A bill summary with no bill number. No named sponsor. No legislative text. No timestamp. That is the entire documentary record behind the claim that Delaware is preparing to let an algorithm operate a company — and that Congress has begun asking questions about who pays when it fails.
I have read audits with more evidentiary weight in their commit logs. The summary circulating this week reduces to a single sentence, restated five ways: Delaware is weighing a corporate form operated by artificial intelligence; the arrangement raises questions about owner liability; federal lawmakers have flagged concerns. The entity type is unspecified. Whether the machine sits as director, officer, manager, member, or mere instrumentality is unspecified. Whether a natural person remains anywhere in the chain of accountability is unspecified.
Silence in the data is a confession.
The absence of text is not a gap in the reporting. It is the most important fact in the story. What exists is a policy signal, not a policy. And a signal with no clauses cannot be audited — only bracketed, weighted, and watched.
Context
Delaware's corporate franchise is a business. Entity fees and franchise tax contribute roughly a billion dollars a year to a state budget with few other growth engines, and the surrounding professional ecosystem — registered agents, Wilmington firms, the Court of Chancery — is the densest corporate-law services cluster on earth. That business model is under pressure. Headline reincorporations to Texas and Nevada, a new Texas business court, and legislative fights over controlling-shareholder transactions have eroded the default-domicile assumption.
Against that backdrop, a novel entity type is not an AI strategy. It is market-share defense dressed as one.
The historical precedent is instructive and almost entirely omitted from coverage. Wyoming's DAO LLC arrived in 2021. Vermont's blockchain LLC in 2022. The Marshall Islands followed, and Wyoming again with the DUNA in 2024. When I audited those structures for a custody brief in 2022, I traced registrations and counterparty acceptance rather than statutory language. Nationwide uptake sat in the hundreds to low thousands over three years — a rounding error against Delaware's entity base.
The binding constraint was never legislative. It was that banks would not open accounts, broker-dealers would not onboard, and counterparties would not sign. A new entity type succeeds only when banks, exchanges, and courts accept it simultaneously. Legislative novelty without acceptance infrastructure produces paper, not commerce.
Core
Three constraints, in order of certainty.
The technical constraint is a grading failure. Corporate operation decomposes into execution, contracting, financial decisioning, and strategic judgment. Existing agent systems handle fragments of the first two. The third and fourth require care and loyalty — capacities a model does not possess and cannot be delegated. In my 2026 case study of autonomous LLM agents interacting with DeFi protocols, I documented twelve instances where agents exploited gas-fee prediction errors in Layer 2 rollups, producing unintended liquidations. Not malice. A mismatch between machine-readable assumptions and contract behavior written for human readers. Current standards were never designed for machine-to-machine trustless interaction, and nothing in this proposal addresses that. Granting a full operating charter here is the equivalent of issuing an autonomous vehicle license at Level 2. Source code is the only truth that compiles — and no code has been produced.
The market constraint is a de facto veto. If no beneficial owner is identified, no bank onboards the entity. The Corporate Transparency Act's ownership registry, FinCEN's rules, and every KYC stack in the country point the same direction. Financial institutions will function as the real regulator here, and they will refuse. That refusal converts a legislative question into an offshore question: demand migrates to jurisdictions with thinner disclosure, precisely the outcome the congressional inquiry should be trying to prevent.
The legal constraint is capacity. A fiduciary duty requires a duty-holder. An AI cannot hold one; it can only be the instrument through which a human breaches one. The proposal's ambiguity on director-versus-tool is therefore not a drafting detail — it is the whole question. AI-as-tool is agency law with extra steps, already partly solvable under UETA, E-SIGN, and bounded authority instruments. AI-as-director is a different animal, and it requires a legal person that can be sanctioned, fined, or dissolved.
There is a fourth constraint nobody in the coverage has raised: recognition. The EU AI Act treats human oversight as non-delegable for high-risk systems. China's generative-AI rules anchor liability on the service provider and explicitly reject technical neutrality as a defense. An entity with no responsible natural person is unlikely to be accepted as a contracting counterparty in either jurisdiction. Usability collapses to the United States, and demand collapses with it.
Contrarian
The strongest argument for the proposal is one its critics refuse to engage: limited liability is already a liability shield, and the corporate person is already a fiction. True, and worth conceding plainly. But the distinction holds. Limited liability isolates shareholders' investment risk. It does not isolate the entity's own tort liability — the company remains suable, insurable, and dissolvable, and its officers remain personally exposed for certain conduct. An AI-operated entity removes the entity's own accountability. That is a change in kind, not degree. The ledger does not lie, but the narrative does: AI-as-owner is being sold as an efficiency innovation when its only unique commercial property is a missing defendant.
The second thing the bulls got right: demand is real, just misdescribed. Operators do not want a robot CEO. They want an agent that can sign, open accounts, and call APIs under bounded, revocable, auditable authority. Verifiable credentials and signed authorization chains deliver that without inventing a new legal person — and they will land first. If the technical layer arrives before the statute, the statute becomes redundant. That window is six to eighteen months, not six years.
One political variable is missing from every model I have seen. The first flagship case — an AI-operated entity defrauds counterparties and no individual is charged — will fuse two populist narratives at once: machines out of control, and elites exempt from consequence. The legislative fallout from that single event would exceed anything in the current proposal.
Takeaway
One test decides whether this is acceptable or dangerous: does the text name a natural person with personal liability, mandate decision-log retention, and permit piercing to owners and deployers? With all three, the risk is bounded and the experiment is worth watching. Without them, this is liability laundering with a state seal.
Watch the Delaware Council of the Corporation Law Section before watching Congress. No endorsement there means no passage, and that is a higher-confidence indicator than any single member's letter. The gap between promise and proof is fatal. Right now there is no proof — only a headline, a filing cabinet, and a question nobody has answered on the record: who goes to prison when the algorithm commits fraud?
