Skip to main content
Skip to content
Agentic Finance

AI Agent Contracts: Authority, Liability and Enforceability

Separate technical execution from legal attribution, contract formation and remedies, using scoped US, UNCITRAL, England and Wales, and EU sources.

AI agent liability and responsibility shown through a human lawyer, humanoid robot, scales of justice and legal documents
Reading options: audio, text size & sections

Evidence review · 12 November 2026 · Named legal scopes, not a global enforceability ranking

What Matters

Automation can produce legally effective transactions without giving software its own legal personality. The useful questions are which law applies, whose act is attributable to whom, what authority existed and what remedy follows a failure. DN’s authority-to-remedy map separates those questions. A signed message or successful smart-contract execution is evidence about an event; it does not decide every legal consequence.

DN Evidence Block

Reviewed 12 November 2026. Sources cover the US federal E-SIGN rule for electronic agents, the UNCITRAL Electronic Communications Convention, the 2024 UNCITRAL Model Law, the Law Commission’s England and Wales analysis, and EU AI Act transparency duties. Their legal status and scope differ. No global court-outcome dataset, liability percentage or independent legal-personhood assessment is claimed.

Original DN work: a practical map from authority evidence to a disputed action and possible remedy. The former policy simulator has been replaced with explicit review questions. These questions help preserve evidence; they do not adjudicate liability.

An agent buys equipment above its spending limit. The payment succeeds, the supplier ships, and the operator says the software was not authorized. There are now several questions: whether a contract was formed, who is bound, whether the internal limit was communicated or legally relevant, whether the product arrived and who bears the cost of the failure. “The transaction is onchain” answers none of those questions by itself.

The opposite shortcut is also unreliable: “No human clicked approve, so there is no contract.” Several legal frameworks expressly address automation. Their recognition rules remove particular objections to electronic or automated processes. They do not abolish every other requirement of the applicable law.

Claim Versus the Actual Legal Source

Four sources that should not be treated as interchangeable
Source What it supports What it does not establish
US 15 USC 7001(h) Electronic-agent involvement alone is not a ground to deny effect where the action is legally attributable to the person bound. Every output is attributable to the deployer or every resulting transaction is valid.
UNCITRAL Convention, Article 12 A recognition rule for automated message systems within the Convention’s applicable scope A universal rule for every domestic, consumer or financial transaction
UNCITRAL Model Law, 2024 A legislative model addressing automated contracting, including AI techniques Self-executing domestic law in every country
Law Commission advice, England and Wales An analysis that existing law can support smart legal contracts A judgment validating every deployed smart contract or a rule for all UK jurisdictions

The distinction changes the research task. A team relying on a model law must identify the relevant domestic enactment and its wording. A team relying on a treaty must check applicability, exclusions and relevant state positions. A technical white paper cannot fill either gap.

The US Rule Requires Attribution

15 USC 7001(h) addresses electronic agents in transactions in or affecting interstate or foreign commerce. The protection against rejection solely because of automation is conditional on legal attribution to the person to be bound. Section 7001(b) also preserves substantive rights and obligations beyond the specified form requirements. This article does not determine whether a particular purchase satisfies that condition.

Operationally, preserve the instruction that existed before the purchase, the identity and authority of the person who issued it, the version of the policy applied, the communications seen by the counterparty and the executed result. Those records help a reviewer investigate attribution. They do not create a rule that the last person to configure the agent always pays.

Suppose a hypothetical agent had authority to buy a specified item for up to $500 but commits to $700 after adding delivery and a subscription. The internal limit is relevant evidence. The legal review still needs the actual terms, communications, applicable law and relationship between the parties. Marking the event “out of policy” is an operational finding; declaring the contract void is a legal conclusion the log cannot supply.

International Automation Rules Have Boundaries

The UNCITRAL, Electronic Communications Convention addresses cross-border electronic contracting and includes an automated-message-system rule. Its scope and exclusions matter: the source identifies personal, family and household contracts and certain financial transactions among the exclusions. Applicability should be established before invoking Article 12. A treaty’s existence is not evidence that all places connected to a cloud service apply it to this transaction.

The UNCITRAL Model Law on Automated Contracting, adopted 11 July 2024, offers rules for legislators and addresses matters including outputs, code and dynamic information. It is not a complete code for every legal issue raised by AI. Distinguish the model’s text from a country’s enacted version and effective date. The phrase “based on UNCITRAL” is a starting point for locating the actual rule, not a substitute for it.

Smart Code and a Smart Legal Contract Are Different Claims

The Law Commission’s 25 November 2021 advice concluded that the existing law of England and Wales could support smart legal contracts. It distinguished software automation from code used to define or perform binding obligations and identified issues including formation, interpretation, remedies, consumers and jurisdiction. This is useful legal analysis for that jurisdiction; it is not a certificate attached to a contract address.

Before deploying, identify what the code is supposed to do in the parties’ agreement. Is it the agreed performance mechanism, a record of a separate promise, or a tool whose behavior can depart from the promise? Then state how version changes, bad inputs and disputes will be handled. If a price feed returns the wrong value, correct execution of the program may still leave a dispute about the agreement.

Consider a hypothetical delivery contract that releases payment after a sensor reports arrival. The event can establish that the specified signal occurred. It may not establish that the goods were complete, undamaged or delivered to the right person. That gap is a reason to define evidence and remedies before deployment, not to assume that settlement ends every possible claim.

DN Authority-to-Remedy Map

Use one row per disputed action; retain facts and unresolved questions separately
Stage Evidence to retain Question for the responsible reviewer
Mandate Instruction, scope, expiry, approver and policy version What authority existed at the relevant time?
Offer and acceptance Terms shown, communications, timestamps and identity evidence Was an agreement formed, and between whom?
Execution Tool call, transaction, chain, amount, recipient and response What actually happened, including retries?
Performance Delivery, acceptance test, omissions and objections Was the promised performance supplied?
Loss and causation Documented harm, intervening events and relevant control failures Which legal claim and causal explanation are supported?
Remedy and forum Contract procedure, governing law, forum, notices and deadlines What relief is available, against whom and through which process?

Do not reduce this map to a responsibility percentage. A missing record can block a conclusion even when the rest of the file is detailed. Equally, an operational control failure does not establish that every claimed loss follows from it. Keep the amount claimed, the evidence supporting it and the disputed assumptions distinct.

DN Authority-to-Remedy Map: downloadable documentation templates

The blank record preserves six unresolved areas. The constructed example marks only its first reference as recorded; it describes no actual organization or dispute. A recorded reference is not verified evidence. Use aliases, not personal names, confidential documents or wallet secrets. These files contain no legal approval or numerical verdict.

Schema version 1.0.0. The source review date concerns this article; the template has no record-creation timestamp. For a user-created copy, keep the evidence statuses accurate and record the creation time separately from any evidence review.

mandate
Instruction and authority at the relevant time
agreement
Terms, communications and parties
execution
Actual action and retry history
performance
Deliverable and acceptance evidence
loss_causation
Claimed loss and causal assumptions
remedy_forum
Applicable procedure, forum and deadlines

AI Regulation Does Not Replace the Liability Analysis

For the EU, the AI Act’s definitions allocate provider and deployer roles to the persons or bodies described in the regulation. This is not a general rule making the AI system itself a party to a damages claim. The separate question of liability requires the relevant cause of action and facts.

The Commission’s 20 July 2026 transparency-guidelines publication states that Article 50 duties apply from 2 August 2026. The duties concern specified systems and roles. For example, the direct-interaction provision includes an obviousness exception; it is not an unconditional requirement to place the same notice on every background machine-to-machine call.

The retained European Commission, AI Act Article 50 transparency obligations and European Commission, AI Act provider and deployer obligations are useful entry points. Check the actual provision, scope and application date. This article makes no claim that all high-risk obligations share the Article 50 date or that a transparency breach automatically determines compensation.

Personhood, Identity and Ownership Should Stay Separate

A durable identifier can help connect actions to a system. A wallet can hold assets under a technical control arrangement. A legal entity can have rights under its governing law. These are different propositions. None of the recognition rules reviewed above says that sophisticated software automatically acquires independent legal personality.

The organization behind the agent may still require a wrapper, documented contracting authority and a plan for control changes. Those issues belong in DN’s DAO and company authority guide. Rights in generated material require their own analysis; the US copyright evidence article separates human contribution, provider terms and rights in inputs.

A team should also distinguish an ethical argument for future machine rights from the legal mechanism it relies on today. Policy debate can be valuable without being described as enacted law. An operational requirement to protect an agent’s credentials is a security decision, not evidence that the software can sue in its own name.

Controls That Preserve Options When Something Goes Wrong

DN recommends explicit mandates, independent revocation, accessible records and a route to a responsible person. These are design recommendations. They should not be presented as a universal statutory appeal right for every automated decision. Where a particular law creates notice, review or redress rights, identify it separately.

Revocation should cover the authority that actually matters: session credentials, scheduled jobs, wallet permissions and any service that can reissue them. Keep the scope precise. Stopping future agent calls does not reverse an already settled payment, cancel a contract or discharge an existing duty. Record which actions were prevented and which still require reconciliation.

After an incident, preserve the initial instructions and event history before rewriting prompts or rotating systems. Separate the evidence copy from operational remediation. A useful incident report explains what changed and when, without treating a later improvement as proof of what the system was authorized to do earlier.

Methodology, Limits and Corrections

This is a dated primary-source review and an original evidence-organizing method. It is not a survey of every jurisdiction or a prediction of a court result. The hypothetical purchases illustrate questions, not real incidents or legal precedents. This review does not recommend a legal provider or rank jurisdictions. It provides information, not individualized legal advice.

12 November 2026 change log: separated treaty, model-law, US statutory and England-and-Wales analysis; narrowed EU timing claims; replaced automatic attribution and policy-simulator conclusions with explicit review questions. The existing canonical URL and media are retained. Send a correction with the relevant jurisdiction, source and affected statement to DN’s corrections desk.

Frequently Asked Questions

Does automation make a contract invalid?

Not by itself under the scoped recognition rules discussed here. Applicability, attribution, formation and other legal requirements still matter.

Does a spending-limit breach mean the operator owes nothing?

No conclusion follows from that fact alone. Preserve the limit, communications and execution record for the applicable authority and contract analysis.

Is the UNCITRAL Model Law binding everywhere?

No. It is a legislative model. Identify the relevant domestic enactment and effective text before relying on it.

Is a successful smart-contract transaction proof of enforceability?

It is evidence of technical execution. The agreement, applicable law, parties and remedy remain separate questions.

Can an agent be sued merely because it has a wallet?

A wallet address does not establish legal personality, standing or the correct defendant. Identify the legally relevant persons and the basis of the claim.

Does stopping the agent settle a dispute?

It can prevent further activity within the revoked authority. It does not automatically unwind transactions or resolve obligations already incurred.


Get the most talked about stories directly in your inbox

Join the Decentralised News briefing for independent crypto, DeFi and AI analysis. No spam, unsubscribe anytime.