Who Is Allowed to Decide

A person’s role may grant authority over one class of decisions but not another. A procurement manager can approve within a budget threshold, while a security officer can accept certain risks, and a customer owner can amend a delivery commitment. Agents need comparably specific boundaries.

Delegation should identify:

• Principal: Who grants the authority?
• Jurisdiction: Which domain, cases, systems, and parties are covered?
• Action: What may the agent read, recommend, prepare, or execute?
• Conditions: Which thresholds, evidence, and policies apply?
• Escalation: When must the agent pause and route the matter?
• Review: Who can inspect, challenge, revoke, or amend the delegation?

These boundaries should be enforced at the point of action. A prompt that tells an agent to act carefully cannot substitute for permissions that prevent an unauthorized transaction.

Organizations already delegate authority through roles, policies, contracts, and controls. Agentic systems add complexity because the delegated actor can interpret a request and select among actions at runtime. The action may not have been enumerated as a fixed rule, even if its scope and constraints were.

This creates a need to distinguish authority to pursue an objective from authority to make every decision that might help achieve it. “Resolve the customer issue” cannot imply permission to change contract terms, disclose restricted information, or promise a refund. The objective must be bounded by explicit action rights and stop conditions.

A system may prepare a recommendation without having authority to execute it. It may execute routine actions under defined conditions while escalating exceptions. Different levels of delegation should be tied to the consequences, reversibility, and uncertainty of the action.

Keeping a human “in the loop” is not enough if the person lacks time, context, or power to reject what the system recommends. Review must be designed so that the accountable owner can understand the evidence, see uncertainty, challenge the interpretation, and stop or reverse action where possible.

Responsibility should not collapse into a vague claim that “the human approved it.” The organization must identify who designed the delegation, who granted authority, who maintained the relevant context and policy, and who owned the underlying commitment. These responsibilities may sit with different people.

This is a governance problem as much as a technical one. Audit records can show what happened; they cannot alone establish that the authority was appropriate or that the institution has met its obligations.

People affected by an agent’s action need a path to question it. A customer should be able to challenge an incorrect decision. An employee should be able to correct a misleading record that shaped an evaluation. A domain owner should be able to contest an agent’s interpretation of policy.

Contestability requires more than an appeal form. The organization must preserve the relevant evidence and decision path, identify someone empowered to reconsider the action, and correct downstream records when needed. The process should also feed recurring errors into policy, context, and system changes.

Some actions cannot be fully reversed. A message may have been sent; information may have been disclosed; a time-sensitive opportunity may have passed. The design should account for those consequences before delegation, rather than relying on post hoc review.

A future computational enterprise will distribute authority through technical controls, operating policies, and human roles. The result could concentrate discretion in a central platform, distribute it among domain agents, or combine shared guardrails with local authority. Each choice carries consequences for power, accountability, and recourse.

Today’s systems provide components such as identity, access control, approval workflows, and audit logs. Extending these into coherent, context-sensitive delegation for agents remains an emerging design and governance challenge.

The executive question is not simply which actions to automate. It is which authority to delegate, to whom, within which jurisdiction, and with what rights of review and challenge.