Summary
- AI regulation doesn’t decide whether an agent’s act binds the company. It adds duties on top: telling people they’re dealing with AI, documentation, logging and accurate claims.
- Complying with a regulation is not the same as having authority. An agent can be fully compliant and still act outside what the company approved, and the reverse.
- In the EU, agents that buy and sell between businesses generally aren’t high-risk. The duty to tell people they’re interacting with AI has applied since Aug. 2, 2026, and it can reach an agent that deals with another company’s employees.
- No federal statute comprehensively regulates private sector AI. The FTC applies its deception powers to what providers claim their agents do. California’s bot law covers automated accounts on which all or substantially all actions or posts are automated. It bars using such an account on a public-facing website or app to mislead people about its artificial identity in order to deceive them into a sale or influence a vote, unless it clearly and conspicuously discloses that it’s a bot.
Five layers
When an agent does something with legal consequences, work outward from the transaction. Most real problems cut across several layers at once, so start at the first one, not the third.
- 1. Authority and contractIs the company bound? Was the agent authorized? Did the other side reasonably rely?The rest of this site
- 2. General lawPrivacy, competition, export controls, payments, fraudApplies with or without AI
- 3. AI-specific statutesEU AI Act, bot disclosure lawsThis guide
- 4. Regulators’ guidance and enforcementFTC orders, supervisory expectationsThis guide
- 5. Technical standardsIdentity, authorization, logging and evaluation standardsGuide 10
Regulation sits in layers 2 through 4. It rarely decides whether the company is bound, but it often shapes the facts that decide it: a required AI disclosure, human oversight step or log can affect what the other side knew, who approved the action and what the company can prove. See Where regulation and authority meet.
Provider, deployer, principal
AI laws assign duties by role. Agent disputes add a third role the statutes don’t name: the person or company the agent acts for. Identify all three every time, because one company can hold more than one.
| Role | Who it usually is | What the law asks of it |
|---|---|---|
| ProviderEU AI Act | Whoever develops the AI system or model, or has it developed, and places it on the market or puts it into service under its own name. The model’s provider and the provider of an agent built on it can be different companies | Design duties: for systems that interact with people, telling them it’s AI (art. 50); for high-risk systems, logging capability, instructions for use, design for human oversight and conformity assessment |
| DeployerEU AI Act | Whoever uses the AI system under its authority in its business | Use duties: for high-risk systems, follow the instructions for use, assign human oversight, keep logs, report serious incidents and give notices to affected people. Other deployer duties, such as disclosing deepfakes, apply more widely (art. 50) |
| PrincipalAgency and contract law | Whoever the agent acts for in the transaction | Bound by the deal, and bears the loss unless a contract moves it, if agency or attribution rules make the deal its own. Being named as the company the agent acts for doesn’t settle that by itself |
Usually the deployer is also the principal, but not always. A company can deploy an agent that acts for its business customers, and then each customer is the principal for its own transactions. Who the deployer is turns on who uses the system under its authority: often the company running the agent, but a customer that directs the agent in its own business may be a deployer too.1 A deployer can also become a provider, for example, by putting its own name on a high-risk system or substantially modifying it.2
The EU AI Act
For business-to-business agents, the Act’s structure matters more than its individual articles. The Act sorts AI systems by risk, and each tier has its own start date. The Digital Omnibus on AI, Regulation (EU) 2026/1744, in force since July 27, 2026, pushed back the high-risk dates.3
| Obligation | Applies from | Relevance to agents |
|---|---|---|
| Prohibited practices (art. 5) | Feb. 2, 2025; additions from Dec. 2, 2026 | Rarely at issue for business-to-business agents. Manipulative designs are the main risk. |
| General-purpose AI models | Aug. 2, 2025; models on the market before that date: Aug. 2, 2027 | Duties fall on the model provider, not the company deploying an agent built on it. |
| Transparency: direct interaction (art. 50(1)) | Aug. 2, 2026 | An agent that interacts with people, including another company’s employees, must be designed so they’re told it’s AI, unless that’s obvious, and, under the Commission’s guidelines, whom it acts for. The separate duty to mark AI-generated content (art. 50(2)) applies from Dec. 2, 2026 to systems already on the market before Aug. 2, 2026. |
| High-risk uses (Annexes I and III) | Annex III (stand-alone uses): Dec. 2, 2027; Annex I (regulated products): Aug. 2, 2028 | Only for listed uses, such as hiring and safety components of regulated products. Ordinary buying and selling between businesses is generally outside the list, but an agent that selects individual workers or contractors can raise Annex III employment issues. |
Article 50 is the duty most likely to reach a business-to-business agent. It protects natural persons, and an employee reading an agent’s messages on behalf of a vendor or customer is one. The Commission’s guidelines say 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, and that it should disclose both that it’s AI and whom it acts for. The guidelines aren’t binding; only the Court of Justice can settle what article 50 means.4 For a case the guidelines leave open, see Open questions.
United States
Federal: old law applied to new systems
Federal AI statutes exist, such as the National Artificial Intelligence Initiative Act of 2020,5 but they fund research and coordinate policy; none comprehensively regulates how private businesses use AI. The FTC applies Section 5 of the FTC Act, which prohibits unfair or deceptive practices, to AI products and claims.6 Its 2024 “Operation AI Comply” sweep included a proceeding against DoNotPay, which had marketed a “robot lawyer.” The final order bars the company from claiming its service performs like a real lawyer or other professional without sufficient evidence.7
For agents, the lesson is direct. What a provider says its agent will and won’t do is a regulated statement. A claim that an agent “never exceeds your budget” can be a warranty in the contract and a deception problem at the FTC.8
Executive Order 14365 (Dec. 2025) directs a Justice Department task force to challenge state AI laws and asks for preemption proposals.9 It doesn’t itself preempt anything. State AI laws remain in force unless a court or Congress says otherwise.
States
California’s bot disclosure law covers an automated online account when all or substantially all of its actions or posts occur without a person taking those actions or making those posts. A person using AI to draft a reply does not, by itself, make the account a bot. It is unlawful to use a bot 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 to encourage a purchase or sale in a commercial transaction, or to influence a vote. A clear and conspicuous disclosure that it’s a bot avoids liability. The law isn’t limited to consumers, since “person” includes businesses, but it reaches only bots on public-facing websites and apps, so a negotiation by private email may fall outside it.10 Disclosure is still the safer design for any sales or negotiating agent.
California has also removed one defense. Since Jan. 1, 2026, a defendant sued for harm allegedly caused by AI it developed, modified or used may not argue that the AI autonomously caused the harm. Other affirmative defenses, evidence on causation or foreseeability, and evidence of anyone else’s comparative fault remain available.11 The statute takes away an excuse; it doesn’t make a developer or deployer automatically liable, and it doesn’t decide whether a contract was formed, whose act a transaction was, or whether an AI system is anyone’s agent.
In September 2026, California also enacted SB 813, which has the Government Operations Agency set up a framework for independent verification organizations that assess AI systems by Jan. 1, 2028, and AB 1405, which creates an AI Auditor Registry and, from Jan. 1, 2029, requires registration to offer, sell or conduct covered AI audits. Executive Order N-9-26 directs the agency to publish the verification requirements by May 1, 2027. By Dec. 1, 2027, the agency must complete its registry duties under Cal. Gov’t Code § 11549.82(a) and begin the work required by subsection (b). Expect provider contracts and diligence to start referring to both.12
Colorado’s new laws on automated decision-making and chatbots, whose duties generally start Jan. 1, 2027, generally don’t reach agents dealing between businesses. The first covers decisions about individuals in education, employment, housing, financial services, insurance, health care and government services; the second excludes tools for commercial transactions, customer service and business operations.13 An April 27, 2026 order in X.AI v. Weiser restricted the Attorney General’s enforcement of the automated-decision law and its predecessor. The order covered investigations and enforcement for alleged violations through 14 days after the court rules on the contemplated preliminary injunction motion. The public docket cited here runs through Oct. 2, 2026.14
Compliance isn’t authority
The most common mistake is treating a compliance review as an authority review. They answer different questions.
The highlighted cell is the one compliance programs miss.
Where regulation and authority meet
- Disclosure supports notice. A required disclosure that “you’re dealing with an AI agent” can also say which company the agent acts for and what the agent’s limits are. Notice of a limit can defeat apparent authority.
- Oversight doubles as approval. A person assigned to oversee an agent is often the person whose sign-off makes an action authorized.
- Logs serve two purposes. Disputes over agent transactions can arise years later, so keep the transaction evidence (mandate version, approvals, confirmations, actions and results) for the limitation period of the contracts involved, and longer if a legal hold applies. Keep sensitive inputs only as long as needed: where GDPR or similar law applies, a contract limitation period can inform, but doesn’t by itself justify, keeping personal data that long.15 Never log working credentials such as passwords, API keys or access tokens.
- Claims can become obligations. Marketing claims about an agent’s limits can attract FTC scrutiny. Claims in documentation can also create contractual obligations or warranties, depending on the agreement and applicable law.
Worked example
Alder’s procurement agent is negotiating a supply contract by email with Westbridge, a vendor with sales staff in Dublin and San Francisco. Westbridge uses its own AI sales agent to answer some messages. Westbridge’s agent offers a 15% volume discount that Westbridge never approved, and Alder’s agent accepts.
- Authority (layer 1). Is Westbridge bound by the discount? Under the tool view the offer is Westbridge’s own, as the airline’s statement was in Moffatt. Under agency rules, the question is whether Alder could reasonably believe Westbridge’s agent could offer it. Is Alder bound by its agent’s acceptance? The same analysis applies in reverse: see Guide 1.
- EU (layer 3). Neither agent is high-risk. Westbridge’s sales agent writes to Alder’s buyers, and Alder’s agent writes to Westbridge’s Dublin staff, so each agent’s provider must design it to disclose that it’s AI where article 50 applies. Under the Commission’s guidelines, an agent’s emails to a person count, and each agent should also say which company it acts for.
- California (layer 3). Westbridge’s agent is trying to make a sale, and Alder’s buyer may be in California. The bot disclosure law reaches bots on public-facing websites and apps, so this email negotiation may fall outside it. If Westbridge’s agent also talks to buyers through its website, the law is in play there.
- FTC (layer 4). If Westbridge’s provider marketed its sales agent as “never offering unapproved discounts,” that claim is now a contract question between them and a potential deception problem.
- The fix. Limit what each agent may offer or accept, which is an authority control. Add AI disclosure to both agents’ messages, which is a compliance control. Neither replaces the other.
What changes in contracts
When an agent’s use is regulated, the contracts for it should include additional terms:
- Roles. Say who is provider and who is deployer for each use.
- Documentation. The provider supplies what the deployer needs to meet its duties: intended uses, known limits, instructions for use.
- Logs. Keep transaction evidence for the longer of any regulatory minimum and the dispute window. Where personal data is involved, set shorter, separate rules for raw inputs and personal data, and use legal holds for disputes.
- Disclosures. Say which side builds and displays AI-interaction notices.
- Changes. Require notice of changes that could alter a system’s risk classification.
- Incidents. Say who tells whom, and how fast, when an agent malfunctions or is compromised, and what happens to transactions made in the meantime. Under the EU AI Act, once its high-risk rules apply (Dec. 2, 2027 for most stand-alone uses), 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. If it can’t reach the provider, the provider’s reporting duty applies to the deployer. The provider has its own deadline to report. A contract can add notices between the parties but can’t replace either party’s own.16
- Cooperation. Require each side to help the other answer a regulator’s inquiry about the agent.
- Records. Say who keeps which records, so neither side assumes the other has them.
Sample language for these terms is in clause 10 of Agent Use Clauses, except for records, which clause 7 covers.
Issue spotting sequence
Use these questions for the regulatory review in step 3 of An Authority Review, Step by Step. The full review also covers authority, controls and contracts.
- Function. What does the agent actually do?
- Consequence. Can it bind, pay, decide or make statements others rely on?
- Principal. Whom does it act for?
- Roles. Who is provider and who is deployer?
- Classification. Does an AI-specific law apply, and in which tier? For the EU, when was the first unit of the same type and model of the system lawfully placed on the EU market or put into service (not only your own go-live date), and has its design changed significantly since?
- General law. Which existing rules apply regardless of AI, such as privacy, export controls or payments rules?
- Authority. What is the agent approved to do, and do its credentials match? See Guide 1.
- Outward signals. What do counterparties see about the agent and its limits?
- Evidence. Will the logs prove both compliance and authority?
Both sides of the deal
If you deploy agents
- Run the authority review and the compliance review separately. Compliance doesn’t make an action authorized, and authority doesn’t make it compliant.
- Disclose AI interaction by default when your agent writes to people at other businesses.
If you build agents
- Ship disclosure and logging features, and document which regulatory duty each supports.
- Keep marketing claims about limits and accuracy to what the logs can show.
- Agree with each customer on the regulatory roles for each use case.
Open questions
- When an agent deals only with the counterparty’s own agent or ordering system, and an employee there sees the result later, does it “interact directly” with that employee under article 50? The Commission’s guidelines exclude mediated exposure and backend machine-to-machine calls, but bring agents in multi-agent setups into scope where another agent deals with people.
- Is an ordering portal or customer app that business customers reach only after signing in “public-facing” under California’s bot law, or does the law stop at the login page?
- Does disclosing that a counterparty is an AI agent change what a reasonable person may assume about its authority?
- Regulation (EU) 2024/1689 (Artificial Intelligence Act) (opens in a new tab), art. 3(3)–(4), OJ L, 12.7.2024. ↩
- Id. (opens in a new tab) art. 25(1). ↩
- Regulation (EU) 2026/1744 (Digital Omnibus on AI) (opens in a new tab), published July 24, 2026, in force July 27, 2026; AI Act arts. 5, 50, 111, 113 (opens in a new tab) as amended; recital 39 of Regulation (EU) 2026/1744 (the art. 111(2) transition attaches to the type and model of high-risk AI system, dated by its first unit placed on the EU market or put into service); see the EUR-Lex consolidated AI Act (July 27, 2026) (opens in a new tab). ↩
- European Commission, Guidelines on the implementation of the transparency obligations for certain AI systems under Article 50 of Regulation (EU) 2024/1689 (opens in a new tab), C(2026) 5054 final, Annex (July 20, 2026), paras. 5, 30(iii)–(iv), 31, 36. ↩
- National Artificial Intelligence Initiative Act of 2020, 15 U.S.C. ch. 119 (opens in a new tab). ↩
- 15 U.S.C. § 45(a) (opens in a new tab). ↩
- In re DoNotPay, Inc., No. C-4812, Decision & Order, pt. I (FTC Jan. 14, 2025) (opens in a new tab); FTC, Operation AI Comply (Sept. 25, 2024) (opens in a new tab). ↩
- See, e.g., U.C.C. § 2-313(1) (opens in a new tab) (express warranties); 15 U.S.C. § 45(a) (opens in a new tab). ↩
- Exec. Order No. 14365, Ensuring a National Policy Framework for Artificial Intelligence, 90 Fed. Reg. 58,499 (opens in a new tab) (Dec. 16, 2025). ↩
- Cal. Bus. & Prof. Code §§ 17940–17943 (opens in a new tab). ↩
- Cal. Civ. Code § 1714.46 (opens in a new tab) (added by Stats. 2025, ch. 672 (A.B. 316)). ↩
- Chaptered texts: S.B. 813 (opens in a new tab) (ch. 179) and A.B. 1405 (opens in a new tab) (ch. 178), both Stats. 2026; Office of the Governor, announcement of Sept. 9, 2026 (opens in a new tab); announcement of Sept. 30, 2026 (opens in a new tab); Cal. Exec. Order N-9-26 (opens in a new tab) (Sept. 18, 2026). ↩
- Colo. S.B. 26-189 (2026) (Automated Decision-Making Technology; signed May 14, 2026) (opens in a new tab); Colo. H.B. 26-1263 (2026) (Conversational Artificial Intelligence Service Operator Requirements; signed May 29, 2026) (opens in a new tab); Colo. Att’y Gen., ADMT and Chatbot Safety Rulemaking (opens in a new tab) (proposed rules Aug. 11, 2026; interim draft Oct. 6, 2026; hearing and comment deadline Oct. 26, 2026). The Attorney General calls the two laws the ADMT Act and the Chatbot Safety Act. ↩
- X.AI LLC v. Weiser (opens in a new tab), No. 1:26-cv-01515-DDD-CYC, ECF No. 24 (D. Colo. Apr. 27, 2026) (minute order) (covering S.B. 24-205 “or any legislation replacing or amending SB24-205 enacted during this legislative session”). ↩
- Regulation (EU) 2016/679 (GDPR) (opens in a new tab), arts. 5(1)(b)–(c), (e), 6, 17(3)(e); AI Act (opens in a new tab), art. 26(6) (deployers keep logs for at least six months “unless provided otherwise in applicable Union or national law, in particular in Union law on the protection of personal data”); see also art. 19(1) (the same qualification for providers). ↩
- Regulation (EU) 2024/1689 (Artificial Intelligence Act) (opens in a new tab), arts. 26(5), 73(1)–(4), 113, as amended by Regulation (EU) 2026/1744 (opens in a new tab). ↩