Direct answer
When an agent acts through its own identity, the agent is the technical actor. When it acts for a person, that person's current and limited delegation must also be visible. When it acts for an organisation, there must be an authority record defining the organisation's purpose, scope, tools, duration and approval boundaries. In one action, the performer, represented subject and accountable party may be different.
In plain language
An assistant might prepare an email draft, send it in your name or issue an approved notice from the company's support address. The assistant performs the work in all three cases, but the authority used and the resulting accountability are not the same.
Why this matters
Confusing the actor with the represented person creates false human attribution, unlimited delegation and a way to hide organisational accountability behind an agent.
Do not confuse
- Technical actor: the identity that performs the action in the system.
- Represented subject: the person or organisation on whose behalf the action is taken.
- Grantor: the party that defines the delegation or organisational assignment.
- Accountable party: the party that carries organisational or legal responsibility for the outcome.
What should you do?
- For every action, record the actor, subject, organisation, initiator, approver and accountable party separately.
- State the delegation's scope, target, tools, duration and revocation conditions.
- Authorise drafting, sending, signing and payment as separate actions.
- Prevent an agent from expanding its own authority.
- In a multi-agent chain, give each sub-agent only the authority it needs.
- Give external users a non-misleading explanation of the AI interaction and the agent's organisational status.
How do you audit it?
- Does the token or log distinguish the actor from the subject?
- Was the person entitled to grant the authority?
- Does the action remain within the delegation's scope and duration?
- Is there an organisational assignment and a human sponsor?
- Did the agent use drafting authority as though it were sending authority?
- Was the identity of the agent or organisation presented misleadingly to the user?
Limit
A technical token may show an access path; it does not by itself prove that the action was legally or organisationally authorised. An AI agent does not become a legal person merely because it has a separate technical identity.
Remember in one sentence
Record separately who acted, on whose behalf and who is accountable for the outcome.
Sources for this record
- S23Information and Communication Technologies Authority (BTK), *Electronic Signature — General Information*Official authority
- S29Microsoft Learn, *Application and Service Principal Objects in Microsoft Entra ID*Provider documentation
- S30Microsoft Learn, *Managed Identities for Azure Resources*Provider documentation
- S31Google Cloud, *Identities for Workloads*Provider documentation
- S32AWS IAM, *Temporary Credentials with AWS Resources*Provider documentation
- S33IETF, RFC 6749, *The OAuth 2.0 Authorization Framework*Standard
- S34IETF, RFC 8693, *OAuth 2.0 Token Exchange*Standard