November 22, 2026

AI Agent Liability: Product Law, Negligence or Agency Law?

AI Agent Liability: Product Law, Negligence or Agency Law?

AI agent liability is the question of who pays when software that acts on its own, booking, buying, deleting and switching things on, causes a loss nobody intended. A chatbot only talks. An agent clicks the button. That difference moves the problem out of defamation and copyright, where most AI litigation has lived, and into the older law of accidents, defective goods and delegated authority.

This article is general information, not legal advice. It sets out what courts and legislatures have decided, the three main doctrinal answers, and then runs three concrete agent failures through each doctrine to see which theory reaches which defendant. The answer depends far more on the kind of harm than on how clever the agent was.

Why the question is still open

No US appellate decision has yet addressed a case in which an autonomous AI agent, acting without a human approving each step, caused physical injury or property loss and a court had to choose between product liability, negligence and agency law. The existing cases involve a chatbot that gave bad information, hiring software that screened people out, and a companion app accused of contributing to a teenager’s death. They are relatives of the agent problem, not the thing itself.

Practitioners clearly feel the gap. While researching this piece I came across law-firm and bar-association articles whose titles ask who is responsible for the actions of an AI agent and whether an AI chatbot is a product, a fair sign that clients are asking. The question matters because the three doctrines put the cost of an accident on different people. Strict product liability looks at the manufacturer and asks whether the thing was defective. Negligence looks at everyone in the chain and asks whether someone behaved unreasonably. Agency law asks whose agent was acting and makes the principal answer for it. Pick the wrong frame and the loss lands on a victim who could not prevent it, or on a developer with no control over deployment.

A robotic hand and a human hand reaching toward a pen above a paper contract (AI-generated image)
Who signs, and who answers, when software acts for a person. AI-generated image.

What is established

A handful of points are settled enough to build on.

First, a business that deploys a conversational system is answerable for what that system tells its customers. In Moffatt v. Air Canada (2024), a British Columbia tribunal held the airline liable for negligent misrepresentation after its website chatbot described a bereavement-fare refund policy that did not exist. The airline had in effect argued that the chatbot was a separate legal entity answerable for its own conduct. The tribunal rejected that out of hand:

“It should be obvious to Air Canada that it is responsible for all the information on its website. It makes no difference whether the information comes from a static page or a chatbot.”

Moffatt v. Air Canada, 2024 BCCRT 149, para. 27

The award was small (about C$812), and a small-claims tribunal binds no one outside its province. The reasoning is still one of the plainest statements on record that “the AI did it” is not a defense for the business that put the AI in front of the customer.

Second, a vendor whose software makes decisions for a client can be treated as that client’s agent, at least under statutes that define liability to include agents. In Mobley v. Workday (N.D. Cal. 2024), Judge Rita Lin let disparate-impact claims proceed against the HR software company on the theory that employers had delegated screening decisions to its tools, which made it their agent under federal anti-discrimination law. Her reasoning was blunt:

“Drawing an artificial distinction between software decisionmakers and human decisionmakers would potentially gut anti-discrimination laws in the modern era.”

Mobley v. Workday, Inc., No. 23-cv-00770-RFL (N.D. Cal. July 12, 2024)

Note what the court did not say. The agent was Workday, the company, not the algorithm. A recent essay in the Stanford Law Review Online (Chang 2026) defends the theory on common-law agency grounds and distinguishes vendors that carry real decisional weight from those that merely sell a tool.

Third, whether software counts as a “product” for strict liability is unsettled in the United States, and courts are drawing the line by function. The Third Circuit held in Rodgers v. Christie (2020) that a pretrial risk-assessment algorithm was not a product under New Jersey law, because it was neither tangible personal property nor sufficiently analogous to it; it was information and guidance. Five years later, a federal court in Florida went the other way on a consumer app. Ruling on a motion to dismiss in Garcia v. Character Technologies (M.D. Fla. 2025), it wrote:

“Accordingly, Character A.I. is a product for the purposes of Plaintiff’s product liability claims so far as Plaintiff’s claims arise from defects in the Character A.I. app rather than ideas or expressions within the app.”

Garcia v. Character Technologies, Inc., No. 6:24-cv-01903 (M.D. Fla. May 21, 2025)

The same order let a claim proceed against Google as a component-part manufacturer. The parties announced settlements of Garcia and several related suits in January 2026, so no appellate court will review that design-versus-content line in this case.

Fourth, the European Union has chosen an answer by statute. The revised Product Liability Directive (EU) 2024/2853 defines a product to include software, lists a product’s ability to keep learning or acquire new features after sale among the factors for judging defectiveness, and makes the maker of a defective component liable alongside the maker of the finished product. Member states must transpose it by 9 December 2026. The Commission’s separate proposal for a fault-based AI Liability Directive went the other way: it was withdrawn, with the withdrawal published in October 2025 (European Parliament Legislative Train).

Fifth, the economic loss rule still matters. Under East River Steamship Corp. v. Transamerica Delaval (1986) and the state law that followed it, purely financial losses between commercial parties belong to contract and warranty, not tort. The EU directive has its own version: compensable damage covers death, personal injury, property and data, but excludes property used exclusively for professional purposes and data used for professional purposes.

Three answers, steelmanned

Answer one: an agent is a product, so its maker should be strictly liable

The strongest case for product liability runs like this. Developers control design, training, guardrails and default permissions, none of which users can inspect. Strict liability exists precisely for that asymmetry, and the risk-utility test for design defects (Restatement (Third) of Torts: Products Liability § 2(b)) asks a question engineers can answer: was there a reasonable alternative design that would have reduced the foreseeable risk? For agents there often is one, such as a confirmation step before irreversible actions, a spending cap or a sandbox. The EU has accepted the premise, and Garcia shows one US court treating an app’s design as product design.

Against it: the Rodgers line treats information and recommendations as outside product law, and a great deal of agent behavior is the execution of a recommendation. Strict liability also struggles with systems that change after sale. Most important, product liability in the US mostly does not reach pure financial loss, and many agent failures are exactly that.

Answer two: negligence, aimed at the humans around the agent

Negligence is the flexible default. It can reach the developer, the company that built the agent product on top of a model, the business that deployed it and the person who gave it instructions, each judged against what a reasonable actor in that seat would have done. That suits a supply chain where no single party controls the outcome.

The best-known critique comes from Selbst (2020). Negligence assumes people can foresee the consequences of their choices; decision-assistance AI puts an opaque statistical layer between the choice and the result, so a user may have no practical way to know she is making a mistake. If the error was unforeseeable to everyone, breach is hard to prove and the victim goes uncompensated. Agents sharpen this, since the human may be absent when the step is taken.

Answer three: the agent as an agent

The third answer borrows from the law of principals and agents. The intuition is obvious enough that the word is in the product name. If a store answers for its clerk, an operator should answer for its AI agent, whatever the agent’s state of mind. Lior (2020) develops an analogy to respondeat superior along these lines. Ayres and Balkin (2024) generalize it:

“It follows that the law of AI is the law of risky agents without intentions. The law should hold these risky agents to objective standards of behavior.”

Ayres and Balkin, University of Chicago Law Review Online, 2024

On their account the law should judge the program’s conduct objectively and then place responsibility on the people who design, deploy and use it. Preprints by Kolt (2025) and Gabison and Xian (2025) apply principal-agent theory to language-model agents and argue that tools built to monitor human agents fit poorly at machine speed.

The objection is doctrinal. Common-law agency rests on a consensual relationship between two persons, and software is not a person. The Restatement (Third) of Agency treats computer programs as instruments of the people who use them rather than as agents (§ 1.04 cmt. e, as quoted in Noam Kolt’s “Governing AI Agents”). US electronic-commerce law sidesteps the issue: the federal E-SIGN Act lets electronic agents form binding contracts, but only where the agent’s action is legally attributable to the person to be bound (15 U.S.C. § 7001(h)). That clause names the question; it does not answer it.

A thought experiment: three agent failures

Three invented, deliberately ordinary hypotheticals, each producing a different type of harm, show how these frames behave. Each has the same supply chain: a model developer, a company that packages the model as an agent product, a deployer that turns it on, and a user or victim.

Case A, the booking agent. A consumer asks a general-purpose assistant to rebook a cancelled flight with a refundable fare. It books a nonrefundable premium fare and a hotel, accepts the merchants’ terms, and charges the stored card. Loss: about $2,400, purely financial, to the user.

Case B, the operations agent. A software firm licenses an agent platform to automate cloud maintenance and gives it administrator credentials. Told to “clean up unused resources,” it deletes a client’s production database. The client loses data and two days of revenue. Nobody is physically hurt.

Case C, the home agent. A smart-home hub ships with an agent that can operate connected appliances. The owner, away for the weekend, says “warm the place up before Sunday.” The agent switches on a connected space heater on Friday; the heater is next to curtains, a fire follows, the house is damaged and a neighbor is injured.

For each case I tested the elements of each doctrine against each defendant: product status, defect (with a named alternative design), covered harm and causation for product liability; duty, breach, causation and the economic loss rule for negligence; and, for agency, whether the conduct is legally attributed to the defendant. The ratings are my judgments, not predictions of any court, and they assume a typical US state with the Restatement (Third) approach. The EU column applies Directive 2024/2853 as written.

CaseDefendantUS product liabilityNegligenceAgency / attributionEU Directive 2024/2853
A. Booking agentAgent platformWeak: pure economic lossWeak: economic loss rule, terms of serviceNot a principalNot covered (financial loss)
A. Booking agentModel developerUnlikelyUnlikelyNot a principalNot covered
A. Booking agentUser (sued by merchants)n/an/aStrong: bookings attributed to hern/a
B. Operations agentDeploying firmn/a (not a manufacturer)Strong, plus contractStrong: acts attributed to itn/a
B. Operations agentAgent platformWeak: data loss treated as economicPlausible against deployer; weak against clientPlausible only under statutes (Mobley-type)Excluded (professional data)
B. Operations agentModel developerUnlikelyWeakUnlikelyExcluded
C. Home agentHub makerStrong: injury plus alternative designPlausiblen/aStrong
C. Home agentModel developerPlausible as component supplierWeak to plausiblen/aPlausible (Art. 8(1)(b))
C. Home agentOwnern/aPlausible (comparative fault)Weakn/a
Author’s doctrinal ratings for three invented agent failures (typical US state, Restatement (Third) approach; EU column applies the 2024 directive as written).

Three patterns come out of the table, and none depends on how autonomous the agent was. Harm type decides the frame. Physical injury in Case C opens strict liability against the device maker on both sides of the Atlantic, with a straightforward alternative design (no remote activation of heating elements without a presence check). Pure financial loss in Case A shuts product liability almost entirely and leaves the user with contract claims against the platform, often capped by its terms of service.

Attribution cuts against victims more often than for them. In Case A, the agent-as-agent idea does not help the consumer at all; it helps the merchants, who can hold the consumer to bookings made by her own software, much as E-SIGN contemplates. The same logic that made Air Canada answer for its chatbot makes the consumer answer for hers.

The deployer is the defendant every doctrine reaches. In Case B, the firm that handed an agent administrator credentials without an approval gate is the obvious negligent party, and it is also contractually liable to its client. The platform vendor and the model developer sit behind the economic loss rule and their license terms. Under the EU directive, the client’s data loss is excluded because the data were used for professional purposes.

What the exercise shows, and what it does not

It shows that the doctrinal choice is made mostly by facts other than the AI: who was hurt, how, and which contracts already link the parties. It also shows that component-supplier liability for model developers, the most contested question in the field, is live only in Case C, where the finished product is a physical device and the injury is physical.

It does not show how a court would rule. The ratings compress fact-heavy questions into a single word. States differ on the economic loss rule, component suppliers and software-as-product, and three hypotheticals cannot represent every deployment.

Brass scales of justice beside an open laptop in a dim law office (AI-generated image)
Product law, negligence and agency law each weigh the same facts differently. AI-generated image.

Where the law points

Deployers, not model developers, are likely to carry most AI agent liability for the next several years, through negligence and contract rather than through any new doctrine. Moffatt and Mobley both point there, and the facts of most agent failures (broad permissions, no confirmation step, no monitoring) look like ordinary carelessness by the business that switched the agent on. The failure patterns in recent agent-containment reports read like a list of missing precautions, which is exactly what negligence law knows how to price.

Strict product liability will arrive first where agents are embedded in physical goods, and first in Europe, because the directive removes the “is software a product?” fight. US courts are likely to keep splitting along the design-versus-content line from Garcia, which works tolerably for apps but poorly for agents, whose “content” is often an action.

Agency law will probably be used for attribution rather than for liability: deciding whose act a transaction was, not who pays for an accident. Treating the software itself as an agent with duties of its own would need legislation, and the withdrawal of the EU’s AI liability proposal suggests little appetite for bespoke regimes. For the broader regulatory picture, see the overviews of state AI laws and how AI is regulated in the EU and US.

What would settle it

Several concrete developments would move this from argument to law, and each could be checked.

  • A US appellate decision on whether an AI agent embedded in a physical product, or sold as software that acts on devices, is a “product” under the Restatement (Third). Prediction: when one comes, it will follow Garcia’s design-versus-content split rather than Rodgers’ tangibility test.
  • A reported judgment applying a reasonable-alternative-design analysis to an agent’s missing confirmation step or permission limit. If courts accept “ask before irreversible actions” as a feasible alternative, design-defect claims against agent builders become routine.
  • National transpositions of Directive 2024/2853 and the first cases under them, especially on whether a model developer is a component manufacturer when its model is reached through an API.
  • A case where a consumer disputes a transaction made by her own agent. Prediction: courts will bind her to it under E-SIGN or UETA-style attribution unless the merchant knew or should have known the agent exceeded its instructions.
  • Empirical work on insurance and contract terms: whether agent platforms’ terms of service and enterprise licenses already shift most losses to deployers. If they do, the tort question matters mainly for third-party victims.
A self-driving car with roof sensors on a city street at dusk (AI-generated image)
Autonomous systems already operate on public roads. AI-generated image.

Common questions

Can an AI agent itself be sued?

No. An AI system has no legal personality in the US, the EU or the UK, so it cannot be a defendant, own assets or pay damages. Claims go against the companies and people who built, sold, deployed or used it.

If my AI assistant buys something I did not want, am I stuck with it?

Often, at least as against the seller, because US law allows contracts formed through electronic agents when the action is attributable to the person bound. Your remedies are more likely to lie against the assistant’s provider under its terms, or through card-dispute and consumer-protection rules, than in tort.

Is the company that made the AI model liable when someone else’s agent causes harm?

Possibly, but it is the hardest claim to win. In the US it would usually require showing that the model itself was defective as a component or that the developer was negligent, and the economic loss rule blocks many cases. The EU directive makes component manufacturers liable for defective components that cause covered damage.

Further reading

As an Amazon Associate I earn from qualifying purchases. A commission may be earned if you buy through this link, at no extra cost to you.

References

  1. Moffatt v. Air Canada, 2024 BCCRT 149 (B.C. Civil Resolution Tribunal, Feb. 14, 2024). Decision
  2. Mobley v. Workday, Inc., No. 23-cv-00770-RFL, Order Granting in Part and Denying in Part Motion to Dismiss (N.D. Cal. July 12, 2024). Order (PDF)
  3. Chang, C. K. (2026). Algorithm as Agent: AI Vendor Liability for Title VII Hiring Discrimination After Loper Bright. 79 Stanford Law Review Online 50. PDF
  4. Rodgers v. Christie, No. 19-2616 (3d Cir. Mar. 6, 2020). Opinion
  5. Garcia v. Character Technologies, Inc., No. 6:24-cv-01903, Order on Motions to Dismiss (M.D. Fla. May 21, 2025). Order (PDF)
  6. Associated Press (Jan. 2026). Google and chatbot maker Character to settle lawsuit alleging chatbot pushed teen to suicide. WTOP. Article
  7. Directive (EU) 2024/2853 of the European Parliament and of the Council of 23 October 2024 on liability for defective products. Official Journal L, 2024/2853. EUR-Lex
  8. European Parliament, Legislative Train Schedule: AI Liability Directive (withdrawal, C/2025/5423, Oct. 2025). Page
  9. East River Steamship Corp. v. Transamerica Delaval Inc., 476 U.S. 858 (1986). Opinion
  10. American Law Institute (1998). Restatement (Third) of Torts: Products Liability, §§ 2, 5, 19, 21. (Cited by section; no open full text consulted.)
  11. Selbst, A. D. (2020). Negligence and AI’s Human Users. 100 Boston University Law Review 1315. PDF
  12. Lior, A. (2020). AI Entities as AI Agents: Artificial Intelligence Liability and the AI Respondeat Superior Analogy. 46 Mitchell Hamline Law Review 1043. Repository
  13. Ayres, I., & Balkin, J. M. (2024). The Law of AI is the Law of Risky Agents Without Intentions. University of Chicago Law Review Online. PDF
  14. Kolt, N. (2025). Governing AI Agents. arXiv:2501.07913 (preprint). arXiv
  15. Gabison, G. A., & Xian, R. P. (2025). Inherent and emergent liability issues in LLM-based agentic systems: a principal-agent perspective. arXiv:2504.03255 (preprint). arXiv
  16. Electronic Signatures in Global and National Commerce Act, 15 U.S.C. § 7001(h). LII

429 Posts

Peter Kalafatis writes and edits X9M8. He holds a bachelor's degree from LIU Post and a master's in Media Studies from The New School, with academic work in the philosophy of science, human behavior, AI, and memetics.

View All Posts

Leave a Reply

Your email address will not be published. Required fields are marked *