An identity is a beginning. Accountability is the goal.
Knowing who an agent is should start a deeper conversation about its authority, its actions, and the people responsible for its outcomes.
An AI agent can authenticate successfully and still make a decision the business did not intend. That is the distinction I want leaders to keep in view as they move from experimenting with agents to putting them to work.
Identity answers an essential question: who is acting? It does not, by itself, establish whether a particular action is appropriate, whether the agent understood the request, or who will take responsibility when the outcome is challenged.
Start with a complete chain of authority
Imagine an agent asked to prepare a supplier payment. Reading an invoice, checking a supplier record, changing bank details, and releasing funds are different business actions. They should not become one undifferentiated permission simply because one workflow can reach all four systems.
The agent needs a distinct identity. It also needs clear authority for the task, boundaries around the resources it can touch, and an explicit handoff when the action requires human judgment.
Every autonomous action should lead back to a clear business owner.
Make the questions operational
I would start a design review with four questions:
- Identity: Can we distinguish this agent from other agents, its developer, and the person requesting the work?
- Authority: What may it do for this request, on whose behalf, and for how long?
- Evidence: Can we connect the request, the permission decision, the tool call, and the result?
- Accountability: Who can approve an exception, stop the workflow, and own recovery?
These questions are useful before a product is selected. They expose the operating decisions that a technology purchase cannot make on the organization’s behalf.
Design for the moment things change
An agent’s responsibilities may expand as people add tools and connect new data. I would treat each expansion as a reason to revisit its authority. Access that was reasonable for a narrow task may be excessive for a broader one.
The goal is to make useful autonomy sustainable. Start with a bounded task, make the actions traceable, and establish how authority can be withdrawn. Then expand with evidence that the controls work.