Are Contracts Made By AI Agents Legally Enforceable?

EXECUTIVE SUMMARY

Agentic wallets empower AI to execute payments, contracts, and business decisions autonomously. But US law offers no legal personhood for AI systems. Every action an AI agent takes must be attributed to a human or corporate actor, even when the AI behaves unpredictably or beyond what the user or developer intended. This strains traditional agency, contract, and tort doctrines, which assume intent, foreseeability, and clear authority. Until new legal frameworks emerge, businesses must proactively manage risk through governance, monitoring, contractual allocation of responsibility, and regulatory compliance.

The problem with agentic wallets is not that they can move money. The problem is that they can move money without becoming legally responsible for what they do.

That distinction matters. Businesses are increasingly experimenting with AI systems that can pay invoices, rebalance portfolios, execute transactions, negotiate commercial terms, and manage financial operations with little direct human involvement. The value proposition is straightforward. Faster execution. Lower operational friction. Continuous decision-making. An autonomous system that can act rather than merely recommend.

The law has no equivalent concept. American commercial law was built around a simple assumption: every meaningful economic action can ultimately be traced to a person or a legally recognized entity. Rights, obligations, liability, ownership, and contractual capacity all begin there. Agentic wallets challenge that assumption:

• Agentic wallets can approve a payment, but they bear no liability for fraudulent ones.
• AI systems can enter transactions, but they cannot be sued when the transaction collapses.
• AI can manage assets but cannot own them.

These platforms perform functions similar to those of employees, agents, and financial managers, yet they possess none of the legal status required to occupy those roles.

The result is a growing mismatch between technical autonomy and legal accountability.

That mismatch becomes difficult to ignore once an agent begins making decisions that neither the user nor the developer specifically anticipated. If an AI agent misinterprets market conditions and executes a catastrophic trade, somebody will absorb the loss. If it routes payments through a sanctioned intermediary, regulators will look for a responsible party. If it enters a binding agreement on behalf of a business, a court will eventually decide whether that agreement can be enforced.

The legal system still needs a defendant. Current doctrine answers these questions through attribution. Every action performed by the agent must be assigned to a human or legal entity somewhere in the chain. That approach works reasonably well when software functions as a tool. It becomes increasingly strained when software functions as a decision-maker. The pressure on existing legal frameworks is only beginning.

The Missing Piece: Legal Personhood

Many discussions surrounding agentic AI eventually arrive at the concept of personhood. That is not an academic distraction. It sits at the center of the problem.

Legal personhood determines who can own property, enter into contracts, assert rights, incur obligations, and bear responsibility for misconduct. Human beings possess legal personhood. Corporations possess legal personhood. Trusts, partnerships, and other recognized entities occupy similar positions under the law. AI does not.

An agentic wallet may hold digital assets, execute transactions, communicate with counterparties, and commit resources. From a functional perspective, it can appear indistinguishable from a commercial actor. From a legal perspective, it does not exist.

Courts cannot impose liability on software. Regulators cannot fine software. Counterparties cannot collect damages from software. Someone else must accept responsibility.

That requirement produces increasingly artificial results as autonomy expands. A user may authorize an objective without understanding the specific path the agent ultimately selects. A developer may create the system without anticipating the particular circumstances under which it operates. Yet both remain candidates for legal responsibility because the AI itself remains outside the legal framework.

Periodic proposals to create some form of electronic personhood have emerged over the years. The idea received attention in Europe during debates over autonomous robotics and advanced AI systems. The theory is simple enough: grant highly autonomous systems a limited legal status capable of holding assets, carrying insurance, and assuming liability.

The objections are equally obvious. Personhood is not merely a liability mechanism. It carries rights. Once lawmakers begin creating legal entities for autonomous software, the discussion moves well beyond commercial accountability. Few regulators appear eager to cross that threshold. American law certainly has not.

No federal statute, state law, or commercial code provision recognizes AI systems as independent legal actors. Existing frameworks continue to rely on attribution, regardless of how autonomous the underlying technology becomes.

That approach may prove adequate for conventional automation. It becomes increasingly difficult to defend when systems begin making consequential decisions that neither their users nor their developers fully control.

The Limits of Agency Law

The instinctive response is to treat agentic wallets as another version of traditional agency. The comparison is apt because it avoids the need to create new legal doctrines. A principal authorizes an agent. The agent acts on the principal’s behalf, and the principal bears responsibility for authorized conduct.

The problem is that agency law was built around human actors. Agency doctrine assumes that agents possess intent. They understand instructions. They exercise judgment. They can distinguish between authorized and unauthorized conduct. Most importantly, they can explain why they acted.

AI systems cannot do any of those things in a legal sense. An agentic wallet does not intend to act on behalf of its user. It processes inputs, pursues objectives, and produces outputs. The distinction may sound philosophical. It becomes practical the moment a dispute reaches court.

Traditional agency relationships also assume a degree of predictability. A company may not know every decision an employee will make, but it generally understands the range of conduct that employment entails. That assumption weakens considerably when machine learning systems begin generating outcomes that neither users nor developers anticipated.

An AI instructed to reduce transaction costs may route payments through intermediaries that create regulatory exposure. An agent tasked with maximizing yield may discover strategies that increase returns while creating substantial legal risk. The system has done precisely what it was asked to do. The problem is how it chose to do it.

Authority presents an equally difficult challenge. Users rarely provide precise instructions. They provide objectives:

• “Maximize yield.”
• “Reduce expenses.”
• “Find arbitrage opportunities.”
• “Optimize treasury management.”

Those directives grant substantial discretion. Courts routinely analyze whether a human agent acted within the scope of authority. The exercise becomes far more difficult when authority is expressed as a broad goal and the underlying decision-making process is largely opaque.

At some point, courts will need to determine whether a particular transaction fell within the user’s authorization. The answer may be obvious in straightforward cases. It becomes much less obvious when an autonomous system develops strategies that no one specifically contemplated.

Current doctrine largely avoids the issue through functional attribution. If the AI executed the transaction under some level of delegated authority, the law is likely to treat the transaction as if the user executed it personally.

That approach preserves existing legal frameworks. It also transfers substantial risk to users, especially in contract law.

Contract formation requires mutual assent. Courts traditionally describe this as a meeting of the minds, although commercial practice has long tolerated automated systems acting on behalf of human parties.

Agentic wallets push that principle much further. An AI system may accept terms, negotiate pricing, execute transactions, or commit an organization to contractual obligations without direct human review of any specific decision. Courts are unlikely to invalidate those agreements merely because a human failed to review every step in the process.

That creates a difficult reality for businesses deploying autonomous systems. A transaction entered by an AI may be binding even when no human specifically approved it. Broad instructions create broad exposure.

The law generally rewards certainty in commercial transactions. Counterparties expect enforceable agreements. Courts share that preference. Businesses may discover that autonomy cuts both ways.

Liability without an Actor

Liability presents the most significant challenge in the entire framework. An autonomous system can cause substantial harm. It cannot bear responsibility for that harm.

So, every liability analysis must decide who stands closest to the conduct. The answer may be the user. It may be the developer. It may be the platform operator. In many cases, it will be some combination of all three.

Existing doctrines offer familiar tools. Negligence. Product liability. Vicarious liability. Fiduciary principles. Regulatory enforcement. Courts will use whatever framework appears most appropriate under the facts presented.

The difficulty is that agentic systems generate failures that do not fit neatly within any of those categories.

Consider a compromised wallet agent that authorizes payments to fraudulent addresses after a malicious actor exploits a vulnerability in its decision-making process. Consider an investment agent that discovers a trading strategy that appears profitable but inadvertently manipulates a market. Consider a treasury management system that enters a series of transactions producing losses that no user intended and no developer anticipated.

Each scenario creates harm, but none fit comfortably within traditional fault models.

Liability doctrine generally relies on two concepts: causation and responsibility. Courts ask who caused the harm and whether the resulting conduct was sufficiently blameworthy to justify liability.

Autonomous systems complicate both inquiries. The chain of causation may involve users, developers, model providers, data sources, platform operators, and third-party integrations. Decision-making processes may be partially opaque even to the organizations that created them. Conduct may emerge after deployment through interactions no participant specifically intended.

The accountability gap appears when harm occurs, but responsibility becomes difficult to locate. Courts are unlikely to tolerate that gap for long.

The most probable response is increased reliance on foreseeability. Judges will ask whether the outcome should reasonably have been anticipated. They will examine system design, available safeguards, monitoring procedures, user conduct, and developer choices.

Foreseeability developed in a world of mechanical failures and human mistakes. Learning systems create different problems. A neural network that develops an unintended strategy six months after deployment presents a far different challenge than a defective machine component.

The legal framework remains the same. The factual analysis becomes substantially harder, and neither extreme solution is likely to prevail.

Holding users strictly liable for every autonomous action would discourage adopting systems they cannot fully control. Holding developers responsible for every outcome would make meaningful autonomy commercially impossible.

Courts are more likely to evaluate liability through a multifactor analysis focused on control, predictability, and safeguards.

• Control: A user who establishes narrow operating parameters, reviews transactions, and actively supervises the system presents a different case than a user who delegates broad authority and disappears.
• Predictability: A transparent system capable of explaining its reasoning creates different expectations than a black-box model generating unexpected outcomes.
• Safeguards: Transaction controls. Approval requirements. Audit logs. Monitoring systems. Emergency shutdown mechanisms. The presence or absence of these measures will become central in future litigation.

Many disputes will ultimately turn on a simple question: who was in the best position to prevent the harm?

Accountability Will Remain Human

Much of the discussion surrounding agentic wallets focuses on capability. The more consequential question is responsibility. The technology is advancing toward increasingly autonomous economic decision-making. The legal system is not moving at the same pace.

Businesses deploying agentic wallets should assume that courts and regulators will continue looking for accountable humans and accountable organizations. The AI itself will not absorb the liability. Someone else will.

Organizations need governance structures capable of constraining autonomous behavior. Spending limits, approval thresholds, audit trails, and monitoring systems are no longer operational conveniences. They are risk-management tools.

Contractual allocation matters as well. User agreements, licensing terms, indemnification provisions, and responsibility frameworks should be drafted with autonomous conduct in mind. The question is not whether a dispute will arise. The question is who will bear the resulting loss.

Compliance obligations remain equally important. Agentic wallets do not operate outside existing financial regulations simply because the decision-maker happens to be lines of code. Securities laws, lending regulations, licensing requirements, reporting obligations, and consumer protection rules continue to apply.

The content of this article is intended to provide a general guide to the subject matter. Specialist advice should be sought about your specific circumstances.