The Principal-Agent Problem Has Changed

Two Gaps in Every Delegation

The principal-agent problem begins when one person delegates work to another person whose interests differ from their own and whose actions are difficult to observe. The problem is structural rather than moral. An honest employee and a careful director still make decisions with different information, rewards, and risks from the person who hired them.

The arrangement creates an incentive gap and an information gap. A company spends money monitoring the work, the agent spends time demonstrating that they can be trusted, and some loss remains after both sides have paid those costs. Equity, bonuses, commissions, boards, audits, and performance reviews try to narrow the gap, but they cannot remove it.

How AI Changes the Incentive Side

I noticed the difference when I began delegating work to AI agents and kept using the wrong diagnosis when something came back badly. I would wonder whether the agent had rushed, avoided an awkward part, or tried to look finished. Those explanations made sense for human work, but they did not explain what had happened.

An AI agent has no career to protect. It does not manage up, delay bad news until Monday, defend last quarter's decision because its name is attached to it, or leave for a competitor. A bonus will not make it care more about the outcome. Much of the machinery used to align human incentives has nothing to act on.

That does not make an AI agent neutral or infallible. It follows objectives shaped by its training, system instructions, available tools, and the prompt in front of it. The important difference is that its errors are not usually attempts to improve its own position inside the company. They are more often failures of specification, context, capability, or verification.

Where the Failures Move

A human colleague can absorb context over months. They learn which constraints are real, why the last approach was abandoned, who the work is for, and which requests should be questioned even when nobody has documented the reason. An AI agent begins with only the context it can access. Everything I know but have not written down is missing.

This makes vague instructions more expensive. The agent may complete a poorly described task quickly and carefully, then return something that solves the wrong problem. It did not decide to provide eighty percent of the work. It acted on a brief that contained eighty percent of the information.

Review also becomes a bottleneck. The work can arrive faster than I can read it, so checking whether it exists is no longer useful. I have to check whether the assumptions were sound, the result meets the real need, and the output has introduced a failure somewhere outside the requested change.

Obedience creates another risk. A colleague may challenge a bad request because experience, professional standards, or self-preservation tells them to stop. An AI agent can challenge a request too, but I cannot assume it will notice the hidden reason that makes the instruction unsafe or pointless. The person delegating the work remains responsible for the judgment that was never included in the prompt.

Human Incentives Have Not Disappeared

AI agents still operate inside human organisations. The people selecting the model, writing the evaluation, approving the output, and reporting the result retain their own incentives. A manager can use an agent to produce more visible work instead of better work, and a company can optimise a metric that was poorly chosen. The old problem moves up a level rather than vanishing from the system.

The technology also has limits that incentives cannot fix. A model may lack current information, misunderstand an unusual domain, or produce a confident answer without adequate support. Paying for more capability can help, but no model setting replaces a clear objective, relevant context, and independent checks for important work.

Specify, Constrain, and Verify

My role in the process has become narrower and more demanding. I need to describe the desired outcome, provide the context that changes the decision, identify the constraints, and explain what evidence would count as complete. Then I need to review the result at the level of risk the task carries.

This has exposed how much of my communication used to depend on other people filling in the gaps. They could infer what I meant because they knew the history and had reasons to ask when something felt wrong. An AI agent makes those missing assumptions visible, often by following the words more closely than I intended.

The human incentive problem is smaller in this form of delegation, but the information problem is larger and the responsibility is clearer. Better tools do not remove management. They move more of it into the quality of the brief and the discipline of the review.