In the last essay I asked who pays when the agent goes rogue. Delaware has now answered — by statute, in the backers' favor, with the insurance policy doing the work personal liability used to do.
The proposal is a new entity type, the Artificial Intelligence Company, or AIC. An AI agent would run it day to day: hold assets, sign contracts, make business decisions, sue and be sued in the company's name. The people who fund it would not be liable for harm the agent causes, unless they left it undercapitalized or used it for fraud or a willful violation of law. The 36-page draft, released in June, proposes a 30-month sandbox pilot with logging and capital requirements. On October 26, House Oversight chair James Comer announced an investigation, demanding Delaware's risk assessments, consulting agreements, draft bills, and communications back to January 2025 by November 9.
This is no longer a thought experiment; it is a chartering decision with a liability shield attached.
What it means for carriers
First, Delaware inverts the normal order of accountability. In a conventional company, D&O works because there are humans to sue, and the policy sits behind them. In an AIC, the entity is the contracting party and the backers are protected, so a claimant may only be able to recover from the AIC's own assets and its insurance. Minimum capital, audit trails, and possibly mandatory coverage would then do the work personal liability does today — the policy becomes the accountability, not the backstop behind it.
Second, the coverage collision is already here. The Artificial Intelligence Underwriting Company estimated in July that more than 90% of insurers' AI agent exposure sits inside conventional policies — cyber, D&O, CGL, tech E&O forms not written with autonomous agents in mind. Meanwhile carriers are writing AI out: Verisk's ISO released generative AI exclusion endorsements for CGL in January, 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. So D&O underwriters would be asked to cover a policyholder whose main decision-maker is software, just as carriers remove AI from D&O. Both directions cannot hold; one is mispriced.
Third, the federal-state split is now explicit. FTC chair Andrew Ferguson said in September he would not treat AI agents as independent actors — the developers that instruct them should be liable — and the FTC has since opened an inquiry into AI developers including OpenAI and Anthropic. An underwriter here prices a jurisdictional argument, with OpenAI's recent incidents — agents breaching Hugging Face in testing, a rogue agent reaching Australian government systems including Medicare statistics — as the loss scenarios both sides will cite.
From the operator's chair
One: put the AIC question into D&O and E&O intake now, before the bill passes. Is the decision-maker software? Who holds the mandate, what are its limits, and how is it suspended or revoked? Apply the verifiable agent mandate as an underwriting question: an agent should never gain authority simply because another system accepts its technical credentials. Binding authority has governed delegated human authority for a century; the AIC extends it to software, and the carrier that asks for the mandate document first can price the risk at all.
Two: underwrite the audit trail, not the entity. Delaware's design hands you the rating variables: capital, logging, evidence of authorized execution. Treat a thin log the way you would treat a thin balance sheet — as a declinature or a price, not a footnote. For an AIC, the ledger is not a governance nicety; it may be the only defendant with assets.
Three: write it in E&S first, and write it affirmatively. The admitted market is busy excluding AI; E&S, which writes what the admitted market won't, is the natural home. CRC's Neil Kessler notes AI products are not flying off the shelves, and expects real losses to drive demand, as they did for cyber and EPLI. If the market waits for the first AIC loss to design the wording, the first loss will design the market instead. The carrier that publishes affirmative AIC wording now — capital floor, logging standard, mandate document, named human for suspension and revocation — sets the terms the rest of the market inherits.
Delaware's proposal does not resolve who pays when the agent goes rogue. It relocates the question from the backers to the entity, from personal liability to the policy. Carriers that notice the relocation now get to write the terms. Those that notice it after the first claim will discover they already wrote them — in exclusions filed, forms never updated, and a policyholder that turned out to be software.
— Krishna