Agents lack Agency
· #ai · #management
Organizations scale through delegation. That sounds obvious, but delegation is often misunderstood. It is not just a way to get more tasks done. It is a way to create more places where judgment can live. A leader cannot personally notice every risk, interpret every ambiguity, remember every promise, and make every tradeoff. So organizations create roles, functions, and layers. Yes, some of that becomes bureaucracy. Sometimes a lot of it does. But the good version exists for a reason.
When we say someone is responsible for an area, we usually mean something deeper than “please complete these tasks.” We mean: understand the intent, use judgment, escalate when needed, and carry context over time. That person becomes a responsibility-bearing node in the organization. They can say, “I’ve got this,” and that sentence actually means something. Their reputation, pride, duty, and future credibility are attached to the outcome. An AI agent can produce the same sentence. But it does not own the obligation. It has agency in the narrow sense: it can act. But it lacks agency in the broader sense: it cannot carry obligation. Agents can do the work. They cannot own the work.
Humans can take on execution and responsibility together. Agents break this: they can only take on execution, never responsibility.
This is not a problem every time someone uses AI. If I use AI to do my own work faster, I am still close to the work. I still own the judgment and the outcome. That is personal leverage. The harder question is delegated execution. Organizational leverage. Agents doing work at a distance from the humans who still have to answer for the outcome.
This is where the “flatter orgs with AI agents” story starts to feel incomplete. If agents take over more execution, but cannot take responsibility, then that responsibility does not disappear. It moves, usually upward. To the human approver, the process owner, the executive, the board. Regulators, courts, and customers will still want accountability. “The agent did it” will not be a satisfying answer. Someone had a duty. Someone had control. Someone should have known. But accountability is not responsibility. Accountability is where the question lands after the fact. Responsibility is the obligation carried inside the work, before and during the fact.
Large organizations already show that broad accountability can work, but only when it is mediated through the right structures. A CEO is accountable for the company, but does not inspect every invoice, customer complaint, legal clause, or product choice. That accountability is made practical by a human responsibility network underneath: leaders, specialists, risk owners, finance controllers, legal counsel, engineering leads. These people are not just sensors and actuators. They are nodes in the system that can absorb responsibility.
Assigning responsibility is not magic. People make mistakes. They follow bad incentives. They protect turf. Sometimes they behave badly. So organizations wrap human agency in governance & controls: reviews, audits, approvals, incentives, culture. But with humans, the control system does not carry the full load. It sits on top of role expectations, professional norms, reputation, judgment, pride, fear, care, and consequence.
A control system wrapped around an AI agent is doing something else. It does not create obligation inside the agent. It only constrains behaviour from the outside. So agent governance is not simply human governance scaled up. Logs, permissions, evals and other agent control mechanisms are necessary & useful. But they do not make the agent responsible.
So the question is: what happens when organizations delegate more and more execution to agents that cannot carry obligation? Where does that responsibility go, the responsibility that made delegation practical in the first place? In the limit, do we hire humans mostly to scale responsibility, not execution? That is the problem hidden inside the phrase agents lack agency.
We may be scaling execution. We are not necessarily scaling delegation.