forms never written with autonomous agents in mind. Meanwhile a review of state filings counted 2,369 generative AI exclusion records in force across 49 states by mid-2026, and W.R. Berkley has added an absolute AI exclusion to its D&O, E&O, and fiduciary products. The market has already split, as Essay Nº 008 described: exclude the risk, or affirmatively underwrite it. An AIC forces that choice into a single decision, because the policyholder itself is now the agent — and no form written for human officers, human attackers, or human professional judgment describes this insured.
From the operator's chair
One: decide the AIC position now, line by line, before the bill arrives. Will you insure an entity whose decision-maker is software? If yes, on what wording — and if no, say so explicitly, because silence will be read as coverage by a claimant, a broker, and eventually a court.
Two: if you insure it, price the control stack, not the shield. Require proof of capitalization, disclosure to counterparties, and an activity log a reviewer can read — what the agent saw, what authority it claimed, what it committed, what stopped it. That log is the verifiable agent mandate from Essay Nº 006, applied to the counterparty. An AIC that cannot produce it is an unpriced risk with a charter.
Three: audit the silent version already in your book. Your insureds are signing contracts with agent-run counterparties and delegating decisions to agents. Find where an agent's commitment could bind your insured, or your insured's agent could bind someone else, and check whether the wording answers that claim or merely assumes a human made it. The bill is a proposal. The counterparty risk is already live.
Four: do not wait for the federal fight to resolve. Comer's November 9 deadline, the FTC inquiry, and the next legislative session will not converge into one clean rule. Write interim terms that survive any outcome: capital, log, disclosure, and a suspension trigger you control.
Delaware proposed giving the agent-run company a charter, a shield, and a sandbox. Before you insure a company with no one to sue, decide what you are insuring — the capital, the log, or the hope that nothing goes wrong. Only the first two can be priced.
— Krishna