From chatbot to agent

A generative system normally produces content in response to a request. An agent may pursue an objective through multiple steps, select tools, interact with external services and perform operational actions.

This is functional autonomy; it does not amount to legal personality.

Attribution of acts

AI does not automatically become a legal person or a representative in the technical legal sense. The question is when the effects of an action should be attributed to the user, the organisation or another actor in the chain.

Analogies with agency and representation are useful, but they do not remove the need to define the agent’s authority and limits.

Consent and machine-to-machine negotiation

Where agents negotiate terms or execute transactions, human consent may be programmed or delegated ex ante within defined parameters. Digital identity, authentication and machine-readable authorisations become important.

Actions beyond those limits raise questions of risk allocation among user, provider and counterparty.

Errors, security and data

An erroneous action may result from ambiguous instructions, poor data, model output, external services, vulnerabilities or attacks such as prompt injection. Causal reconstruction requires logging and traceability.

Agents should also follow least-privilege principles, limiting access to data, credentials and tools.

Graduated autonomy and accountability

Not every action requires the same level of oversight. Financial thresholds, categories of acts, human approval and automated stopping rules can scale autonomy to risk.

The core legal question is not whether a machine can ‘will’ like a person, but when and under what conditions the legal system should attribute to a person or organisation the effects of actions autonomously performed by a machine entrusted with the capacity to act.