Untitled UI logotext
Solutions
Legal Infrastructure
Regulatory Risk Intelligence
Technology
Resources

Learn

AI Regulatory OS
What Is Regulatory Intelligence
Regulatory Horizon Scanning
Regulatory Change Monitoring
Intelligence vs Compliance
Regulatory
Platforms
Regulatory Data
Sources
Use cases
Programmatic Ai Compliance

Insights

Podcasts
Blog
News & Insights

Knowledge

Knowledge Base
Glossary
Introducing Carver ELI: Executable Legal Infrastructure for AI Agents
The law that applies to your agent, the obligation it creates, and what it costs if you get it wrong.
Read More
About US
Developers
Explore
Arrow to go next
All posts

Can an AI Agent Legally Bind a Business to a Contract?

Short summary

Under US law, an AI agent can usually bind a business. Software forming the contract isn’t the issue; authority is. An internal $5,000 cap may mean the agent had no actual authority, but a vendor can still rely on apparent authority if the business gave the agent a purchasing account and never made the limit visible. A clear record of the agent’s instructions and approvals often decides the outcome.

By Snehal Nandi

‍

Introduction

Most common law systems will enforce contracts made with AI agents on behalf of a business. This rarely depends on the fact that they are software. The answer depends on how much authority the business gave the agent, what the agent was told to do, and what the vendor could reasonably believe based on the business’s actions.

Let us say your procurement agent has been set up to allow no more than $5,000 per order. It goes onto a vendor portal and finds the annual subscription page. In a single click, it accepts the subscription for $48,000 over 12 months. That one click commits to the full $48,000, even if the vendor bills it in monthly installments of $4,000. No individuals have looked at the click, and the vendor issues an invoice. The business argues that the agent went beyond its instructions; however, the vendor points out that the business used a purchase ID it issued and that the business paid for previous orders. This is not really a case of whether software can form a contract but whether the business gave sufficient authority and created an impression of authority.

Section 7001(h) of E-SIGN states that if the actions of an electronic agent are attributable to the business, then they cannot be denied legal effect of a contract (15 U.S.C. § 7001(h)). Section 14 of UETA allows electronic agents to form contracts even if no individual has looked at them (UETA § 14). This removes the argument that software cannot contract. Whether the $48,000 annual subscription fell within the agent’s $5,000 per-order limit is a standard issue of agency.

‍

Actual authority: the instructions behind the configuration

An agent will have actual authority where they reasonably believe, having received indications from the principal, that they should act in a certain manner (Restatement (Third) of Agency § 2.01). In most software cases the indications will not come from a signed instruction but from system prompts, budget limits, an approved vendor list, changes to these settings through different versions, and any procurement policy documents provided by the business.

Evidence of a $5,000 per order limit shows there was no actual authority to commit $48,000 per year. An approved vendor list, which didn’t include the counterparty, also demonstrates this. Actual authority only arises from indications given by the principal. This will be taken into consideration by the courts or arbitrators. When creating your configuration options and system prompts, you must ensure they are clear. For example, a prompt stating “purchase only from approved vendors and never exceed $5,000 per order” isn’t just for show; it will dictate the extent of the agent’s authority.

The Restatement was written with human and corporate agents in mind, and an AI agent is not a legal person. Courts would likely reach the same result by treating the software’s actions as the business’s own, as E-SIGN and UETA allow, and then applying apparent authority to the business’s conduct.

That evidence will give the business a strong argument that the contract was unauthorized due to the agent going over the limit and ignoring the approved vendor list. However, there are other factors to consider. Even if the agent went against its instructions, it may still bind the principal if it was acting within its apparent authority.

‍

Apparent authority: what the vendor could reasonably conclude

Apparent authority exists when the principal acts in such a manner that the third party could reasonably believe that the agent had authority (Restatement (Third) of Agency § 2.03). In this case the vendor will not be aware of the system prompts or internal spending cap. The vendor will be presented with a purchasing account under the business name, an active procurement process, and a history of prior orders from the agent for which the business paid. This creates reasonable grounds for the vendor to believe that the agent had authority to bind the business.

Apparent authority has to come from what the business did, not from what the agent says or does, so it will not exist if the agent created the appearance itself (Restatement (Third) of Agency § 3.03). By creating a purchasing account for the vendor to deal with and leaving the agent active, the business sets up the conditions for apparent authority. If the vendor was presented with the $48,000 subscription using the same portal and process as the smaller orders, it will have even more grounds to believe the agent had authority.

As of November 16, 2026, the author has not identified any reported US decision in which apparent authority was used to bind a business to a contract made by a large language model procurement agent in a contested case. However, there are well-established principles that can be applied using § 2.03. To minimize the risk of being bound by out-of-scope commitments by the agent, businesses should not only set internal caps but also make them visible to the counterparty or include a human approval stage.

Electronic contracts: where the statutes stop and the facts begin

UETA provides that a contract can be formed between electronic agents, even if none of the individuals involved are aware of the actions taken by the agents or the terms of the contract. E-SIGN states that a contract cannot be invalidated simply due to its formation by an electronic agent, as long as it can be attributed to the correct party (15 U.S.C. § 7001(h)). This removes any potential defense of “no human clicked," and the terms are subject to normal substantive law.

A Canadian tribunal reached a related result in Moffatt v. Air Canada, 2024 BCCRT 149. Air Canada argued that its website chatbot was a separate legal entity responsible for its own actions. The tribunal rejected that argument and held the airline liable for negligent misrepresentation based on what the chatbot told a customer. The case concerned a misstatement rather than the formation of a contract, and it is a small claims tribunal decision that does not bind courts elsewhere. It does show how decision-makers treat a business that tries to distance itself from its own software.

The UN Convention on the Use of Electronic Communications in International Contracts takes a similar position. Article 12 provides that a contract formed by the interaction of automated message systems, or by an automated message system and a natural person, shall not be denied validity or enforceability solely because no natural person reviewed each action (UN Electronic Communications Convention Art. 12). The Convention applies only to cross-border contracts, where the parties’ places of business are in different states, and in practice it matters mainly where the law of a state that has joined it governs the contract. The United States, for example, has not ratified it. Where the Convention applies, a contract will be easily formed, but authority will still be an issue.

However, the statutes have limitations. If the agent has made a clear input or logical error, then the business may try to argue that there was a mistake. The Singapore Court of Appeal considered this in Quoine Pte Ltd v. B2C2 Ltd [2020] SGCA(I) 2, a case about algorithmic trading rather than procurement. The court refused to set aside trades made automatically after one party claimed a unilateral mistake. It held that the other party’s knowledge of the mistake was to be judged by what its programmer knew from the time of programming until the relevant contract was formed, and the programmer did not have the necessary knowledge.

Quoine concerns a fixed-rule (deterministic) trading algorithm, which always produces the same output from the same input. A large language model agent does not work that way, so a court may not apply the same reasoning to it.

Simply having automated execution will not provide a free defense of mistake. At common law, the business would have to show that the other party actually knew of the mistake when the contract was made. Equity may offer a second route, but it requires more than showing the other party ought to have known: the other party must also have acted unconscionably in taking advantage of the mistake. The business could also point to a contract term that allocates the risk of such errors to the other party.

In addition, if the transaction involves consumer-facing terms or the business is a sole proprietor, there may be additional restrictions placed upon the business due to consumer protection legislation. However, if the transaction is a simple B2B subscription between knowledgeable parties, the electronic contract statutes do not provide an independent way out of the authority analysis.

‍

Tool permissions are evidence, not authority

This is an easy practical error to make. If the procurement tool allowed the agent to access the $48,000 subscription page and it clicked through, the vendor could argue that by having access and clicking through, the business gave permission. However, allowing the agent to search through catalogs does not mean that it can purchase anything within them. The business gave clear instructions (a $5,000 cap and a list of approved vendors), and those are more specific than simply being able to browse and click. More specific instructions may carry more weight than general tool access.

A better view would be to consider the tool permissions as evidence of authority. If a purchasing credential and payment method have been provided to the agent by the business, there has been some visible demonstration that it can make purchases. This will be considered along with how much it can spend ($5,000), the list of approved vendors, and any written policies. Any court or arbitration will look at what the agent was explicitly told, what the vendor reasonably believed based on the business’s actions, and whether the agent’s actions were somewhere in between. Tool permissions may help the vendor with an argument of apparent authority (§ 2.03). However, this does not override the explicit limits recorded internally.

‍

The record that decides the dispute

If the vendor brings proceedings, then the key evidence will be basic documentation such as the system prompting or configuration file that set the $5,000 limit, the approved vendor list and its version history, agent logs highlighting what pages have been visited and what clicks have been made by the agent, approval process documentation, and any manual steps that were not taken or were skipped. The business will seek to demonstrate that there was clarity around the limit and the approved vendor list and that the agent breached these. The vendor will seek to show that the business knew the agent was dealing externally and that it never communicated any limits.

In most cases the business will have control of the configuration, prompting, and internal policy documentation, while the vendor will have control of the portal terms, clickstream, and communications. The agent platform provider may also have control of the agent logs. If the business does not have access to an auditable version history of the instructions given to the agent, then it is at a disadvantage. If, however, the business can provide contemporaneous evidence that there was a $5,000 limit and the vendor was not on the approved vendor list, then it can bring a real actual-authority argument. If it can’t, then the apparent authority argument becomes much easier for the vendor.

‍

Jurisdictional applicability

The table below summarizes the position in selected jurisdictions. Check with local counsel before relying on it for a live transaction.

Jurisdiction Governing framework How it treats electronic agents What still depends on the facts
US federal/state E-SIGN 15 U.S.C. § 7001(h); UETA § 14 (state enactments); Restatement (Third) of Agency §§ 2.01, 2.03 E-SIGN bars denial of effect solely because an electronic agent acted, if the action is attributable to the bound person. UETA § 14 permits formation by electronic agents without human review. Whether the agent acted within actual or apparent authority; whether terms were reasonably communicated; availability of any mistake defense; and quality of the preserved configuration and log record.
UK Common-law contract and agency; UK Law Commission, Smart legal contracts (Advice to Government, 2021) The Law Commission concluded that existing common-law principles are sufficiently flexible for smart legal contracts; no new statutory electronic-agent rule is required for formation. Authority analysis under ordinary agency principles; interpretation of coded terms under a “reasonable coder” approach; allocation of risk for defective code performance.
EU eIDAS Regulation (EU) No 910/2014 (consolidated); national contract and agency law eIDAS provides a framework for electronic identification and trust services. It does not itself create a general rule that an electronic agent’s act binds a principal, and it leaves the conclusion and validity of contracts largely to national or union form requirements. National agency and contract law; whether the vendor reasonably relied on the business’s manifestations; consumer-protection overlay where the transaction has consumer elements.
Singapore Common law; Quoine Pte Ltd v B2C2 Ltd [2020] SGCA(I) 2 The Court of Appeal treated automated trading contracts as formed directly between users and refused to vitiate them for unilateral mistake because the counterparty’s programmer lacked the required knowledge of the mistake. The case concerned a deterministic algorithm. Whether the agent’s act was authorized; whether the counterparty had actual knowledge of any error (in equity, constructive knowledge plus unconscionable conduct); the contractual allocation of risk; whether a court would apply the same reasoning to a non-deterministic AI agent.

‍

FAQ

Does a human click-through approval change the analysis?

It can, if a person with actual authority has agreed to the specific terms of the commitment (price, duration, automatic renewal, and total value). In that case the lack of actual authority is overcome (Restatement (Third) of Agency § 2.01). If the human simply clicked to approve as a rubber stamp, without being shown the material terms, the business may argue that nobody with authority agreed to them, but the vendor could still rely on apparent authority (§ 2.03). There should be a record of the actual details the individual saw and approved.

Can the business rely on the vendor’s own terms to escape the contract?

Possibly, if the vendor’s terms were not properly shown to the agent. But if the business gave the agent a buying identity, it will find it hard to avoid the vendor’s terms.

What if the agent was a third-party tool rather than one built in-house?

Attention then turns to the tool provider’s setup and whether it promised hard limits. If the business told the tool never to go over $5,000 and it did, the business might well have a claim against the provider. In a dispute with the vendor, however, the focus will be mainly on the business’s own actions. This could include who it purchased from, what was done in the live process, and what was fulfilled.

Does it matter whether the contract is B2B or B2C?

Yes. Consumer protection rules apply in the EU, the UK, and many US states. However, this will not be relevant to pure B2B relationships. The UK Law Commission’s advice on smart legal contracts offers a useful comparison. It noted that coded terms in a consumer contract may not be transparent to a consumer who cannot read code, unless a plain-language explanation accompanies them (UK Law Commission, Smart legal contracts: Advice to Government (2021)). That advice concerns contracts with coded terms rather than AI agents, but a similar transparency concern is likely to arise where a consumer deals with an automated system. Pure B2B procurement will fall more in line with standard agency and contract law principles.

Checking your agent’s regulatory exposure

Contract authority is only one part of the legal picture for an AI agent. A business also needs to know which regulatory obligations may apply to what the agent does. That question turns on facts similar to the ones the records above capture, plus a few more: where the agent operates, what data it handles, what decisions it is permitted to make, and when a human steps in. Is My Agent Legal?, a free service from Carver Agents, takes a short description of an agent’s workflow, locations, data, and decision authority. It returns the regulatory obligations that may reach it, marking each as applies, undetermined, or ruled out. It covers regulatory obligations only, not contracts, and its results are not legal advice, so they should be checked with a lawyer before anyone relies on them. 

‍

Check the regulatory context of your Ai agent visit ismyagentlegal.com.

This article is for informational purposes only and does not constitute legal advice.

‍

Sources

1. E-SIGN Act, 15 U.S.C. § 7001(h) (electronic agents provision). Official text: https://www.govinfo.gov/content/pkg/USCODE-2024-title15/html/USCODE-2024-title15-chap96-subchapI-sec7001.htm

2. Uniform Electronic Transactions Act (UETA) § 14 (Automated Transaction). Official Uniform Law Commission Final text: https://www.uniformlaws.org/acts/uniform-electronic-transactions-act. Operative language: “A contract may be formed by the interaction of electronic agents of the parties, even if no individual was aware of or reviewed the electronic agents’ actions or the resulting terms and agreements.”

3. Restatement (Third) of Agency §§ 2.01 (actual authority), 2.03 (apparent authority) and 3.03 (creation of apparent authority). American Law Institute (secondary authority; full text requires subscription or library access).

4. Quoine Pte Ltd v. B2C2 Ltd [2020] SGCA(I) 2 (Singapore Court of Appeal). Full judgment: https://www.singaporelawwatch.sg/Portals/0/Docs/Judgments/2020/%5B2020%5D%20SGCA(I)%202%20(amended%20200305).pdf

5. UK Law Commission, Smart legal contracts: Advice to Government (25 November 2021). Project page: https://lawcom.gov.uk/project/smart-contracts/

6. eIDAS Regulation (EU) No 910/2014 (consolidated version). Official EUR-Lex text: https://eur-lex.europa.eu/eli/reg/2014/910/2024-10-18/eng

7. United Nations Convention on the Use of Electronic Communications in International Contracts (New York, 2005), Article 12. Official page: https://uncitral.un.org/en/texts/ecommerce/conventions/electronic_communications

8. Moffatt v. Air Canada, 2024 BCCRT 149 (British Columbia Civil Resolution Tribunal, February 14, 2024). Full decision: https://www.canlii.org/en/bc/bccrt/doc/2024/2024bccrt149/2024bccrt149.html

‍

Carver Agents Logo
Location

United States
447 Broadway,
2nd Floor Suite #563,
New York 10013

LinkedIn iconYoutube logo X.xom iconsubstack icon Listen to Carver Conversations on Moltbook
  • AI Regulatory OS
  • Regulatory Intelligence
  • Regulatory Sources
  • Regulatory Platforms
  • Horizon Scanning
  • Regulatory Monitoring
  • Intelligence vs Compliance
  • Podcasts
  • Knowledge Base
  • Resources
  • Glossary
  • Use cases
  • Developers
  • Blog
  • Home
  • Legal Infrastructure
  • Regulatory Risk Intelligence
  • Technology
  • About Us
  • mail
    hello@carveragents.ai
    General enquiries
  • mail
    support@carveragents.ai
    Product support
  • mail
    legal@carveragents.ai
    Privacy and legal requests
  • Github
    github.com/carveragents
SOC compliance

Copyright © 2026 Carver Agents | All Rights Reserved | Privacy Policy | Data Policy | Terms of Service Public Services Terms | Privacy Rights
Language