The new delegation

AI agents are moving from answering questions to taking action. They can research a market, prepare a contract, change software, coordinate a project, and keep working after the original conversation has ended. This is a profound increase in leverage. It is also a change in the location of responsibility.

Most organizations still treat AI as a faster interface. The more accurate frame is delegation: a principal gives an agent a goal, information, tools, and some range of discretion. Once we see the system this way, the important question is no longer whether the model is intelligent. The question is who has the authority to decide, act, and accept the consequences.

Authority is a system

Human authority is not preserved by inserting an approval button at the end of every workflow. A tired person clicking approve on an opaque proposal is not meaningful oversight. Authority requires context: the purpose of the action, the evidence behind it, the constraints that apply, the alternatives that were rejected, and the consequences of being wrong.

The system therefore needs explicit boundaries. Some actions can be observed, some can be prepared, some can be executed reversibly, and some require a named person to authorize them. These levels should be designed around consequence, not around how technically difficult an action appears.

Consider an agent that prepares a contract change overnight. In the morning a manager receives a summary, a recommendation, and an approve button. The next meeting is already late; the exception clauses are collapsed; the rejected alternative is not shown. Clicking approve looks like oversight. It is a signature on a decision whose purpose, evidence, and reversibility were never inspected. Authority failed before the button was pressed.

A decision contract

A useful agent workflow begins with a compact decision contract: the outcome being pursued, the evidence that counts, the constraints that cannot be crossed, the person who owns the decision, and the proof required to call the work complete. What counts as evidence is already a decision about meaning. Meaning Before Automation is the prior distinction: if that meaning stays implicit, the contract authorizes a confusion.

This contract gives the agent room to move while keeping the human role substantive. The agent can search broadly, compare options, produce artifacts, test assumptions, and surface conflicts. The human remains responsible for the meaning of the goal and for the irreversible choices made in its name.

The opposite failure is as real. If every reversible draft, search, and comparison must wait for a person, attention is spent on motion rather than on irreversible choice. People click through queues they cannot actually read. The aim of a decision contract is not more gates. It is fewer, named authorizations where consequence is high, and more room for the agent where a mistake can still be undone.

Learning, not just control

Good governance should not merely prevent failure. It should improve the next decision. Every important action should leave behind a trace: what was believed, what was approved, what happened, and what the organization learned. Without that loop, oversight becomes bureaucracy and automation repeats yesterday’s misunderstandings at greater speed.

The objective is not to keep humans in every mechanical step. It is to keep humans in the places where purpose, legitimacy, judgment, and accountability are real. AI proposes. People approve, execute, and learn. Over time, both the people and the system become better at knowing which is which. Building Companies That Learn is the test of that loop: the next decision must be able to change.

Back to writing