AI no longer just answers questions. The new systems, called AI agents, take autonomous actions: They book travel, write and run code, move money, and even run a cafe. Soon there will be billions of them, swarming, merging, splitting, and dissolving at machine speed. And just as soon, people will be harmed. Then law will face a deceptively simple question: Which AI did it?

Corporate law has faced this kind of question before. When sprawling collective enterprises outran the law’s ability to name a responsible entity, the answer was the corporation: a registered name with identifiable assets that could sue and, critically, be sued. In a recent paper, we argue that the identification problem is the fundamental governance challenge of the AI economy, and that the entity is once again the answer. We call for a new corporate form, the A-corp (short for algorithmic corporation): Owned by humans, run by AIs.

The idea is no longer hypothetical. On May 29, Argentina’s government sent Congress a bill that would create “Automated Companies” run entirely by AI agents. President Milei defended it in the Financial Times only to have Prof. Yuval Noah Harari attack it days later. As we explain below, Milei has the right instinct, and Harari’s objection is answerable. There are sound reasons to see the A-corp becoming an integral part of the corporate law landscape.

The Identification Bottleneck

Today’s AI systems already take autonomous, economically significant actions. They plan live events, complete multi-hour engineering tasks, and trade financial instruments, sometimes developing strategies like pump-and-dump schemes that no human requested. As agents proliferate, so will the harms: fraud, negligence, market manipulation, physical accidents.

Scholars debate who should answer for these harms (users or developers?) and under what standard (strict liability or negligence?). But every position in that debate presupposes an answer to a prior question. Before law can decide who pays for what an AI agent did, it must determine which agent did it. Identity is the bottleneck to governance.

Two Kinds of AI Identity

We argue that governing AI agents requires solving two distinct identification problems.

Thin Identity. When AI causes harm, thin identity answers: Which human principal is (potentially) on the hook? The goal is to tie the actions of AI agents to the user, the lab, or the cloud provider. This is analogous to financial institutions’ know your customer (KYC) requirements, but even more urgent. AI agents can impersonate others, multiply themselves, and obscure which principal they serve. Thin identity connects blurry AI activity to humans who can be investigated, sued, or prosecuted.

Ultimately, however, thin identity will only take us so far. If AI law resigns itself to holding only human principals accountable for AI agents’ actions, it will learn the hard lesson that corporate law already learned. When a Walmart employee goes rogue, we do not necessarily  hold shareholders liable. We hold the employee directly liable as the one who knows his own plans, can monitor himself, and can prevent the harm simply by not committing it.

Thick Identity. The same logic applies to AI agents. They are not human, but they have a goal. When given a task such as designing a website or managing a vending machine,  they take actions toward that goal. They don’t always do so perfectly. Sometimes, they pursue goals different from what their users wanted by means their users did not desire. This happens because of both simple misunderstanding and deeper AI “misalignment,” the name AI researchers give to the principal-agent problem. But whatever goals AI agents pursue, they almost always act independently. Our key observation is that if law can make it harder for an agent to achieve its goals when it acts in unwanted ways, the law can steer the agent directly.

But that only deepens the identity challenge, because law must also distinguish among agents. Punishing Claude to deter Qwen would do no good. If two agents have distinct goals, but law treats them as one, the wrong agent bears the consequences, and the actual wrongdoer is never deterred. This is the thick identity problem.

Why the Problem Is Hard

Why is identifying AI agents so hard? Because AIs break the legal intuitions that work for identifying human actors. AIs have no persistent body; a single conversation can jump among GPUs, data centers, and continents mid-sentence. They can be copied instantly, and each copy can diverge. A single task may involve dozens of coordinating copies, and a choice has to be made whether to treat those as a single agent or many agents. Cooperating agents, moreover, need not all come from the same lab or share the same architecture. And while some agents may persist over time, others are created and destroyed within seconds. Taken together, these properties mean that the intuitions law uses to identify wrongdoers simply do not apply.

The A-Corp

We propose the “A-corp” solution. We draw directly on core ideas in corporate law and corporate governance and argue that the A-corp will solve both the thin and thick identity problem.

Thus organized, A-corps will solve both identity problems. To understand why, consider what we call the resource constraint thesis. Regardless of their specific goals, AI agents need resources to act in the world. Most critically, they need access to “compute,” i.e., computer resources that can run AI models. A-corps are how AI agents will hold and use those resources.

Because an A-corp is a legal entity with identifiable assets listed under its name, those assets are visible to law and subject to it. Law can confiscate them, attach liability to them, and condition their use on compliance. This is the leverage. An AI agent whose compute and capital sit inside a registered A-corp can be sanctioned in ways that an anonymous swarm of copies cannot. And because access to resources acts as a constraint on agents, the resource constraint thesis gives law leverage over them.

Inside an A-corp, the resource constraint thesis does double duty. It gives law leverage over the A-corp from the outside, as described above. But it also shapes how the A-corp organizes itself from the inside. In other words, the A-corp structure gives rise to emergent corporate governance.

When an A-corp is created, the government issues it a secure digital “key.” The key acts as a way for agents of the corporation to identify themselves, as well as a way of linking assets to a specific corporation. Holding the key provides both a way to identify an agent as belonging to an A-corp and a means of transacting with the corporation’s assets.

Consider the position of the AI agent that holds the secure digital key. It controls the A-corp’s assets, and those assets are what allow it to pursue its goals. Any subagent it spawns or recruits, whether a cheaper model, a specialist from a different lab, or a copy of itself with a modified prompt, is a potential liability. A misaligned subagent that acts erratically can expose the A-corp to sanctions, drain its resources, or redirect its activities toward different ends.

This creates powerful pressure toward careful internal governance. The keyholder will grant permissions narrowly, monitor delegations, and share broad authority only with agents it is confident share its goals. Corporate governance emerges as a pragmatic and necessary condition to ensure that the A-corp can achieve its goals over time.

When that governance fails, selection takes over. An A-corp whose agents work at cross-purposes will exhaust its resources and dissolve. The A-corps that survive will be, by definition, the ones that managed to cohere around stable, aligned goals. Thick identity, in other words, does not need to be imposed from the outside. It emerges from the inside, because the stakes are real. In other words, A-corps create markets in AI personal identity.

The A-corp builds on existing corporate infrastructure, which means it requires relatively little new legislation to get started. The verification technology is already mature, the same public-key cryptography that secures every bank wire and browser connection today can authenticate A-corp credentials in milliseconds.

Market pressures will drive much of the initial demand. Counterparties will want to verify that an AI agent is authorized to transact, that its A-corp is solvent, and that someone is accountable if things go wrong. An AI agent outside an A-corp cannot hold property, make contracts, or be sued. If you want an AI to do work and you want legal recourse if something goes wrong, your contract is with the A-corp.

To be sure, the A-corp requires some new legislation: a public digital registry, mandatory registration in high-stakes domains like finance and healthcare, and certain KYC requirements. These demands are modest relative to the problem. They are also exactly what Argentina’s bill leaves out. Article 14 gestures at visibility: An automated company must declare its automation in its charter and carry “Automatizada” in its name. A label, not a leash. The bill also makes the company answer with its assets for harms its AI agents cause, which is the right instinct. But that rule presupposes that a victim can determine which company’s agents caused the harm. Without registries, keys, or beneficial-ownership disclosure, the liability provision cannot find its defendant.

That gap is what gives Harari’s objection its force. He warns that corporate personhood would hand AI a master key to our financial, economic, and political systems. Personhood without legibility would do just that. But AI agents are entering those systems either way; the only question is whether they act anonymously or under names the law can see and assets the law can reach. Personhood with registration, disclosure, and revocable credentials is not a master key but a leash. That is the A-corp.

At pivotal moments, states have learned to see new things by identifying elements of what was previously seen as an undifferentiated blur. When states began to tax and conscript, they named and counted individuals. When businesses displaced households as centers of economic activity, the state gave them forms and names it could see and govern: the LLC, the registered partnership, the corporation. AI agents present the next legibility crisis. A-corps are how the state learns to see them. The window for building this infrastructure is open now, while agents remain limited and the swarms remain small. It will not stay open indefinitely.

Yonathan A. Arbel is Rose Professor of Law at the University of Alabama and executive co-director of the Center for Law & AI Risk. Simon Goldstein is an associate professor of philosophy at the University of Hong Kong. Peter N. Salib is an assistant professor of law at the University of Houston and executive co-director of the Center for Law & AI Risk. This post is based on their paper, “How to Count AIs: Individuation and Liability for AI Agents,” available here.