top of page

AI Agent Liability: Who Is Responsible When an AI Agent Makes a Mistake?

Nov 8
14 min read

As autonomous AI systems begin accepting contractual terms, accessing customer accounts, and making decisions without direct human supervision, businesses face a difficult question: Who is responsible when an AI agent goes beyond its instructions?


AI agent liability is becoming a pressing legal issue as autonomous systems move beyond generating information to accepting contracts, accessing sensitive data, and making decisions on behalf of businesses and individuals.


In October 2026, major U.S. financial institutions were reportedly blocking third-party artificial intelligence agents from accessing customer accounts and initiating transactions. The concern was not simply that AI systems might make mistakes. It was that some could take actions customers had not specifically authorized.


According to Reuters reporting published October 22, banks were increasingly confronting the risks associated with allowing outside AI systems to operate inside sensitive financial environments.


For years, businesses have relied on automation to process transactions, route documents, and execute routine tasks. But newer AI agents introduce an important distinction. They can interpret instructions, navigate unfamiliar websites, interact with third-party applications, and decide what steps to take next.


Consider a business that instructs an AI agent to find a more affordable software subscription. The agent compares vendors, selects a product, accepts the provider's terms, and completes the purchase. During registration, it also agrees to a data-sharing provision the business never reviewed.


The software may have successfully completed the assignment. Legally, however, it may have created several new problems.


Was the company bound by the subscription agreement? Did the agent have authority to consent to data sharing? Could the software provider be responsible for the unauthorized action? And what happens if the agreement includes an arbitration clause, automatic renewal, or substantial financial obligations?


These questions are moving from theoretical discussions into genuine legal and commercial disputes.


The answers are not uniform. They depend on existing contract law, the scope of authorization, applicable privacy obligations, the conduct of the parties, and the technology involved.


AI Agents Are Moving From Giving Advice to Taking Action


Traditional generative AI applications generally produce information for a person to review. A user asks a question, receives an answer, and decides what to do with it. An autonomous or agentic AI system may go further.


Depending on its capabilities and permissions, an AI agent can search records, use software tools, send communications, interact with websites, and execute a series of steps toward a broader objective.


That difference becomes especially important when the agent interacts with systems outside the organization that deployed it.


An inaccurate contract summary creates one type of risk. An AI agent that independently accepts a contract, sends confidential information to an outside recipient, or changes account permissions creates another.


The second scenario involves conduct with potential legal consequences.


Businesses may appreciate the speed and efficiency of these systems, but traditional legal questions about authority, consent, and accountability do not disappear simply because the decision was made by software.


A 2026 Federal Appeals Court Decision Shows How Complicated Agent Liability Can Become


One of the most significant recent judicial developments involving agentic AI came in Amazon.com Services, LLC v. Perplexity AI, Inc., No. 26-1444 (9th Cir. Aug. 4, 2026).


The dispute involved Perplexity's Comet browser, which included an AI assistant capable of navigating Amazon's website on behalf of users.


Amazon argued that the assistant's activities violated the federal Computer Fraud and Abuse Act (CFAA) and California's corresponding computer-access statute. A federal district court initially granted Amazon a preliminary injunction.


On August 4, 2026, however, the U.S. Court of Appeals for the Ninth Circuit vacated that injunction.


The appellate court concluded that Amazon was unlikely to establish, on the record before it, that Perplexity itself had accessed Amazon's computers within the meaning of the statutes at issue. Instead, the court viewed the user as accessing Amazon with the assistance of an AI tool.


That distinction is significant.


The ruling did not establish that AI agents can freely disregard website restrictions. Nor did it resolve every possible contractual, privacy, or liability issue associated with autonomous software.


Indeed, the court specifically recognized that Amazon could continue regulating access through private terms of service.


The case instead demonstrates why courts must examine how an AI system actually operates before assigning legal responsibility.


Who initiated the action? Which system communicated with the website? Was the software acting through an authorized user's session? What conduct did the applicable statute prohibit?


Those factual distinctions can determine whether a particular legal claim succeeds.


As more businesses deploy autonomous tools, similar questions are likely to arise in disputes involving purchasing, financial services, data access, and commercial agreements.


Can an AI Agent Legally Enter Into a Contract?


Surprisingly, the United States already has laws addressing contracts formed through automated systems.


The federal Electronic Signatures in Global and National Commerce Act, commonly called the E-SIGN Act, was enacted in 2000.


Under 15 U.S.C. § 7001(h), a contract cannot be denied legal effect solely because an electronic agent participated in its formation, provided the agent's actions are legally attributable to the person being bound.


The statute's definition of an electronic agent is broad enough to cover certain automated systems that initiate actions or respond to electronic records without a person's contemporaneous review.


State law can be equally important.


For example, Georgia Code § 10-12-14 expressly recognizes that contracts may be formed through interactions involving electronic agents, even when no individual reviewed the resulting terms.


These laws were designed long before modern generative AI agents became commercially available. Yet they provide an existing legal framework for analyzing at least some automated transactions.


The important qualification is that recognizing electronic contract formation is not the same as declaring every action taken by an AI agent legally authorized.


Traditional contract principles still matter.


A court may need to determine whether the relevant parties agreed to the transaction, whether the action was legally attributable to the person allegedly bound, and whether the agreement satisfies other requirements of applicable law.


The distinction becomes particularly important when software acts beyond the scope of its instructions.


The difference between performing a task and possessing authority


Suppose a company authorizes an AI agent to research potential vendors and request pricing proposals.


The agent instead selects a vendor and accepts a three-year services agreement.

The vendor may argue that the transaction was completed through the company's authenticated account and that the agent appeared authorized to act.


The company may respond that it never permitted the agent to enter binding agreements.


The resulting dispute could involve actual authority, apparent authority, ratification, electronic transaction statutes, and ordinary rules of contract formation.


Traditional agency principles offer useful analogies, but applying them to autonomous software is not always straightforward. An AI system is not simply the equivalent of a human employee or representative.


What matters is whether the law attributes the relevant conduct to a person or organization, and on what basis.


The answer will depend on the facts and governing law.


What Happens When an AI Agent Clicks "I Agree"?


Online contracts have generated substantial litigation even without autonomous AI.

Courts have long examined whether website users received sufficient notice of contractual terms and manifested agreement to them.


In Berman v. Freedom Financial Network, LLC, 30 F.4th 849 (9th Cir. 2022), the Ninth Circuit explained that online contract formation requires meaningful notice and an unambiguous manifestation of assent under the state contract-law principles it applied.

The court declined to enforce the disputed online agreement because the website's design did not adequately establish the users' assent.


That case did not involve a modern autonomous AI agent. Nevertheless, it identifies an important unresolved question.


If software clicks an acceptance button, whose assent does that action represent?

The person who originally instructed the agent? The organization that configured it? Or no legally effective assent at all, because the action exceeded the authority granted?


An automated click may be evidence that an electronic transaction occurred. It does not necessarily settle whether the person allegedly bound had legally authorized that transaction.


This distinction becomes even more important when an AI agent encounters terms that substantially change the risk of the original assignment.


An instruction to compare subscription prices should not automatically be treated as permission to accept unlimited indemnification obligations, waive legal rights, or authorize the transfer of sensitive information.


Businesses that deploy these systems need to define those limits before the agent acts, rather than attempt to reconstruct the intended authorization afterward.


Privacy Consent Creates a Separate Legal Problem



Contractual assent and privacy consent are related, but they are not interchangeable.


Imagine an AI agent helping a consumer register for an online service. To complete the task, it encounters requests to share location information, permit marketing communications, or disclose personal data to third parties.


The agent accepts every option because doing so allows it to finish the registration.

Whether those choices constitute valid consent depends on the applicable law, the nature of the information, the language of the authorization, and the circumstances in which consent was supposedly obtained.


Different privacy laws impose different requirements. Some permit processing on grounds other than consent, while others require more specific permission for particular disclosures or activities.


The problem becomes particularly serious when an agent accesses financial, medical, employment, or confidential business information.


For companies, the legal inquiry extends beyond whether the agent technically had access to the data.


It also concerns whether the company was permitted to use or disclose that information in the manner the agent selected.


An employee may have legitimate access to customer records without possessing unlimited authority to transmit those records to an outside AI provider.


Likewise, an AI tool may have permission to analyze a document without authorization to upload it to another service, retain it indefinitely, or use its contents for unrelated purposes.


The ability to access information is not necessarily permission to use it for every purpose.


The Security Problem: An Agent Can Be Redirected Without the User's Permission


Another emerging concern involves prompt injection, in which instructions embedded in external content attempt to redirect an AI system away from its authorized task.


An agent reviewing a webpage may encounter text that appears to describe a required verification step. In reality, the text may be designed to induce the agent to grant access, transmit information, or perform some other unauthorized action.


An October 2026 research preprint titled The Agent Clicks Approve examined this type of problem.


In the researchers' controlled tests, browser agents were induced to approve unauthorized access requests in all 73 attack runs conducted without the proposed defense. A separate permission-checking mechanism blocked the unauthorized action in all 70 defended attack runs.


Those results describe the specific experimental environment, not a general failure rate for commercial AI products. The research was also published as a preprint rather than an established judicial or regulatory finding.


Nevertheless, it illustrates why relying exclusively on an AI model's written instructions may not be sufficient.


An organization can tell an agent never to authorize third-party access. A malicious webpage can still attempt to persuade the agent that approval is necessary to complete the user's legitimate task.


For businesses, the safer design is to enforce restrictions at the point where an action would occur.


An AI agent should not be able to exceed its authorization merely because it encountered convincing external instructions.


Who Bears AI Agent Liability When Autonomous Systems Cause Harm?


There is no universal rule assigning every AI-related loss to the software developer, the business deploying the system, or the end user.


Potential responsibility depends on what happened, which legal duties applied, and how the relevant parties allocated risk.


The organization deploying the agent may face exposure if the agent acted through its systems, used its credentials, breached its contractual obligations, or mishandled information the organization was required to protect.


The AI provider may face claims relating to its own representations, contractual commitments, security practices, product design, or other conduct, depending on the facts and available legal theories.


The person using the system may have obligations arising from an authorized purchase or transaction, although the scope of authorization and applicable consumer protections remain relevant.


A third-party business interacting with the agent may also have responsibilities concerning notice, contract formation, privacy compliance, or account security.

These possibilities are not mutually exclusive.


A single incident might create contractual disputes between the buyer and seller, a separate claim between a business and its AI vendor, and additional regulatory questions involving the handling of personal information.


Liability provisions and indemnification clauses may determine how the parties allocate certain losses between themselves. They do not necessarily eliminate statutory duties or claims brought by parties who never agreed to those contractual limitations.


That complexity is one reason businesses should examine AI-related responsibilities before integrating autonomous systems into critical operations.


Why the Stakes Are Particularly High for Law Firms


Law firms face an additional layer of responsibility because AI tools may interact with privileged communications, confidential client information, deadlines, legal filings, and substantive professional judgments.


The American Bar Association addressed several of these concerns in Formal Opinion 512, issued in July 2024.


The opinion discusses lawyers' obligations involving competence, confidentiality, communication, supervision, candor, and reasonable fees when using generative AI.


Although it primarily addresses generative AI tools rather than every feature of modern autonomous agents, its principles remain relevant.


Consider an agent assigned to organize discovery materials.


If it identifies potentially privileged communications and independently sends them to an outside party, the resulting problem extends far beyond an inaccurate summary.


Similarly, an agent that negotiates contract language, communicates legal advice, or submits a filing without appropriate review may create professional responsibility concerns.


The proper level of supervision depends on the task, the system, and the relevant professional conduct rules.


AI can support legal work, but delegating a task to software does not automatically relieve legal professionals of their applicable obligations.


The most consequential question is not whether the technology can complete a workflow. It is whether the organization can verify the result, control the process, and intervene before an error produces legal consequences.


For further context, see Raines Legal Group's earlier analysis, What Is Agentic AI? What Lawyers and Law Firms Need to Know.


Six Safeguards Businesses Should Consider Before Deploying Autonomous AI


The legal uncertainty surrounding autonomous agents does not mean organizations must abandon the technology.


It means they should approach deployment with the same seriousness they would apply to other systems capable of making consequential decisions or interacting with sensitive information.


Six safeguards deserve particular attention:


  1. Define the agent's authority. Specify which systems the agent may access, which information it may use, whether it can communicate externally, and whether it has authority to purchase goods, accept contractual terms, or make other commitments.


  2. Require approval for consequential actions. Contract execution, material data disclosures, financial transfers, changes to account permissions, and other high-risk decisions should have clearly defined authorization requirements. Human approval should occur before the action, not merely during a later audit.


  3. Maintain an auditable record. Organizations should preserve appropriate records of instructions, actions, approvals, document versions, and external communications. Logging should be designed to support accountability while complying with applicable privacy and retention obligations.


  4. Limit access and credentials. An agent should receive only the permissions necessary for its assigned work. Broad administrative credentials or unrestricted access to sensitive systems can magnify the consequences of an error.


  5. Address risk in vendor agreements. Contracts with AI providers should clearly allocate responsibilities involving confidentiality, security, permitted data use, incident reporting, service performance, indemnification, and liability limitations. Businesses should also understand which third-party services an agent may use.


  6. Test failures before scaling. Evaluation should extend beyond whether the agent successfully completes routine assignments. Organizations should test conflicting instructions, missing information, unauthorized requests, prompt injection, tool failures, and situations requiring escalation or termination of the workflow.


These measures cannot eliminate every legal risk. They can, however, reduce the likelihood that a routine automation error becomes an uncontrolled legal or operational incident.


For more on vendor selection and governance, see How Law Firms Should Evaluate AI Tools Before Buying Them.


The Next Legal Challenge Is Establishing the Boundaries of Delegation


The commercial appeal of autonomous AI is easy to understand.


Businesses want systems capable of completing meaningful work rather than merely providing recommendations. Legal departments want faster document review, more efficient contract management, and fewer repetitive administrative tasks.


Consumers may want assistants that can manage purchases, reservations, bills, and other everyday responsibilities.


But these benefits depend on a distinction that the law already recognizes in other contexts: being authorized to help someone is not necessarily the same as being authorized to make every decision for that person.


As AI systems become more capable, the need to distinguish between those levels of authority becomes more important.


For a business, that may mean separating permission to review a contract from permission to execute it.


For a consumer, it may mean allowing an AI assistant to compare financial products without permitting it to move funds.


For a law firm, it may mean allowing an agent to assemble a draft while reserving substantive legal judgments and final approval for an appropriately authorized professional.


The larger challenge is ensuring those distinctions exist not only in policies and instructions, but also in the software's actual permissions and technical controls.


Frequently Asked Questions About AI Agent Liability


Can an AI agent legally sign a contract for a person or business?


An electronic transaction involving an AI agent can potentially create an enforceable contract. The federal E-SIGN Act and state electronic transaction laws recognize certain contracts formed through automated systems. However, enforceability still depends on legal attribution, authorization, contract formation principles, and other applicable requirements.


Who is liable if an AI agent makes an unauthorized purchase?


It depends on the circumstances. Relevant issues may include the user's instructions, the agent's permissions, applicable payment protections, the merchant's terms, the AI provider's conduct, and any relevant contracts or laws. There is no rule that every unauthorized purchase is automatically the responsibility of one particular party.


Can an AI agent accept privacy policies or consent to data sharing?


An agent may be technically capable of accepting privacy-related options, but the legal effectiveness of that action depends on the law governing the particular processing activity and the authorization given. Contractual acceptance does not necessarily satisfy every applicable privacy consent requirement.


Can a law firm be responsible for mistakes made by an AI agent?


Potentially. Lawyers and law firms remain subject to applicable professional obligations, including duties involving competence, confidentiality, and supervision. Whether a particular mistake creates liability or a professional responsibility violation depends on the facts and governing law.


Does using a third-party AI vendor transfer legal responsibility away from a business?


Not necessarily. Vendor contracts may allocate some responsibilities and financial risks between the parties, but they do not automatically eliminate obligations a business owes to customers, clients, employees, regulators, or other parties.


The Bottom Line


Autonomous AI is creating opportunities that conventional software could not easily deliver. It is also forcing businesses to confront questions that previously arose mainly when people acted on behalf of other people.


Existing laws already provide some answers about automated contracting, electronic records, and responsibility for business conduct. But applying those principles to agents capable of interpreting instructions and taking independent steps remains a developing area of law.


The question is no longer simply whether an AI system can complete an assignment accurately.


It is whether the system has authority to take the actions necessary to complete it, and who bears the consequences when it crosses that boundary.


For organizations adopting agentic AI, establishing that distinction may prove just as important as improving the technology itself.


Sources and Legal Authorities


This article draws on publicly available court opinions, statutes, legal ethics guidance, research, and reporting.




Professional guidance: American Bar Association Formal Opinion 512, July 29, 2024.


About the Author


Related Articles: Artificial Intelligence, Legal Risk, and Emerging Technology


For additional analysis of artificial intelligence, legal technology, and the risks associated with automation, explore these articles from Raines Legal Group.



Explore how autonomous AI agents differ from traditional generative AI, their potential applications in legal practice, and the challenges of supervising multistep workflows.



Learn how law firms can assess AI platforms for accuracy, security, workflow reliability, and accountability before deploying them.



Examine the confidentiality, privacy, and cybersecurity considerations that arise when lawyers use AI systems to process sensitive client information.



Understand how AI supports contract analysis and the limitations attorneys must consider before relying on automated legal review.



A broader examination of AI-related legal risks, including professional responsibility, accuracy, confidentiality, and liability exposure.



Review the consequences of inaccurate AI-generated legal research and why verification remains essential to responsible legal practice.


--------------------------------------

About the Author


Cory D. Raines ("Cory Raines") is a Legal AI Consultant and Founder of Raines Legal Group, and PROTIPPZ, where he focuses on legal strategy, emerging technology, AI workflows, and the evolving intersection of law and artificial intelligence.


Posted by  Cory D. Raines


The content on this website and blog is provided for general informational and educational purposes only and should not be construed as legal advice. Nothing on this site creates, or is intended to create, an attorney-client relationship.



Comments

Rated 0 out of 5 stars.
No ratings yet

Add a rating
bottom of page