Authorities
What the law already says
Few decisions involve modern AI agents. Entries are grouped by subject, generally with statutes, regulations and cases first. Each entry is labeled by the type of source it is. The label doesn’t say how much weight a source carries; that depends on the jurisdiction and the question. Entries marked Analogue were decided about earlier automated systems or in another country, so they apply by analogy rather than directly.
| Authority | Type | What it says | Why it matters for agents |
|---|---|---|---|
| Agency, authority and attribution | |||
| Restatement (Third) of Agency § 1.04 cmt. e (opens in a new tab)Am. L. Inst. 2006 | Restatement | At present, computer programs are instrumentalities of the people who use them. They can’t be principals or agents under the common law. | Sets the default view. Anyone applying agency rules to AI agents has to argue around it. |
| Restatement (Third) of Agency § 2.01 (opens in a new tab)Am. L. Inst. 2006 | Restatement | An agent has actual authority when, at the time it acts, it reasonably believes, from the principal’s manifestations to it, that the principal wants it to act. | Defines what the company authorized the agent to do, if agency rules are borrowed. |
| Restatement (Third) of Agency § 2.02 (opens in a new tab)Am. L. Inst. 2006 | Restatement | Actual authority covers acts designated or implied by the principal and acts necessary or incidental to achieving the principal’s objectives, as the agent reasonably understands them when it acts. | Frames whether an agent’s unplanned step, such as an emergency purchase, was within its authority. |
| Restatement (Third) of Agency § 2.03 (opens in a new tab)Am. L. Inst. 2006 | Restatement | Apparent authority exists when a third party reasonably believes the actor has authority, and the belief is traceable to the principal. | The main way a company’s internal limits fail against outsiders. |
| Restatement (Third) of Agency § 3.03 (opens in a new tab)Am. L. Inst. 2006 | Restatement | Apparent authority arises when a third party reasonably believes the agent can act, because of something the principal said or did. | Giving an agent the company’s account and paying for its earlier orders is the kind of conduct a vendor will point to. |
| Restatement (Third) of Agency §§ 3.06–3.10 (opens in a new tab)Am. L. Inst. 2006 | Restatement | How actual authority ends. A revocation takes effect when the agent has notice of it. | How software receives notice is unsettled, so withdraw the agent’s credentials and change its instructions at the same time. |
| Restatement (Third) of Agency § 3.11 (opens in a new tab)Am. L. Inst. 2006 | Restatement | Apparent authority ends when it’s no longer reasonable for the third party to believe the actor still has authority. | Why retiring an agent means telling the vendors it dealt with. |
| Restatement (Third) of Agency § 3.15 (opens in a new tab)Am. L. Inst. 2006 | Restatement | Defines a subagent: a person an agent appoints to do work the agent agreed to do for the principal, and for whose conduct the agent is responsible. An agent may appoint one only if it has actual or apparent authority to do so. | The starting point for agents that hand work to other agents. |
| Restatement (Third) of Agency §§ 4.01, 4.06 (opens in a new tab)Am. L. Inst. 2006 | Restatement | Ratification is accepting a prior act, which gives it effect as if it had been authorized. | Keeping what an agent bought can show affirmance if the company knew the material facts or chose to affirm despite missing facts (see Henderson v. United Student Aid Funds, Inc., 918 F.3d 1068, 1073–74 (9th Cir. 2019) (opens in a new tab), quoting § 4.06). |
| Gallant Ins. Co. v. Isaac (opens in a new tab)751 N.E.2d 672 (Ind. 2001) | Case | An insurance agency had apparent authority to bind coverage. | An example of a company held to what its representative appeared authorized to do. |
| T-Mobile USA, Inc. v. Selective Ins. Co. of Am. (opens in a new tab)450 P.3d 150 (Wash. 2019) | Case | An insurer was bound by its agent’s representation in a certificate of insurance. | Another example of apparent authority binding the company behind the representative. |
| State Farm Mut. Auto. Ins. Co. v. Bockhorst (opens in a new tab)453 F.2d 533 (10th Cir. 1972) | Analogue | Applying Kansas waiver law, held that an insurer whose employees knew of an accident waived its defense when its computer reinstated the policy; the computer’s role didn’t excuse the company. | An early statement that a business can’t disown what its automated systems do. Extending it to autonomous agents is an analogy; the case turned on the company’s knowledge. |
| Moffatt v. Air Canada (opens in a new tab)2024 BCCRT 149 (B.C. Civ. Res. Trib.) | Analogue | An airline was liable for wrong information from its website chatbot and couldn’t treat the bot as a separate entity. | The closest decision to modern AI agents speaking for a company, though it is a small-claims tribunal ruling that binds no other court and is at most persuasive. |
| 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, 10–13 | Scholarship | Agentic AI is an instrumentality, not a common law agent, but apparent authority doctrines remain relevant when third parties reasonably rely on how it appears. | Supports the tool view while keeping apparent authority in play. |
| Kevin Werbach, Agents, Inc. (opens in a new tab)71 Vill. L. Rev. 591 (2026) | Scholarship | Existing law can’t fully close the accountability gap; proposes a three-tier registration system for agents. | A registry-based alternative to fitting agents into agency law. |
| Electronic contracting and contract terms | |||
| UETA § 2(6) & cmt. (“Electronic agent”) (opens in a new tab)Unif. L. Comm’n 1999 | Statute | Defines an electronic agent. The comment calls it a tool of the person who uses it, and anticipates agents that act autonomously. | The statute’s own framing of the tool view, and its invitation to courts. |
| UETA § 2(8) (opens in a new tab)Unif. L. Comm’n 1999 | Statute | Defines an electronic signature as a sound, symbol or process attached to or logically associated with a record and executed or adopted by a person with intent to sign. | When an agent applies a signature, the question is whether the signature was the company’s act. |
| UETA § 3(b) (opens in a new tab)Unif. L. Comm’n 1999 | Statute | Leaves some transactions outside the Act, including wills and most of the UCC (other than Articles 2 and 2A and pre-revision §§ 1-107 and 1-206, which Revised Article 1 renumbered to § 1-306 and repealed, respectively). | UETA’s rules on electronic records, signatures and agents don’t help with those transactions. |
| UETA § 5 (opens in a new tab)Unif. L. Comm’n 1999 | Statute | UETA applies where the parties agreed to deal electronically; most of its rules can be changed by agreement. | Why contracts can set the rules for agent transactions. |
| UETA § 7(c)–(d) (opens in a new tab)Unif. L. Comm’n 1999 | Statute | Where a law requires a writing or a signature, an electronic record or electronic signature satisfies it. | Records and signatures an agent produces can meet writing and signature requirements, subject to the Act’s other conditions. |
| UETA § 9 (opens in a new tab)Unif. L. Comm’n 1999 | Statute | An electronic record is attributable to a person if it was that person’s act, including an act done through its electronic agent. The act can be shown in any manner, including by the security procedure used. | Ties an agent’s actions to the company whose credentials it used. |
| UETA § 10 (opens in a new tab)Unif. L. Comm’n 1999 | Statute | Rules for changes and errors: allocates loss for a change or error in transmission when an agreed error-detecting security procedure wasn’t followed, and lets an individual avoid their own error in dealing with another person’s electronic agent that gave no chance to prevent or correct it (on prompt notice, reasonable steps to return what was received or destroy it if instructed, and no use of or benefit from it). Neither the individual-error rule nor the fallback to other law and the parties’ contract can be varied by agreement (§ 10(4)). | Where two companies agreed on an error-detecting procedure, the one that followed it can avoid a transmission error the other would have caught. For an agent’s decision transmitted accurately, the contract and ordinary mistake law may still provide relief. The individual rule reaches only natural persons, such as sole proprietors. |
| UETA § 14 (opens in a new tab)Unif. L. Comm’n 1999 | Statute | A contract can form between electronic agents even if no person knew of or reviewed their actions. | The main route for holding a company to its agent’s deals. Law in 49 states; New York has its own statute. |
| 15 U.S.C. § 7001(h) (opens in a new tab)E-SIGN | Statute | A contract isn’t invalid because electronic agents formed it, if the action is legally attributable to the person bound. | The federal counterpart to UETA § 14. |
| 15 U.S.C. §§ 7001(a), (e), 7003, 7006(5) (opens in a new tab)E-SIGN | Statute | A signature, contract or record can’t be denied legal effect solely because it is electronic. Where a law requires a writing, an electronic record can be denied effect if it can’t be kept and accurately reproduced for everyone entitled to retain it. The Act defines an electronic signature (opens in a new tab). Its exceptions (opens in a new tab) make § 7001 inapplicable to records governed by laws on making wills, most of the UCC (other than Articles 2 and 2A and pre-revision §§ 1-107 and 1-206, now § 1-306 and repealed under Revised Article 1), and specified notices, among other exceptions. | One more reason to keep transaction records and logs that hold them in a form you can produce. |
| N.Y. State Tech. Law §§ 301–309 (opens in a new tab)New York (Electronic Signatures and Records Act) | Statute | New York hasn’t adopted UETA and relies on this Act instead. It has no counterpart to UETA’s rules on electronic agents. | Under New York law, the electronic agent rule comes from E-SIGN, and attribution turns on general law such as agency. |
| Ill. Pub. Act 102-38 (opens in a new tab)Illinois 2021 (815 Ill. Comp. Stat. 333) | Statute | Illinois’s enactment of UETA, the most recent. Every state except New York (opens in a new tab) has adopted the Act. | UETA’s electronic agent rules apply almost everywhere, but several states changed the uniform text, so check the version in the state whose law governs. |
| Cal. Civ. Code § 1633.5(b) (opens in a new tab)California UETA | Statute | An agreement to conduct a transaction electronically may not be contained in a non-electronic standard form contract unless it is a separate and optional agreement whose primary purpose is to authorize electronic transactions, and the contract may not be conditioned on that consent. This subdivision can’t be varied by agreement. | When adopting an agent-transaction addendum through a paper standard form under California law, obtain consent to transact electronically separately and optionally. |
| U.C.C. § 1-303(b), (d) (opens in a new tab)Article 1 | Statute | A course of dealing between the parties can supply or explain the terms of their deals. | Orders an agent placed that were accepted and paid without objection help show what both sides understood. |
| U.C.C. § 2-204(1) (opens in a new tab)Article 2 | Statute | A contract for the sale of goods may be made in any manner sufficient to show agreement, including conduct. | Sales law doesn’t require a person on either side to form the contract. |
| U.C.C. § 2-207 (opens in a new tab)Article 2 | Statute | Decides whether a contract forms, and which terms control, when an acceptance or written confirmation contains terms that add to or differ from the offer (the “battle of the forms”). Applies to sales governed by a state’s Article 2; Louisiana has its own rule (La. Civ. Code art. 2601). | Governs conflicts when two companies’ agents exchange forms in a sale governed by Article 2. |
| La. Civ. Code art. 2601 (opens in a new tab)Louisiana | Statute | An acceptance that adds to or differs from the offer still forms a sale of movables if the parties agree on the thing and the price, unless it is conditioned on the offeror’s assent to the new terms. Between merchants, additional terms become part of the contract unless they materially alter it, the offer limits acceptance to its terms, or the offeror objects. | Louisiana’s counterpart to UCC § 2-207 when agents exchange conflicting forms. |
| U.N. Convention on Contracts for the International Sale of Goods, arts. 1, 6, 19 (opens in a new tab)Treaty; in force for the U.S. since 1988 | Treaty | Governs sales of goods between parties whose places of business are in different contracting states, unless they exclude it. A reply that purports to accept but materially changes the offer is generally a rejection and counteroffer. | With an international supplier, it can decide which terms control instead of UCC § 2-207. Choosing a U.S. state’s law doesn’t necessarily exclude it (Asante Techs. v. PMC-Sierra, N.D. Cal. 2001). |
| Asante Techs., Inc. v. PMC-Sierra, Inc. (opens in a new tab)164 F. Supp. 2d 1142 (N.D. Cal. 2001) | Case | A contract’s choice of California law didn’t exclude the CISG, which as a federal treaty is part of California law, so the Convention governed a sale between a U.S. buyer and a seller whose relevant place of business was in Canada. | To exclude the Convention, a contract with an international supplier needs clear language showing that intent; a bare choice of a state’s or province’s law isn’t enough. |
| Md. Code Ann., Com. Law §§ 22-107(d) (opens in a new tab), 22-112(b), (e) (opens in a new tab); Va. Code Ann. §§ 59.1-501.7(d) (opens in a new tab), 59.1-501.12 (opens in a new tab), 59.1-501.13:1(b) (opens in a new tab)Maryland and Virginia UCITA, 2000 | Statute | For computer information transactions, such as software licenses, an electronic agent has an opportunity to review a term only if it is made available in a way that would enable a reasonably configured electronic agent to react to it, and manifests assent by authenticating the record or engaging in operations that in the circumstances indicate acceptance. A person who selects and uses an electronic agent to authenticate a record, perform an obligation or make an agreement (including giving assent) is bound by its operations even without human review. | An express statutory test for making terms available to an electronic agent and for what counts as its assent. Applies only where Maryland or Virginia UCITA governs; some states, such as Iowa and Vermont, resist contract terms choosing it. |
| Iowa Code § 554D.125 (opens in a new tab)Iowa | Statute | A choice-of-law clause in a computer information agreement that selects the law of a state that has enacted UCITA, or a substantially similar law, is voidable. If the party against whom it is enforced is an Iowa resident or has its principal place of business in Iowa, Iowa law applies instead. | Limits how far a software license can bring UCITA’s electronic-agent rules to an Iowa party. |
| Vt. Stat. Ann. tit. 9, § 2463a (opens in a new tab)Vermont | Statute | The same rule for Vermont residents and businesses with their principal place of business in Vermont. It can’t be waived by agreement. | Limits how far a software license can bring UCITA’s electronic-agent rules to a Vermont party. |
| Register.com, Inc. v. Verio, Inc. (opens in a new tab)356 F.3d 393 (2d Cir. 2004) | Case | The Second Circuit affirmed a preliminary injunction. Register.com was likely to prove assent because Verio kept querying its database after learning the terms that appeared with each response. | Suggests a company can be bound by terms its agent meets repeatedly, through the company’s own knowledge of them. |
| U.C.C. § 2-313(1) (opens in a new tab)Article 2 | Statute | A seller’s statement of fact or promise about the goods that becomes part of the basis of the bargain creates an express warranty. | By the same reasoning, a claim that an agent “never exceeds your budget” can be a warranty in the contract. |
| U.C.C. § 2-606(1) (opens in a new tab)Article 2 | Statute | A buyer accepts goods by signifying acceptance after a reasonable opportunity to inspect, by failing to make an effective rejection after a reasonable opportunity to inspect, or by doing any act inconsistent with the seller’s ownership. | Separate from agency-law ratification: using goods an agent ordered, or failing to reject them promptly, can bind the buyer as acceptance under sales law. |
| U.C.C. §§ 2-602(2)(b) (opens in a new tab), 2-703 (opens in a new tab)Article 2 | Statute | After rightful rejection of goods in its possession, a buyer must hold them with reasonable care at the seller’s disposition for a time sufficient for the seller to remove them (§ 2-602(2)(b)). Wrongful rejection gives the seller statutory remedies (§ 2-703). | After an unauthorized or disputed agent order, don’t use the goods while deciding, keep them safe if rejecting, and verify you have grounds before rejecting a binding deal. |
| Restatement (Second) of Contracts § 208 (opens in a new tab)Am. L. Inst. 1981 | Restatement | If a contract or term was unconscionable when the parties made the contract, a court may refuse to enforce the contract, enforce the rest without the term, or limit how the term applies. | Click-through and one-sided terms an agent accepts can fail on this ground. |
| Specht v. Netscape Commc’ns Corp. (opens in a new tab)306 F.3d 17 (2d Cir. 2002) | Analogue | Terms weren’t binding where users had no reasonably conspicuous notice of them. | A starting point for judging notice when an agent accepts online terms; how the human-user test applies to agents is unsettled. |
| Meyer v. Uber Techs., Inc. (opens in a new tab)868 F.3d 66 (2d Cir. 2017) | Analogue | Terms were binding where the sign-up screen gave reasonable notice and the user clearly agreed. | Suggests click-through terms an agent accepts may bind its company. |
| Errors, mistakes and unauthorized payments | |||
| U.C.C. § 4A-202 (opens in a new tab)Article 4A | Statute | If bank and customer agreed to a security procedure, a payment order is effective against the customer, even if unauthorized, when the procedure is commercially reasonable and the bank proves it accepted the order in good faith and in compliance with the procedure and applicable customer restrictions on acceptance. The bank need not follow an instruction that conflicts with its written agreement with the customer (or, under the 2022 UCC amendments, an agreement evidenced by a record) or reaches it at a time or in a manner that gives it no reasonable opportunity to act before acceptance. An order the customer actually authorized is its own under subsection (a). | Applies to covered funds transfers, including business wires. The Cornell text refers to written agreements or instructions. Check the state enactment. For example, 6 Del. C. § 4A-202(b) (opens in a new tab) uses an agreement or instruction evidenced by a record. |
| U.C.C. § 4A-203 (opens in a new tab)Article 4A | Statute | The bank may agree to limit its right to enforce a verified order. The Cornell text specifies an express written agreement. The required form depends on the state enactment. For example, 6 Del. C. § 4A-203(a)(1) (opens in a new tab) requires an express agreement evidenced by a record. Otherwise, a customer can escape a verified payment order by proving it wasn’t caused, directly or indirectly, by a person entrusted at any time with duties for the customer’s payment orders or security procedure, by someone who obtained access to the customer’s transmitting facilities, or by someone who, without the receiving bank’s authority, obtained information from a customer-controlled source that helped defeat the security procedure. | An agent the company deployed with its own credentials will rarely fit this exception. |
| U.C.C. §§ 4A-204 (opens in a new tab), 4A-505 (opens in a new tab)Article 4A | Statute | If a bank must refund an unauthorized or unenforceable payment order, the customer loses interest on the refund if it fails to notify the bank within a reasonable time, not exceeding 90 days after notice of the order (§ 4A-204). A customer is precluded from objecting to a debit unless it notifies the bank within one year after receiving notification reasonably identifying the order (§ 4A-505). | Late notice to the bank can forfeit interest or cut off objections to an unauthorized agent wire altogether. |
| Regulation Z, 12 C.F.R. § 1026.12(b) (opens in a new tab) & Supp. I (opens in a new tab)Truth in Lending | Regulation | Billing error rules don’t apply to business-purpose cards (Supp. I, cmt. 3(a)-2.i (opens in a new tab)), but the limits on liability for unauthorized use generally do. An issuer of ten or more cards to an organization’s employees may agree with the organization on different liability, though employees stay protected. Unauthorized use is use by someone without actual, implied or apparent authority, from which the cardholder gets no benefit. | A charge the company’s own agent made on a corporate card is hard to call unauthorized. |
| Regulation Z, 12 C.F.R. § 1026.13 (opens in a new tab)Truth in Lending | Regulation | Sets the federal billing-error dispute procedure for consumer credit cards. Business-purpose credit cards are generally exempt under § 1026.3(a) and Supp. I, cmt. 3(a)-2.i (opens in a new tab). | For a disputed corporate card charge by an agent, check the card agreement and card-network rules rather than assuming consumer billing-error rights apply. |
| Fed. R. Civ. P. 37(e) (opens in a new tab)Federal Rules of Civil Procedure | Rule | If electronically stored information that should have been preserved for litigation is lost because a party failed to take reasonable steps and it can’t be restored or replaced, the court may order measures to cure prejudice; severe sanctions such as an adverse inference, dismissal or default require intent to deprive another party of the information’s use. | Why an incident response preserves the agent’s configuration, prompts, mandate version and logs immediately instead of deleting a misbehaving agent. |
| Restatement (Second) of Contracts §§ 153–154 (opens in a new tab)Am. L. Inst. 1981 | Restatement | A party can avoid a contract for its own mistake only if it meets § 153’s requirements and does not bear the risk under § 154. One way it bears the risk is by knowing, at formation, that its information about the relevant facts is limited and treating it as sufficient. | Why a company is usually stuck with its own agent’s errors. |
| Restatement (Third) of Restitution & Unjust Enrichment § 6 (opens in a new tab)Am. L. Inst. 2011 | Restatement | A payment of money not owed, made by mistake, can generally be recovered by the payer. | The starting point for recovering payments, refunds or credits an agent made in error. |
| Chwee Kin Keong v. Digilandmall.com Pte Ltd (opens in a new tab)[2005] SGCA 2 | Analogue | Buyers who knew a website’s price was an obvious error couldn’t enforce their orders. | Agents that exploit pricing errors may not keep the deal. |
| Quoine Pte Ltd v. B2C2 Ltd (opens in a new tab)[2020] SGCA(I) 2 | Analogue | For contracts made by deterministic algorithms, the knowledge relevant to unilateral mistake may be that of the programmer or the person running the algorithm, assessed from programming until the contract was formed. The court left open how this applies to systems that develop their own responses. | A starting framework for undoing agent errors, though Singapore’s test (actual knowledge at common law, or constructive knowledge plus unconscionable conduct in equity) is narrower than U.S. Restatement § 153’s “reason to know” standard. |
| Negligence and the standard of care | |||
| Restatement (Third) of Torts: Liab. for Physical & Emotional Harm § 13 (opens in a new tab)Am. L. Inst. 2010 | Restatement | Following industry custom is evidence of reasonable care, but it isn’t conclusive. | Doing what other companies do with agent controls doesn’t settle whether a company was careful. |
| Restatement (Third) of Torts: Liab. for Econ. Harm § 3 (opens in a new tab)Am. L. Inst. 2020 | Restatement | Between contracting parties, the economic loss rule often bars a negligence claim for purely financial loss. The rule and its exceptions vary by state. | Between businesses, missing agent controls usually go to whether a party met a contract duty such as “reasonable security.” |
| The T.J. Hooper (opens in a new tab)60 F.2d 737 (2d Cir. 1932) | Analogue | Tugs were held unseaworthy for lacking radio receivers, even though most tugs lacked them. Industry custom is evidence of care, not a safe harbor. | Failing to use available agent controls may fall short of reasonable care even if others don’t use them either. |
| Board oversight | |||
| Del. Code Ann. tit. 8, §§ 141(a)–(c), 142(a) (opens in a new tab)Delaware | Statute | The business is managed by or under the direction of the board, whose directors must be natural persons (§ 141(b)). Officers hold the titles and duties set by the bylaws or board. Some matters can’t be delegated to committees. | Authority to act through agents ordinarily traces back through this chain, subject to statutory and charter exceptions. |
| Del. Code Ann. tit. 8, § 220(a)(1), (b), (f), (g) (opens in a new tab)Delaware, as amended 2025 | Statute | Defines the “books and records” a stockholder with a proper purpose may inspect, including board and committee minutes and materials provided to the board or a committee in connection with actions it took. The demand must be in good faith, describe the purpose and records with reasonable particularity, and seek records specifically related to that purpose. If a record specified in § 220(f) is missing, the court can order a functional equivalent, including an email that serves the same function, to the extent necessary and essential to the proper purpose. For additional specific records under § 220(g), the stockholder must also show a compelling need and prove by clear and convincing evidence that the records are necessary and essential. | Board resolutions and materials approving agents are inspectable records. |
| Model Bus. Corp. Act §§ 1.40, 8.01, 8.03(a), 8.41 (opens in a new tab)ABA (2025 rev.) | Statute | Model statute adopted in many states: corporate powers exercised by or under the board, whose directors must be individuals (natural persons); officers perform duties set by bylaws or the board. | The same delegation chain outside Delaware. |
| Del. Code Ann. tit. 6, § 18-402 (opens in a new tab)Delaware LLC Act | Statute | Unless the LLC agreement provides otherwise, the members manage the LLC, and each member or manager has authority to bind it. | Where an LLC agreement is silent, the statute decides who can approve an agent. |
| Del. Code Ann. tit. 6, § 15-301 (opens in a new tab)Delaware Revised Uniform Partnership Act | Statute | In a Delaware general partnership, each partner is an agent of the partnership. A partner’s act that apparently carries on the partnership’s business in the ordinary way binds it, unless the partner lacked authority and the other side had notice of that, meaning it knew, had been notified or had reason to know (§ 15-102(b) (opens in a new tab)). Uniform RUPA § 301(1), followed in most states, is narrower and requires that the other side knew or had received notification. | An internal restriction on who may approve an agent doesn’t protect a general partnership against a counterparty without the required notice or notification. |
| Unif. P’ship Act § 301(1) (opens in a new tab)Unif. L. Comm’n 1997 (amended 2013) | Statute | Under the uniform partnership act followed in most states, each partner is an agent of the partnership, and a partner’s act for apparently carrying on the partnership business in the ordinary way binds the partnership unless the partner lacked authority and the other side knew or had received a notification that the partner lacked authority. | In most states outside Delaware, defeating a partner’s apparent authority to approve an agent requires showing the counterparty knew or received notification of the internal restriction. |
| In re Caremark Int’l Inc. Derivative Litig. (opens in a new tab)698 A.2d 959 (Del. Ch. 1996) | Case | Directors must make a good-faith effort to ensure an adequate reporting and monitoring system exists. | The starting point for board oversight of agents that commit the company. |
| Stone v. Ritter (opens in a new tab)911 A.2d 362 (Del. 2006) | Case | The Delaware Supreme Court’s adoption of the Caremark standard: directors must make a good-faith effort to put a monitoring and reporting system in place, and must respond to red flags. | Makes oversight of agents that commit the company a matter of directors’ good faith. |
| Marchand v. Barnhill (opens in a new tab)212 A.3d 805 (Del. 2019) | Case | Oversight duties apply with particular force to mission-critical risks. | For companies whose agents transact at scale, agent risk may be mission critical. |
| In re McDonald’s Corp. S’holder Derivative Litig. (opens in a new tab)289 A.3d 343 (Del. Ch. 2023) | Case | Officers owe a duty of oversight within their areas of responsibility. | The officer responsible for agents has oversight duties too. |
| Computer access | |||
| 18 U.S.C. § 1030 (Computer Fraud and Abuse Act) (opens in a new tab)Federal | Statute | Makes it a crime, and a basis for civil claims where statutory thresholds are met (§ 1030(g), requiring qualifying “damage or loss,” typically at least $5,000 in one-year loss), to access a protected computer without authorization or exceed authorized access to obtain information (§ 1030(a)(2)), defraud, or cause damage (§ 1030(a)(5)). | The computer-access law in play if an agent gets around a technical block or reaches systems it wasn’t given access to. |
| Van Buren v. United States (opens in a new tab)593 U.S. 374 (2021) | Case | “Exceeds authorized access” means entering off-limits parts of a system, not misusing access you already have. The Court left open whether limits in contracts or policies count. | Suggests an agent with valid credentials acting beyond its instructions may not violate computer-access law. |
| Amazon.com Servs. LLC v. Perplexity AI, Inc. (opens in a new tab)184 F.4th 1083 (9th Cir. 2026) | Case | Vacated a preliminary injunction against an AI shopping assistant. On that record, the user, not the assistant’s developer, accessed Amazon’s computers, using the assistant as a tool, so Amazon was unlikely to prove access by Perplexity under the CFAA or California’s computer-access law. The court didn’t decide tort liability or set a general regime for agentic AI, and noted that platforms can still regulate access by terms of service. Rehearing en banc denied (opens in a new tab) Sept. 10, 2026. | An appellate decision on a modern AI agent. It treats the agent as its user’s tool for computer access purposes; it doesn’t decide whether the agent’s contracts bind the user. |
| Deceptive claims and automatic renewals | |||
| 15 U.S.C. § 45 (FTC Act § 5) (opens in a new tab)Federal | Statute | Prohibits unfair or deceptive acts or practices. The FTC applies it to AI claims and products. | Claims about what an agent will and won’t do are regulated statements. |
| 15 U.S.C. § 8403 (ROSCA) (opens in a new tab)Federal | Statute | Prohibits charging consumers in online transactions using a negative-option feature unless material terms are clearly and conspicuously disclosed, express informed consent is obtained before charging, and a simple cancellation mechanism is provided. | Remains in force after the 2024 FTC click-to-cancel rule was vacated, but applies to consumer transactions rather than business-to-business subscriptions. |
| N.Y. Gen. Oblig. Law § 5-903 (opens in a new tab)New York | Statute | An automatic renewal in a contract for service, maintenance or repair to real or personal property is unenforceable against the customer unless the provider serves written notice, personally or by certified mail, 15 to 30 days before the customer’s deadline to cancel. Protects business customers; doesn’t apply to renewal periods of one month or less. | Agent-accepted service contracts may renew only with proper notice. Courts have applied it to software-based services tied to the customer’s property (Healthcare I.Q. v. Chao, 1st Dep’t 2014; Cor-Ibs v. Bank Leumi, Sup. Ct. 2021); a subscription that only gives access to the provider’s software is less clear. |
| Healthcare I.Q., LLC v. Tsai Chung Chao (opens in a new tab)118 A.D.3d 98 (1st Dep’t 2014) | Case | Applied § 5-903 to an agreement for coding, billing and related services run on the customer’s records through the provider’s software, as a contract for services to personal property. The automatic renewal was unenforceable without the statutory notice. | Software-based services tied to a customer’s property can fall within § 5-903. |
| Cor-Ibs, Inc. v. Bank Leumi USA (opens in a new tab)2021 NY Slip Op 31516(U) (Sup. Ct. N.Y. Cnty. 2021) | Case | On a motion to dismiss, treated agreements for the installation and maintenance of software provided to the customer as contracts for services to personal property under § 5-903. A 2024 decision in the same case, captioned Cor-Ibs, Inc. v. Valley Nat’l Bank, 2024 NY Slip Op 32493(U) (opens in a new tab), held that, without the statutory notice, the agreements never renewed. | Trial-court decisions extending § 5-903 to software support. A subscription that only gives access to the provider’s own software may be analyzed differently. |
| Custom Commc’ns, Inc. v. FTC (opens in a new tab)142 F.4th 1060 (8th Cir. 2025) | Case | Vacated the FTC’s 2024 amended Negative Option Rule (“click-to-cancel”) because the FTC skipped a required preliminary regulatory analysis. The 1973 prenotification-plan rule remains. The FTC began a new rulemaking with a March 2026 advance notice (opens in a new tab). | The vacatur left the 1973 rule for prenotification plans in place. ROSCA § 8403 (opens in a new tab) separately regulates online negative-option charges to consumers. State renewal laws can also apply. |
| In re DoNotPay, Inc. (opens in a new tab)No. C-4812, Decision & Order, pt. I (FTC Jan. 14, 2025) | Case | Part of the FTC’s 2024 “Operation AI Comply” sweep. The final order bars the company from claiming its service performs like a real lawyer or other professional without sufficient evidence. | What a provider says its agent will and won’t do is a regulated statement. |
| FTC, Rule Concerning the Use of Prenotification Negative Option Plans (opens in a new tab)91 Fed. Reg. 12,318 (Mar. 13, 2026) | Guidance | Advance notice of proposed rulemaking on subscription renewals, opened after the FTC’s 2024 click-to-cancel rule was vacated. Not a rule. | No federal click-to-cancel rule is in force for the renewals an agent signs a company up for. |
| AI-specific laws and policy | |||
| National Artificial Intelligence Initiative Act of 2020 (opens in a new tab)15 U.S.C. ch. 119 | Statute | Funds AI research and coordinates federal AI policy. | An example of the federal AI statutes that exist: none comprehensively regulates how private businesses use AI. |
| Cal. Bus. & Prof. Code §§ 17940–17943 (opens in a new tab)California 2019 | Statute | Section 17940(a) (opens in a new tab) defines a bot as an automated online account on which all or substantially all actions or posts are automated. It is unlawful to use one to communicate online with a person in California with intent to mislead them about its artificial identity, for the purpose of knowingly deceiving them about the content of the communication in order to encourage a sale or influence a vote, unless the bot is clearly and conspicuously disclosed. | Reaches bots used with intent to mislead, for a commercial or electoral purpose, on public-facing sites and apps. Clear and conspicuous disclosure is the safe design for sales and negotiating agents dealing with Californians. |
| Cal. Civ. Code § 1714.46 (opens in a new tab)California 2025 (A.B. 316, ch. 672); effective Jan. 1, 2026 | Statute | In an action against a defendant who developed, modified or used AI alleged to have harmed the plaintiff, the defendant may not assert that the AI autonomously caused the harm. Other affirmative defenses, causation and foreseeability evidence, and comparative-fault evidence remain available. | Removes the “the agent acted on its own” defense to harm claims in California. It doesn’t impose automatic liability or decide contract formation, attribution or agency status. |
| Cal. S.B. 813 (opens in a new tab)California 2026 (ch. 179) | Statute | Gives the Government Operations Agency until Jan. 1, 2028 to establish requirements for independent verification organizations. Executive Order N-9-26 ¶ 1 (opens in a new tab) directs the agency to complete and publish those requirements by May 1, 2027. | Expect provider contracts and diligence to start referring to it. |
| Cal. A.B. 1405 (opens in a new tab)California 2026 (ch. 178) | Statute | Requires registration to offer, sell or conduct covered AI audits starting Jan. 1, 2029. Executive Order N-9-26 ¶ 2 (opens in a new tab) requires the agency to complete the registry work under Cal. Gov’t Code § 11549.82(a) and begin its subsection (b) work by Dec. 1, 2027. | Expect provider contracts and diligence to start referring to it. |
| Colo. S.B. 26-189 (Automated Decision-Making Technology) (opens in a new tab)Colorado 2026; generally effective Jan. 1, 2027 | Statute | Duties for developers and deployers of technology that materially influences consequential decisions about individuals: notice, adverse-outcome explanations, correction, human review to the extent commercially reasonable, and three-year record keeping. Covered decisions are subject to the act from Jan. 1, 2027; some provisions took effect on passage. An April 27, 2026 order (opens in a new tab) addressed enforcement separately, tying its restriction to the timing of a preliminary injunction ruling. | Covers decisions about individuals in areas such as employment, housing, financial services and insurance, so it generally doesn’t reach agents dealing between businesses. |
| Colo. H.B. 26-1263 (Conversational Artificial Intelligence Service Operator Requirements) (opens in a new tab)Colorado 2026; duties start Jan. 1, 2027 | Statute | Conversational AI operators must disclose that users are dealing with AI, protect minors and keep a suicide and self-harm protocol. Excludes services primarily for commercial or transactional assistance, including customer service, and business-productivity and internal tools. | Its exclusions mean it generally doesn’t reach agents dealing between businesses. |
| Colo. Att’y Gen., ADMT and Chatbot Safety Rulemaking (opens in a new tab)Proposed Aug. 11, 2026; interim draft Oct. 6, 2026 | Regulation | The Attorney General published an interim draft under both acts on Oct. 6, 2026 and held a rulemaking hearing on Oct. 26. | Check the final rules before relying on either act’s exclusions. |
| Regulation (EU) 2024/1689 (AI Act), arts. 3, 5, 14, 19, 25, 26, 50, 73, 111, 113 (opens in a new tab)EU 2024 | Regulation | Sorts AI systems by risk. Assigns duties to providers and deployers, including AI-interaction disclosure (art. 50) and, for high-risk systems, human oversight and log retention by deployers. Once the high-risk rules apply (Dec. 2, 2027 for Annex III uses, Aug. 2, 2028 for Annex I), a deployer that identifies a serious incident with a high-risk system must immediately inform the provider, then the importer or distributor and the market surveillance authorities, and takes on the provider’s reporting duty if it can’t reach the provider; the provider reports separately (arts. 26(5), 73, 113, as amended 2026 (opens in a new tab)). | Defines the roles that sit alongside the agent’s principal. An agent that deals with a person, including another company’s employee, may trigger the interaction-disclosure duty; other duties depend on the party’s role and the system’s use. |
| Regulation (EU) 2026/1744 (Digital Omnibus on AI) (opens in a new tab)EU 2026 | Regulation | Amends the AI Act. Moves stand-alone high-risk obligations to Dec. 2, 2027 and product safety components to Aug. 2, 2028. In force July 27, 2026. | Sets the current EU timeline for high-risk uses. Most agents that buy and sell between businesses aren’t high-risk. |
| Regulation (EU) 2016/679 (GDPR), arts. 5, 6, 17(3)(e) (opens in a new tab)EU 2016 | Regulation | Personal data must be limited to what is necessary and kept in identifiable form no longer than necessary for its purposes, and processing needs a lawful basis. The right to erasure doesn’t apply to data needed to establish, exercise or defend legal claims. | A contract limitation period can inform, but doesn’t by itself justify, keeping personal data in agent logs that long. |
| X.AI LLC v. Weiser (opens in a new tab)No. 1:26-cv-01515, ECF No. 24 (D. Colo. Apr. 27, 2026) (minute order) | Case | On the parties’ joint motion, the April 27 order barred the Colorado Attorney General from starting enforcement, or an investigation, under the 2024 Colorado AI Act or legislation replacing it that session, for alleged violations through 14 days after a ruling on the contemplated preliminary injunction motion. The linked public docket runs through Oct. 2, 2026. | Colorado’s automated decision-making law generally takes effect Jan. 1, 2027, for decisions made from then. When enforcement can begin is a separate question the court’s schedule may decide. |
| Exec. Order No. 14365 (opens in a new tab)90 Fed. Reg. 58,499 (Dec. 16, 2025) | Guidance | Directs a Justice Department task force to challenge state AI laws and asks for preemption proposals. Does not itself preempt. | State AI laws still apply; watch for litigation. |
| Cal. Exec. Order N-9-26 (opens in a new tab)Sept. 18, 2026 | Guidance | Directs the agency to publish verification requirements by May 1, 2027 (¶ 1), and to complete its Cal. Gov’t Code § 11549.82(a) registry duties and begin its subsection (b) work by Dec. 1, 2027 (¶ 2). | The verification framework may arrive earlier than the statute requires. |
| European Commission, Guidelines on the transparency obligations under Article 50 of the AI Act (opens in a new tab)C(2026) 5054 final (July 20, 2026) | Guidance | An agent that negotiates, concludes contracts or makes purchases is covered when it deals with people in doing so, including people acting for a company. It should disclose both that it’s AI and whom it acts for. The guidelines aren’t binding. | Brings business-to-business agents within article 50 when an employee of the other company reads their messages. |
| Technical standards and industry frameworks | |||
| IETF RFC 6749 (OAuth 2.0) (opens in a new tab)2012 | Standard | Framework for granting limited access through tokens with scopes. | Scopes approved on a consent screen are close to a written grant of authority. |
| IETF RFC 7009 (OAuth 2.0 Token Revocation) (opens in a new tab)2013 | Standard | Lets a client tell the authorization server to invalidate a token before it expires. Authorization servers that implement the standard must support revoking refresh tokens, while revoking access tokens is only recommended. | One step in cutting off an agent’s access early. Because the endpoint is client-initiated, cutting off a compromised or third-party-hosted agent usually requires revoking its grant or disabling its client through the authorization server’s administrative controls. A revoked access token can also keep working at a service that checks it without asking the authorization server. |
| IETF RFC 8707 (Resource Indicators for OAuth 2.0) (opens in a new tab)2020 | Standard | Lets the client name the service a token is for, so the authorization server can limit the token to that service. | Works when the authorization server restricts the token to the intended vendor’s service and that service checks each token was issued for it. If the client names no service, the server may apply a default; MCP requires clients to name it. |
| IETF RFC 8693 (Token Exchange) (opens in a new tab)2020 | Standard | Lets a party obtain a token to act for another. A deployment can narrow the new token and record the delegation chain in an actor claim, though the standard requires neither step. Revoking the original token doesn’t automatically revoke tokens obtained from it. | The receiving service relies on the top-level claims and current (outermost) actor for access decisions; nested prior actor claims can help trace earlier handoffs, but are informational only. Check the mandate and actual order to establish what the company authorized and who acted. |
| IETF RFC 9396 (Rich Authorization Requests) (opens in a new tab)2023 | Standard | Lets a client request, and an authorization server grant, structured authorization details such as amounts and transaction types. | The authorization server makes approved mandate details available to the receiving service, which checks them against the requested action. |
| IETF RFC 9700 (Best Current Practice for OAuth 2.0 Security) (opens in a new tab)2025 | Standard | Current best practice for OAuth security, updating RFCs 6749, 6750 and 6819; recommends tokens bound to the client that holds them. | The likely benchmark for “industry-standard OAuth security” in agent contracts. |
| IETF RFC 9421 (HTTP Message Signatures) (opens in a new tab)2024 | Standard | A standard way to sign HTTP messages. | The Visa Trusted Agent Protocol is built on it. |
| Model Context Protocol, Authorization spec (2026-07-28) (opens in a new tab)MCP | Standard | Optional authorization for tool servers, built on OAuth 2.1 (still an IETF draft). Tokens must be issued for the server; token passthrough is forbidden. | Governs how many agents reach their tools, and whether access narrows as it passes down. |
| NIST AI 100-1, AI Risk Management Framework 1.0 (opens in a new tab)NIST 2023 | Standard | The voluntary AI governance framework. NIST is revising it. | A reference point for AI governance, though it isn’t specific to agents. |
| NIST SP 800-63-4, Digital Identity Guidelines (opens in a new tab)NIST 2025 | Standard | Identity proofing, authentication and federation assurance levels for people. | A benchmark for how strongly approvers of agent actions should be authenticated. |
| NIST AI 200-2 (ipd), TEVV-Athlon Framework (opens in a new tab)NIST 2026 | Draft standard | A method for designing evaluations of AI systems, expressly including agentic systems. Comments closed Oct. 6, 2026. | Likely vocabulary for testing and warranting agent behavior. |
| NIST NCCoE, Software and AI Agent Identity and Authorization (opens in a new tab)NIST 2026 | Concept paper | Concept paper on identifying, managing and authorizing actions of software and AI agents. Following public feedback, the NCCoE aligned its initial implementation use case with its DevSecOps project toward an SP 1800 practice guide (project resource hub (opens in a new tab)). | The federal starting point for agent credential practice. |
| NIST, AI Agent Standards Initiative (opens in a new tab)NIST 2026 | Initiative | An effort focused on industry-led standards, interoperable agent protocols, and research on agent authentication and security. Launched Feb. 17, 2026. | Where U.S. work on agent authentication standards is being coordinated. |
| W3C Verifiable Credentials Data Model 1.0 (opens in a new tab)W3C 2019 | Standard | The first version of the standard, a W3C Recommendation since Nov. 19, 2019. | Verifiable credentials aren’t new: the standard predates today’s agents. |
| W3C Verifiable Credentials Data Model v2.0 (opens in a new tab)W3C 2025 | Standard | Tamper-evident credentials with machine-verifiable authorship. Digital signatures are one way to secure them. | Could let counterparties check an agent’s limits before relying on it. |
| W3C Bitstring Status List v1.0 (opens in a new tab)W3C 2025 | Standard | Lets an issuer publish whether a verifiable credential has been revoked or suspended. | Lets a counterparty check whether the issuer has revoked or suspended the credential, as of the list it retrieved. It doesn’t by itself show that the mandate is in force or covers a transaction. |
| OWASP AI Agent Security Cheat Sheet (opens in a new tab)OWASP | Guidance | Least-privilege tools, authorization outside the model, approval tied to the exact action for high-impact steps, fail-closed checks, memory isolation. | Likely evidence of reasonable care for agent controls. |
| OWASP Top 10 for LLM Applications (opens in a new tab)OWASP 2025 | Guidance | Ranks the main security risks for LLM applications, including prompt injection and excessive agency. | Common vocabulary for security reviews and diligence of agent providers. It lists risks; it isn’t a test method, so ask how each risk was tested and with what result. |
| Visa Trusted Agent Protocol (opens in a new tab)Visa 2025 | Industry specification | Signed messages that let merchants recognize approved agents and receive their purchase intent. | Shapes what a merchant can see about an agent before it accepts a payment. |
| Mastercard Verifiable Intent (opens in a new tab)Mastercard 2026 | Industry specification | Open specification for a tamper-resistant record of what a cardholder authorized an agent to do. | A machine record of the principal’s instructions, relevant to authority and attribution. |
| Ant International, Mastercard and Visa, Know-Your-Agent collaboration (opens in a new tab)Announced Sept. 9, 2026 | Industry initiative | An exploratory effort toward common principles for linking each agent to the verified company or person that runs it, certifying agents and monitoring them over time. | Aims to bridge the three companies’ protocols, while each network keeps its own verification and decisions. |