Updated 2026-10-12
from the news desk
Can AI Be a Legal Agent? Agency law says no, but AI acts can still bind a principal
Deborah A. DeMott says agentic AI is not an agent in the legal sense because agency requires a consensual relationship between persons and duties an AI cannot owe. Mark O. Riedl and Deven R. Desai write that agency law still helps analyze…
Deborah A. DeMott says AI cannot be a legal agent in the common-law sense she describes, because agency requires a relationship between persons and duties an AI cannot owe. Riedl and Desai likewise write that a software agent is not the same as a human agent, even when it acts for a user. The harder legal question is usually not whether the AI itself is an agent, but when a person or company is bound or exposed because it used AI as an intermediary.
The short answer under agency law
DeMott puts the answer directly. In Duke Law’s notice about her work on agentic AI, she says: “Importantly, despite its name, nothing about agentic AI establishes a legal relationship of agency.” The same notice says a software tool is not a person, and that agency requires an ongoing consensual relationship between principal and agent.
Riedl and Deven R. Desai make the same basic distinction in their paper. They write that the law of agency concerns one human acting on behalf of another in a fiduciary relationship. Fiduciary here means a relationship structured around duties to act for another’s benefit. Riedl and Desai add that software agents should not be anthropomorphized, and that human agents can face legal penalties in a way software cannot.
The American Law Institute says the Restatement of the Law Third, Agency covers actual authority, apparent authority, contracts with third parties, tort liability, and duties between agent and principal. One implication is that agency is a defined legal framework, not just a casual label for any system that follows instructions.
Why common-law agency does not fit AI
DeMott’s reason for saying no is structural, not semantic. The Duke Law notice says agency stipulates a consensual relationship between two persons, principal and agent, and that this category includes legal persons such as corporations. The same notice says AI itself is not capable of owing a duty, meaning a legally enforceable obligation, to anyone.
Riedl and Desai reinforce that point by reproducing the Restatement definition of agency as a relationship in which one person manifests assent to another person to act on the principal’s behalf and under the principal’s control, and the other person consents so to act. Their paper treats that two-person structure as a central difference between legal agency and software agency.
That no answer does not depend on denying that AI systems can act autonomously in a practical sense. DeMott’s point is narrower. An autonomous system may plan, communicate, or transact, but those features do not themselves create the legal relationship called agency.
There is still a live push in the other direction. DeMott notes that some people argue for personhood for agentic AI, and she answers with a practical question: “Some people argue for personhood for agentic AI, and then people like me ask, 'Well, how would an AI buy liability insurance?'” One reading of her response is that the argument for AI as a legal agent would need more than better software. It would need some route to legal personhood or a comparable legal status, because the barrier she identifies is built into the definition of agency she describes.
That personhood question is separate from this site’s explainer on electronic personhood. DeMott’s narrower claim is that under the agency law she is discussing, AI is not already an agent just because people call it one.
How AI can still have legal effects for a person or company
Saying AI is not itself a legal agent does not mean its conduct is legally irrelevant. DeMott says: “Agency doctrine can also be relevant when people and enterprises choose to interact with others via AI agents.” Her point is that agency doctrine can apply around the AI-mediated interaction even when the software is not the agent in law.
The Duke Law notice illustrates that with Moffatt v. Air Canada. The notice says a chatbot on Air Canada’s website told a customer he could apply for a bereavement discount after buying a ticket, and that a small claims court held the airline responsible for the misleading information. DeMott says the reasoning has parallels in apparent authority and apparent agency, because the airline created what appeared to be a legally consequential way for third parties to communicate with it. She sums up the principle this way: “If you encourage people to interact with something that you've held out as your legally consequential intermediary, that can carry legal consequences.”
Riedl and Desai make a related point from the side of contract and disclosure rules. They write that when an agent discloses the principal and has authority, the contract is treated as one between the third party and the principal. They also write that if the principal is not disclosed, the third party may treat the agent as the principal. Their warning for AI is that these issues arise when software interacts with outsiders, not only with the user behind the screen. As they put it: “AI agents raise these potential problems precisely because they interact with third parties.”
One reading of DeMott and of Riedl and Desai is that AI often functions less like a rights-bearing legal actor and more like an intermediary whose conduct can still be attributed to a human or firm. That is close to, but not the same as, whether AI can sign a contract.
Why agency concepts still matter for autonomous AI
Riedl and Desai argue that agency law remains useful because software agents and legal agents share some practical problems even when they are not the same thing. They write: “Both computer science and the law recognize the problems of under-specification for an agent.” In their account, both fields have to deal with under-specification, meaning that instructions leave gaps and the actor must fill them in.
Their paper compares legal ideas such as actual and implied authority with AI ideas such as value alignment. Value alignment is their term for training or designing a system so it stays within human aims when the objective is incomplete. Riedl and Desai say today’s alignment work has focused on helpfulness, honesty, and harmlessness, but has paid less attention to loyalty and disclosure, both of which agency law treats as central.
That is the strongest case for borrowing agency concepts without declaring AI a legal agent. If a system can shop, negotiate, reserve, or communicate on someone’s behalf, then questions about whose interests it must serve, what it must reveal to third parties, and when a principal is bound do not disappear. Riedl and Desai say loyalty especially matters because a platform may have interests of its own that conflict with the user’s interests. They write: “Loyalty extends the concept of helpfulness in a way that also intersects harmlessness.”
This is also where the topic differs from whether AI can be an employee or from debates about moral blame. The legal issue here is not whether the model deserves praise or punishment as a moral actor. It is whether the law should treat the system as the agent, or instead treat the person or company behind it as the legally relevant actor.
Where the real disagreement sits
On the sources here, the direct answer to can AI be a legal agent is no. DeMott says agentic AI does not establish an agency relationship, and Riedl and Desai say software agents are not the same as human agents under agency law.
The more ambitious counter-position is better framed as a proposal than as a description of present doctrine. If agency law requires persons, then one path toward AI as a true legal agent would be to change the legal status of AI so that it could count as a person for these purposes. DeMott acknowledges that people make personhood arguments, even as she questions how that would handle ordinary agency features such as liability.
Even without that step, both DeMott and Riedl and Desai say agency law is still useful because it helps sort out authority, appearance, disclosure, loyalty, and liability when AI acts in the world. So the real dispute is often not over whether the chatbot or software tool is already an agent in law. It is over whether existing doctrines are enough to attribute its acts to a principal, or whether new rules would be needed if societies ever wanted AI itself to occupy the agent’s legal role.
Frequently asked questions
If AI cannot be a legal agent, can its actions still bind a company?
Yes, in some situations. DeMott says agency doctrine can matter when a company holds out AI as its legally consequential intermediary, and the Duke Law notice says the court in Moffatt v. Air Canada held the airline responsible for misleading chatbot information on its website.
Would AI personhood change whether AI can be a legal agent?
It could change the analysis described by DeMott, because she says agency requires a relationship between persons and includes legal persons such as corporations. But DeMott presents personhood as an argument some people make, not as the legal rule she says governs AI now.
Why do lawyers use agency law for AI if AI is not an agent in law?
Riedl and Desai say agency law still helps with problems of authority, loyalty, disclosure, and third-party dealings when AI systems act for users. DeMott likewise says agency doctrine can be relevant when people and enterprises choose to interact with others via AI agents.
Does agency law treat AI more like a person or more like a tool?
DeMott’s discussion points toward tool or instrumentality, not person. The Duke Law notice says AI cannot owe duties and compares one possibility to a non-culpable instrumentality that still creates consequences for the human behind it.
Caveats: This explainer relies mainly on Deborah A. DeMott’s account in Duke Law’s 27 July 2026 notice and on Mark O. Riedl and Deven R. Desai’s 2025 paper. Specific examples such as Moffatt v. Air Canada appear in the Duke Law notice, and the extended discussion of loyalty and disclosure appears in Riedl and Desai’s paper.