Summary
- We have not identified a U.S. decision resolving whether, or how, common law agency principles govern an AI agent’s contracts, or whether the agent is simply a tool of the company that runs it, as of November 2026.
- Most precautions don’t depend on the answer. Limits enforced in systems and contracts, notice to counterparties and fast rejection of unwanted deals help under either view.
- Today’s default is the tool view. The Restatement says software can’t be an agent, and the electronic transactions statutes let contracts form through automated systems without human review. Whether a particular act is the company’s, and what effect it has, still depends on the facts, the agreement and other law. In 2026, the Ninth Circuit treated an AI shopping assistant as its user’s tool for computer-access law.
- The agent view would let companies rely on limits and the other side’s knowledge. Courts may borrow parts of it even while keeping the tool view.
- Plan for the tool view. It’s the likelier rule, and the harsher one for companies that deploy agents.
What holds under either view
A company that regularly pays for orders placed through its account can give a vendor reason to believe the next similar order is authorized. That history matters under either view.
- Assume the tool view. Plan as if what your agent does through your accounts may bind you. It’s the more conservative assumption, and in our view the likelier rule, though we haven’t found a decision settling it, as of November 2026.
- Put limits where they work under either view: in the systems the agent uses, and in contracts with the parties it deals with.
- Make limits visible. Notice to counterparties helps under both views.
- Act fast on errors. Give prompt notice disputing an unwanted deal and reserving rights, and then assess any right to cancel. Under the agent view, prompt notice helps defeat ratification; under the tool view, the deal may already bind you, so a flat rejection could itself be treated as a breach of contract. That is why the notice should dispute the deal and reserve rights, not just reject it.
For how to set these limits, see Technical Permission vs. Legal Authority (limits in the systems your agent uses, and making them visible) and Agent Use Clauses (limits in your contracts).
Where the answer changes the result
The views part ways in a few situations. In each, the facts worth establishing are the same; what changes is how much they count.
| Situation | Tool view | Agent view | Facts to establish |
|---|---|---|---|
| The other side knew, or should have known, the agent was outside its limits | Matters mainly through mistake, fraud and other defenses, and through any limits or security procedure the parties agreed on | Can defeat apparent authority, so the company may not be bound | What the counterparty was told and shown, and when |
| The agent went past an internal limit | Mostly a matter between the company and its own people and vendors | The limit defines actual authority | The written mandate, approvals and instructions in force |
| The company kept or used what the agent obtained | Often already bound, so this may add little | Knowing use can affirm an unauthorized deal | Which material facts the company knew, or whether it chose to affirm despite gaps |
| The agent handed work to another agent | Trace the disputed action to the party that used the system to take it | The subagent has actual authority only if the first agent could appoint one; the company can still be bound through its own conduct toward the other side, or by ratifying | The handoff instructions, the contract and the log showing which system placed the order |
Why the question matters
When an AI agent does something its company didn’t want, the company’s first question is whether it’s bound. The answer depends on which body of law a court reaches for.
If the agent is a tool, first identify the person or company that used the system to take the disputed action. Once that is established, ordinary contract rules and defenses determine its legal effect. If the agent is treated like a human agent, the court asks what the company approved, what the other side could reasonably believe, and whether the company accepted the result afterwards. Those questions give companies more ways out, and more ways to be caught.
View 1: it’s a tool
The Restatement of Agency takes this position directly. It says that, at present, computer programs are instrumentalities of the people who use them, and can’t be principals or agents under the common law.1 The reasoning is that only a person, meaning someone who can hold legal rights and owe legal duties, can be a principal or an agent. A program can’t, and if it malfunctions, even in ways no one anticipated, the consequences for the person using it are the same as when any other instrumentality malfunctions.
The electronic transactions statutes start from the same place. UETA allows contracts to form through electronic agents with no human review, and its official comments describe an electronic agent as a tool of the person who uses it.2 E-SIGN says a contract can’t be denied effect solely because electronic agents formed it, so long as the agent’s action is “legally attributable” to the person bound.3 That removes the objection to using agents; the contract must still be valid on other grounds.
Courts have reasoned this way before. In Bockhorst, an insurer’s computer reinstated a lapsed policy back to a date before an accident, after the company’s employees had learned of the accident. Applying Kansas waiver law, the court rejected the insurer’s attempt to avoid a knowing waiver merely because the reinstatement was processed by computer: a computer “operates only in accordance with the information and directions supplied by its human programmers.” The case turned on what the company knew; applying it to an autonomous agent is an analogy, not its holding.4 More recently, a Canadian tribunal refused to let an airline treat its chatbot as a separate entity responsible for its own statements.5
An appellate court applied the same view to a modern AI agent for computer-access law. Amazon sued Perplexity over an AI assistant that shopped on Amazon for its users. The Ninth Circuit held that the user, not Perplexity, accessed Amazon’s computers, “using the Assistant as an AI tool,” and vacated a preliminary injunction under the federal and California computer access statutes. The court limited the holding: it didn’t decide whether Perplexity could avoid liability in other contexts, including tort claims, and said it was not establishing “a new legal regime governing agentic AI.” It also noted that Amazon can still regulate access through its terms of service. The case says who “accesses” a computer; it doesn’t say an agent’s purchases bind its user, or that agents may ignore a platform’s contract terms.6
Recent scholarship agrees. Deborah DeMott concludes that agentic AI is “an instrumentality created by its developer and deployed by its user, and not the common law agent of either,” while arguing that agency law’s doctrines of apparent authority remain relevant when third parties reasonably rely on how the system appears.7
What it means: use of the company’s accounts is strong evidence that what the agent did is the company’s act. Whether a transaction binds the company also depends on the agreement, the rules on contract formation and any defenses, such as mistake. Internal limits the other side never saw usually don’t help; limits written into the parties’ agreement can. And if the agent causes harm rather than an unwanted contract, California no longer lets a developer or user defend on the ground that the AI acted autonomously (see Regulation & standards).
View 2: treat it like an agent
The case for this view is that modern agents do things a vending machine can’t. They interpret goals, choose tools, negotiate and take several steps in a row. Instructions to an AI agent look a lot like a principal’s instructions to a human agent, and the problems that follow look like agency problems: acting beyond instructions, misleading third parties, handing work to others.
Agency law already has tools for those problems: actual authority, apparent authority, ratification, subagency and rules on when authority ends.8 Borrowing them would let a company argue that an order was outside what it approved, and that the seller knew it.
The obstacles are real. The Restatement rejects the view. Agency law assumes an agent who owes duties and can be liable, and software can be neither. And UETA’s drafters, while anticipating agents that act “autonomously, and not just automatically,” chose to leave it to courts to read the definition of electronic agent to cover them, not to apply agency law.9
The likely middle ground
Courts may not have to choose. Several agency ideas turn on the principal’s conduct, not the agent’s nature, and those could apply even under the tool view.
- Appearances created by the company. Apparent authority depends on the principal’s manifestations to third parties. The Restatement defines it as a power held by “an agent or other actor,” which may leave room for non-agents.10 If a vendor sees the company identify the account as its purchasing channel, that can give the vendor reason to rely on orders from it. Insurers, for example, have been held to coverage their agents appeared authorized to bind.11
- Accepting the result. A company’s decision to keep a purchase after it learns the material facts can show that it adopted the deal. A warehouse receipt alone tells a vendor little about what the company knew.
- Notice. Under either view, a counterparty that knew the agent was acting outside its limits has a weaker claim.
The same facts can answer two different questions. Under the tool view, account records and past dealings can help show whose order it was. To establish apparent authority, the vendor must actually and reasonably believe the agent had authority for this order, and that belief must be traceable to the company’s words or conduct.
Others look past agency law altogether. Kevin Werbach argues that tort, contract and agency law can’t by themselves close the accountability gap agents create, and proposes a three-tier registration system for agents, from a unique identifier up to a posted bond, on the model of corporate registration.12
What doesn’t carry over. A human agent owes the principal duties of loyalty, care and obedience, and can be sued for breaching them. Software owes nothing. If courts borrow agency law, they will borrow the rules about the principal and third parties, such as authority, notice and ratification, not the agent’s own duties. Those duties fall on people instead: the employees who configure and supervise the agent, and the provider, to the extent its contract promises how the agent will behave. The rules on ending authority do carry over. Actual authority ends when the agent has notice that the principal has revoked it, but apparent authority can last until it is no longer reasonable for third parties to believe the agent still has authority.13 See Ending authority.
Side by side
| Question | Tool view | Agent view |
|---|---|---|
| Is the company bound by the agent’s deal? | Usually, if the action is attributable to it, subject to the parties’ agreement, formation rules and defenses | Yes, if the agent had actual or apparent authority |
| Do internal limits matter? | Mostly between the company and its own people and vendors | Yes, they define actual authority |
| Does the other side’s knowledge matter? | Through the parties’ agreement, such as agreed limits or security procedures, the rules on formation, and defenses such as mistake and fraud | Yes, it can defeat apparent authority |
| Can the company adopt a deal afterwards? | Usually unnecessary if the company already formed a contract through its system | Yes, by ratification |
| Who can be liable besides the company? | The provider, under its contract | The same; the software itself can’t be liable |
One instruction, several grants of authority
“The agent” is rarely given one kind of authority. Suppose Acme gives a separate cloud purchasing agent its own mandate and tells it to “reduce our cloud storage costs by 15% while keeping service levels.” To finish the job, the agent may need several different powers, and Acme can grant some without the others.
| Step | What the agent does | Authority it needs | Level |
|---|---|---|---|
| Research | Examines usage and market prices | Read access only; no effect outside Acme | 0 |
| Request bids | Asks vendors for quotes | Permission to contact suppliers for Acme | 1 |
| Negotiate | Bargains on price and terms | Authority to negotiate within set limits | 2 |
| Recommend | Proposes a vendor for a person to approve | None; a person makes the decision | 0 |
| Select | Chooses the vendor and tells it | Authority to choose among suppliers | 3–4 |
| Accept terms | Agrees to contract terms, including online terms | Authority over specific clauses | 3–4 |
| Sign and pay | Commits Acme and starts payment | Signing authority up to a limit, and use of a payment credential | 4 |
Levels refer to the authority scale used across this site: 0 Recommend, 1 Communicate, 2 Negotiate, 3 Agree subject to confirmation, 4 Execute within limits. Where a step shows 3–4, it is 3 if a person must confirm before Acme is bound, and 4 if the agent commits on its own.
The first four steps are generally lower risk to automate, if the system is built, and counterparties are told, so that only an authorized person can commit Acme. Even then, an outgoing offer, acceptance or concession can create obligations, and a false statement can create liability whether or not a contract forms, so limit what the agent may say and disclose. Calling the agent a negotiator doesn’t by itself tell the other side that its word isn’t final. The last three are where the legal structure is needed. Written down, the grant might read: the agent may buy cloud storage from approved vendors, up to $10,000 per order and $50,000 a month, on approved terms, but may not agree to auto-renewal, IP ownership terms, uncapped indemnities, arbitration or nonstandard data processing terms. Lawyers, procurement, security and engineering can all work from a sentence like that, and it helps under either view of what the agent is. See the Agent Mandate template.
A situation where both views agree
A vendor that has filled and been paid for similar orders has evidence of both attribution and apparent authority.
Acme’s agent places an order that breaks an internal rule. Is Acme bound?
Short answer: Probably, under either view. Acme paid earlier orders from the same account, giving the vendor a reasonable basis to believe this purchase was authorized.
Acme’s cloud purchasing agent uses Acme’s account with an approved vendor to order $8,000 of cloud storage at list price. Acme has paid for earlier orders placed the same way. Acme’s policy requires quotes from at least two vendors for orders over $5,000, but the rule appears only in that internal policy; the vendor never saw it.
Under the tool view, the use of Acme’s account and its history of paying similar orders support attributing this order to Acme,14 and nothing in these facts gives Acme a defense such as mistake. Under agency rules, giving the agent Acme’s account and paying for its earlier orders is the kind of conduct that creates apparent authority, and a limit the vendor never saw doesn’t defeat it.15 On these facts, Acme is likely bound under either view. The broken rule is a matter between Acme and the people who set up the agent.
The views part ways when the vendor had reason to doubt the order, for example because Acme had told it the limit, or because the order was far outside Acme’s usual pattern. See Where the answer changes the result.
- If you deploy the agent: Enforce rules like the two-quote requirement in the ordering system, not only in policy. If you want a vendor to be bound by a limit, tell the vendor.
- If you sell to the agent: An order that fits the buyer’s history is generally safe to fill. Confirm with a person when an order is unusual for that account.
Both sides of the deal
If you deploy agents
- Plan as if you’re bound by what your agent does through your accounts.
- Recover losses through your contract with the provider, since the law points at you first.
If you build agents
- Under the tool view your customer is likely bound, so its recovery runs through your contract. Set your position on indemnities and caps before a buyer asks.
- Describe capabilities accurately. Promises about what your agent won’t do become warranty claims.
- If your own sales or support agent talks to customers, you can be liable for what it says.
Open questions
- Will a court read UETA’s definition of electronic agent to cover systems that choose their own steps, as the comment invites?
- Does “other actor” in the Restatement’s definition of apparent authority reach software?
- If an agent misleads a business customer, is the company’s liability direct (its own statement) or vicarious (for its agent’s)? The answer may affect defenses and insurance.
- Restatement (Third) of Agency § 1.04 cmt. e (opens in a new tab) (Am. L. Inst. 2006). ↩
- Unif. Elec. Transactions Act §§ 2(6) & cmt. (“Electronic agent”), 14 (opens in a new tab) (Unif. L. Comm’n 1999). ↩
- 15 U.S.C. § 7001(h) (opens in a new tab). ↩
- State Farm Mut. Auto. Ins. Co. v. Bockhorst (opens in a new tab), 453 F.2d 533 (10th Cir. 1972). ↩
- Moffatt v. Air Canada (opens in a new tab), 2024 BCCRT 149 (B.C. Civ. Res. Trib.). ↩
- Amazon.com Servs. LLC v. Perplexity AI, Inc. (opens in a new tab), 184 F.4th 1083, 1091–94 & n.5 (9th Cir. 2026), reh’g en banc denied (opens in a new tab), No. 26-1444 (9th Cir. Sept. 10, 2026). ↩
- Deborah A. DeMott, When Agentic AI Met the Common Law of Agency (opens in a new tab), Harv. J.L. & Tech. (forthcoming 2026), manuscript at 1 (quotation), 10–13 (apparent authority). ↩
- Restatement (Third) of Agency §§ 2.01, 2.03, 3.11, 3.15, 4.01 (opens in a new tab). ↩
- Unif. Elec. Transactions Act § 2 cmt. (“Electronic agent”) (opens in a new tab). ↩
- Restatement (Third) of Agency § 2.03 (opens in a new tab). ↩
- See, e.g., Gallant Ins. Co. v. Isaac (opens in a new tab), 751 N.E.2d 672 (Ind. 2001) (agency had apparent authority to bind coverage); T-Mobile USA, Inc. v. Selective Ins. Co. of Am. (opens in a new tab), 450 P.3d 150 (Wash. 2019) (insurer bound by agent’s representation in a certificate of insurance). ↩
- Kevin Werbach, Agents, Inc. (opens in a new tab), 71 Vill. L. Rev. 591 (2026). ↩
- Restatement (Third) of Agency §§ 3.06–3.11 (opens in a new tab). ↩
- See Unif. Elec. Transactions Act §§ 9, 14 (opens in a new tab); Restatement (Third) of Agency § 1.04 cmt. e (opens in a new tab). ↩
- Restatement (Third) of Agency §§ 2.03, 3.03 (opens in a new tab). ↩