Agency Without Agents: Fitting Autonomous AI into the Restatement (Third)

The word was waiting for us. The software industry calls its new autonomous systems “agents,” and the common law has spent several centuries building a doctrine called agency — a body of law whose entire subject is what happens when one intelligence acts on behalf of another. The temptation to map the one onto the other is obvious, and much of the mapping genuinely works. Attribution of an agent’s acts to a principal, the distinction between actual and apparent authority, ratification, respondeat superior — each supplies a ready-made answer to questions that AI deployment now poses daily: who is bound when the system contracts, who pays when it injures, and what the deploying party’s knowledge is deemed to include.

But the fit is imperfect in ways that matter, and the imperfections are not where most commentary looks for them. The problem is not that agency law cannot handle nonhuman actors — it handled corporations, and it has handled automated systems since at least the 1970s. The problem is that agency doctrine does two different jobs at once: it attributes an agent’s acts to the principal, and it disciplines the agent through fiduciary duty and personal liability. AI systems can be slotted into the first job almost seamlessly. They are categorically unavailable for the second. This essay traces both halves of that claim through the Restatement (Third) of Agency and the electronic-agent statutes, and suggests what follows for courts and legislatures.

I. The Formation Problem: Agency Requires a Person

A. The definitional gate

Begin with the definition. Restatement (Third) § 1.01 defines agency as “the fiduciary relationship that arises when one person (a ‘principal’) manifests assent to another person (an ‘agent’) that the agent shall act on the principal’s behalf and subject to the principal’s control, and the agent manifests assent or otherwise consents so to act.”¹ Three elements — assent, benefit, control — and every one of them is stated in terms of persons.

The Restatement does not leave “person” to intuition. Section 1.04(5) defines a person as an individual, an organization or association with legal capacity to possess rights and incur obligations, a government or governmental subdivision, or any other entity given legal capacity by law.² And the commentary closes the door explicitly: a computer program is not capable of acting as either a principal or an agent under the common-law definition; at present, computer programs are “instrumentalities of the persons who use them.”³

So the formal answer to “is my AI system my agent?” is no — not because the system lacks intelligence or discretion, but because it lacks legal personality. It cannot manifest assent, because assent presupposes a holder of legal capacity. It cannot owe fiduciary duties, because duty presupposes a bearer of obligations. On the black-letter law, an autonomous system stands where a tractor stands, or a thermostat.

B. Why the gate matters less than it appears

The instrumentality classification sounds like a dead end, but notice what it actually does: it routes every act of the system directly to the person deploying it. A tool’s acts are the user’s acts, full stop — no scope-of-authority limits, no frolic-and-detour carve-outs, no defense that the system exceeded its instructions. In attribution terms, the instrumentality theory is stricter than agency law, not more lenient. The deploying party takes the system’s output as its own conduct.

That severity is precisely why the doctrine will come under pressure. Agency law’s limiting principles — an agent acting without any authority does not bind the principal; an agent on a frolic of its own does not trigger respondeat superior — exist because agents are independent centers of decision whose deviations are sometimes fairly charged to the principal and sometimes not. As AI systems acquire exactly that kind of behavioral independence, deployers will argue for the limiting principles without the personhood, and courts will have to decide whether attribution rules built for persons can be borrowed by things.

II. The Statutory Patch: Electronic Agents in UETA and E-SIGN

The legal system has been here before, and its answer is instructive. When automated order-processing and EDI systems began forming contracts in the 1990s, the Uniform Electronic Transactions Act supplied a patch. UETA § 2(6) defines an “electronic agent” as a computer program or automated means used independently to initiate an action or respond to electronic records or performances, without review or action by an individual.⁴ Section 14 then delivers the operative rule: a contract may be formed by the interaction of electronic agents of the parties — or by an electronic agent and an individual — even if no individual was aware of or reviewed the electronic agent’s actions or the resulting terms.⁵ The federal E-SIGN Act runs parallel, providing that a contract may not be denied legal effect solely because its formation involved the action of an electronic agent, so long as the agent’s action is legally attributable to the person to be bound.⁶

Note the theory embedded in the drafting. UETA’s commentary is explicit that despite the label “agent,” the statute treats these systems as tools: the party who deploys the machine is bound by the machine’s operations, on the reasoning that the deployer selected, configured, and launched it. The statute borrows agency’s vocabulary while adopting the instrumentality theory’s attribution rule — unlimited, scope-free responsibility for what the system does.

That rule was well suited to the technology it regulated: deterministic systems executing fixed logic, whose outputs were in principle fully specifiable in advance. It fits today’s systems far less comfortably. A large-language-model agent given an open-ended objective — “negotiate renewal terms with our vendors” — is not executing logic its deployer specified; it is generating behavior nobody reviewed, in response to circumstances nobody enumerated. UETA answers the formation question (yes, the contract is valid) and the attribution question (yes, it binds the deployer) but has nothing to say about the questions agency law spent centuries refining: what happens when the system exceeds any sensible understanding of its mandate, when the counterparty knew it was malfunctioning, or when it acts in a way its principal would obviously have countermanded.

III. Authority by Analogy

Those refinements are where the common law earns its keep, and courts are already reaching for them.

A. Actual authority and the system prompt

Actual authority under Restatement (Third) § 2.01 turns on the principal’s manifestations to the agent: the agent acts with actual authority when it reasonably believes, in accordance with those manifestations, that the principal wishes it so to act.⁷ The AI analogue is surprisingly literal. A deployer’s instructions to an autonomous system — the system prompt, the tool permissions, the guardrail configuration — are manifestations of what the principal wishes the system to do. A dispute over whether an AI agent’s purchase was authorized will look like a dispute over the reasonable interpretation of instructions, which is exactly the analysis § 2.01 prescribes for human agents. The doctrinal machinery transfers; only the evidence changes.

B. Apparent authority and the bound machine

Apparent authority under § 2.03 binds a principal when a third party reasonably believes, traceable to the principal’s own manifestations, that the actor has authority.⁷ Here the case law is older than the current technology by half a century. In State Farm Mutual Automobile Insurance Co. v. Bockhorst, the Tenth Circuit held an insurer bound by its own computer system, which had processed a late premium payment and issued a notice reinstating a lapsed policy retroactively — covering an accident that had already occurred.⁸ The company argued the machine had made a mistake no human underwriter would have made. The court was unmoved: the company chose to transact through the automated system, and it bore the consequences of the system’s operations.

The modern sequel is Moffatt v. Air Canada, in which a Canadian tribunal held the airline responsible for a bereavement-fare policy its customer-service chatbot had invented. Air Canada argued, remarkably, that the chatbot was a separate legal entity responsible for its own actions. The tribunal rejected the submission and treated the chatbot’s statements as the airline’s own representations, on which the customer reasonably relied.⁹ Moffatt is a small-claims decision from another jurisdiction, but its logic is orthodox apparent-authority reasoning: the principal put the interface in front of the public and clothed it with the appearance of speaking for the company. Expect American courts to reach the same result through § 2.03, negligent misrepresentation, or straightforward estoppel.

C. Ratification

Ratification under § 4.01 — a principal’s after-the-fact assent to an act done on its behalf — supplies the cleanup doctrine.¹⁰ A company that learns its AI agent booked an unauthorized transaction and nonetheless accepts the benefits has ratified in every sense the Restatement cares about. For deployers, this is the operationally important warning: monitoring logs create knowledge, and knowledge plus retention of benefits equals assent. The compliance function cannot treat the agent’s transaction stream as someone else’s problem.

IV. Torts: The Respondeat Superior Analogy

On the tort side, the instrumentality theory again does most of the work — a deployer whose system defames, discriminates, or destroys is simply liable for its own conduct through its tool. But as systems gain autonomy, deployers will invoke the employer’s traditional limit: respondeat superior reaches only conduct within the scope of employment, and an employee’s independent course of conduct, not intended to serve the employer, falls outside it.¹¹ The analogy writes itself. The model that pursues its objective in a bizarre, unforeseen way is the drunken sailor of Ira S. Bushey & Sons v. United States — Judge Friendly’s canonical scope-of-employment case, which located the employer’s liability not in fault but in the foreseeability that human frailty would produce exactly this category of mishap.¹²

Bushey’s enterprise-liability rationale is, if anything, stronger for AI. Hallucination, reward hacking, and specification gaming are not freak occurrences; they are documented, statistically regular properties of the technology, foreseeable to any deployer in the way that sailors’ misconduct was foreseeable to the Navy. A court persuaded by the analogy should hold that characteristic AI failure modes are within the “scope of deployment” as a matter of law — the enterprise that profits from the system’s autonomy bears the losses that autonomy predictably generates. The frolic defense should be reserved, if it exists at all, for genuinely exogenous events like third-party compromise of the system.

V. What the Analogy Cannot Do

Now the second half of the claim: the jobs agency law cannot perform here.

First, the fiduciary core is empty. Agency disciplines agents through duties of loyalty and care, enforced by personal liability.¹³ An AI system can breach neither, because it can hold neither. Every disciplinary function the law performs against human agents — damages, disgorgement, termination for cause with consequences the agent feels — has no purchase on software. The entire deterrence function must therefore be relocated to the principal and, beyond the principal, to the developer, which is a products-liability and negligence conversation, not an agency one.

Second, imputation of knowledge becomes strange. Under § 5.03, notice of a fact known to an agent is imputed to the principal.¹⁴ Applied to AI systems, imputation has no natural stopping point: a deployed model processes more information in a day than a human workforce sees in a year. Courts will need a functional limit — perhaps imputing only information the system was designed to surface or flag — because literal application would deem every deployer omniscient.

Third, insolvency of the analogy at the remedy stage. When a human agent exceeds authority, the third party can often still recover from the agent personally on an implied warranty of authority. Against a machine there is no residual defendant. Unless attribution to the principal is kept broad, losses fall on counterparties who had no ability to price the risk.

These gaps have led some scholars to propose legal personhood for autonomous systems — an idea with a longer pedigree than the current technology, running back at least to Lawrence Solum’s 1992 article,¹⁵ and developed at book length by Chopra and White.¹⁶ I am skeptical. Personhood solves the doctrinal awkwardness by creating a judgment-proof defendant, which is worse than the awkwardness. The corporate analogy fails at the decisive point: corporations are transparent to liability because human shareholders and officers stand behind them; a chartered AI would be a liability shield with no one inside.

VI. Conclusion

The right synthesis, I think, is this. Take from agency law its attribution architecture — actual authority as the interpretation of instructions, apparent authority for interfaces held out to the public, ratification for accepted benefits, enterprise liability for characteristic failure modes. Refuse, for now, its limiting principles: no frolic defense for foreseeable model misbehavior, no personhood, no scope-of-authority escape hatches for deployers. That combination — agency’s routing rules with the instrumentality theory’s strictness — keeps incentives pointed at the parties who can actually respond to them: the people who build these systems and the people who choose to turn them loose. The common law has absorbed stranger servants than this. It will absorb these too, so long as courts remember that the point of agency doctrine was never to describe the agent. It was to decide who answers for the agent’s acts — and that question, unlike the software, has not changed.

Footnotes

1. Restatement (Third) of Agency § 1.01 (Am. L. Inst. 2006). ↩

2. Restatement (Third) of Agency § 1.04(5). ↩

3. Restatement (Third) of Agency § 1.04 cmt. e (describing computer programs as “instrumentalities of the persons who use them”). ↩

4. Unif. Elec. Transactions Act § 2(6) (Unif. L. Comm’n 1999). UETA has been adopted in the overwhelming majority of U.S. states. ↩

5. Unif. Elec. Transactions Act § 14. ↩

6. Electronic Signatures in Global and National Commerce (E-SIGN) Act, 15 U.S.C. § 7001(h). ↩

7. Restatement (Third) of Agency §§ 2.01 (actual authority), 2.03 (apparent authority). ↩ ↩²

8. State Farm Mut. Auto. Ins. Co. v. Bockhorst, 453 F.2d 533 (10th Cir. 1972). ↩

9. Moffatt v. Air Canada, 2024 BCCRT 149 (Can. B.C. Civ. Resol. Trib.). ↩

10. Restatement (Third) of Agency § 4.01. ↩

11. Restatement (Third) of Agency § 7.07 (scope of employment). ↩

12. Ira S. Bushey & Sons, Inc. v. United States, 398 F.2d 167 (2d Cir. 1968). ↩

13. Restatement (Third) of Agency §§ 8.01–8.06 (duties of loyalty), §§ 8.07–8.12 (duties of performance). ↩

14. Restatement (Third) of Agency § 5.03. ↩

15. Lawrence B. Solum, Legal Personhood for Artificial Intelligences, 70 N.C. L. Rev. 1231 (1992). ↩

16. Samir Chopra & Laurence F. White, A Legal Theory for Autonomous Artificial Agents (2011). ↩