Estonia plans state-issued ID codes for AI agents with scoped, auditable permissions

The first country to give AI agents their own legal identity is probably going to be Estonia. A plan approved by Prime Minister Kristen Michal on June 17 proposes state-issued “AI ID codes,” and it resurfaced across AI communities this week as agents became a mainstream product category.

What the plan proposes

The core idea is separation. An AI ID code identifies an agent as distinct from the person or company it acts for. Today an agent that books, buys or files on someone’s behalf usually does so by borrowing that person’s full credentials. Estonia’s model replaces inherited access with scoped, auditable permissions: view this data, prepare that document, initiate a payment within a set limit. Every action would be verifiable and attributable to a specific agent acting under a specific grant of rights.

Michal’s framing puts the accountability question first: “It must be clear who is acting on whose behalf with what rights, and who is ultimately responsible.”

Implementation details are still open. The government has not published the rollout mechanics, so how codes would be issued, to whom, and how permissions would be enforced across services remains to be defined.

Why it matters now

The timing of the renewed attention is not accidental. Agents are already booking, buying and filing for people, and the dominant way they do it is by using full human credentials, which creates two problems at once. It is a security hole, because an agent with a user’s complete access can do anything the user can. And it is a responsibility vacuum, because when something goes wrong there is no clean record of whether a person or a program took the action, or under what authority.

Estonia, the country that made e-residency a thing, is proposing the missing primitive: an identity layer for software actors, issued by the state, with permissions that can be limited and audited.

What to watch

The proposal is a policy direction rather than a working system, and its weight will depend on the mechanics the government has yet to publish. The questions that matter are practical: who can register an agent, how a spending cap or data scope is enforced at the point of action, and how liability flows when an agent with a valid ID exceeds or misuses its grant. If those are answered well, the model is exportable in the way e-residency was. If they are not, the ID code becomes a label without teeth. Either way, Estonia has put a concrete answer on the table to a question most agent deployments are currently answering with a shared password.