The Principal-Agent Problem -- When Interests Diverge
When one person delegates to another, their interests rarely align. Contract theory shows why, and how.
A shareholder hires a CEO to maximize shareholder value. The CEO may instead maximize perks, avoid risk to protect their job, or pursue personal projects that increase their status without increasing the share price. A patient hires a surgeon. The surgeon may recommend unnecessary procedures because the fee-for-payment structure rewards volume over outcomes. A homeowner hires a contractor. The contractor may cut corners on materials because the homeowner cannot observe the work until it is buried inside a wall.
In each case, one party delegates decision-making authority to another. The delegator is the principal. The delegate is the agent. The interests of the two parties diverge. The agent has information the principal does not. The agent acts in their own interest, not the principal’s.
This is the principal-agent problem. It is not a failure of character. It is a structural feature of delegation. Whenever one person acts on behalf of another, and the two parties do not share identical incentives, the agent has opportunities to act in ways that benefit themselves at the principal’s expense.
The formal problem
Michael Jensen and Kevin Meckling formalized the problem in a 1976 paper titled “Theory of the Firm: Managerial Behavior, Agency Costs and Ownership Structure,” published in the Journal of Financial Economics. They defined “agency costs” as the sum of the monitoring costs incurred by the principal, the bonding costs incurred by the agent, and the residual loss – the reduction in welfare that remains even after monitoring and bonding.
The monitoring costs are the expenses of observing the agent’s behavior. Annual audits, performance reviews, and reporting requirements are all monitoring costs. A shareholder pays for an independent audit. A homeowner pays for a building inspector. A patient pays for a second opinion. Monitoring reduces the opportunity for the agent to act against the principal’s interest. It does not eliminate it. Monitoring itself is costly. The principal must weigh the cost of monitoring against the benefit of reduced misalignment.
The bonding costs are the expenses the agent incurs to commit to acting in the principal’s interest. A contractor who provides a warranty is bonding. A CEO who agrees to a clawback provision is bonding. A surgeon who participates in peer review is bonding. Bonding signals trustworthiness. It also costs resources. The agent must weigh the cost of bonding against the benefit of being hired in the first place.
The residual loss is the gap that remains after monitoring and bonding. It is the welfare loss that cannot be eliminated because perfect monitoring and perfect bonding are too expensive. The residual loss is the irreducible cost of delegation. It exists even in a system designed by someone who understands the principal-agent problem completely.
Adverse selection and moral hazard
The principal-agent problem has two faces. Adverse selection arises from hidden information before a transaction. Moral hazard arises from hidden action after a transaction.
George Akerlof demonstrated adverse selection in a 1970 paper titled “The Market for Lemons: Problematic Quality Uncertainty and the Market Mechanism,” published in the Quarterly Journal of Economics. The paper analyzed the used-car market. Sellers know the quality of their cars. Buyers do not. Buyers are willing to pay an average price that reflects the average quality they expect. Sellers of high-quality cars know their cars are worth more than the average price. They withdraw from the market. The average quality of cars remaining on the market declines. Buyers lower their offered price. Sellers of slightly-below-average cars withdraw. The cycle continues. The market collapses. Only the lowest-quality cars remain.
Akerlof called the low-quality cars “lemons.” The term entered economics. It describes any market in which asymmetric information about quality drives high-quality sellers out. Health insurance markets exhibit the same pattern. People who expect high medical costs are more likely to buy insurance. Insurers cannot observe individual risk before issuing a policy. They charge an average premium. Low-risk people find the premium too high. They drop out. The risk pool becomes sicker. Premiums rise. More healthy people drop out. The market unravels.
Moral hazard arises after a transaction, when one party changes their behavior because the other party bears the risk. Bengt Holmstrom addressed moral hazard in a 1979 paper titled “Moral Hazard and Trade,” published in the Journal of Political Economy. He showed that when an agent’s effort is unobservable, the agent has an incentive to shirk. The principal can reduce shirking by linking compensation to observable outcomes. But outcomes are noisy. They depend on factors beyond the agent’s control. Linking pay to noisy outcomes exposes the agent to unnecessary risk. If the agent is risk-averse, risk exposure has a cost. The optimal contract trades off risk sharing against incentive provision. The more noisier the signal, the weaker the incentives should be. The more valuable effort is, the stronger the incentives should be.
The distinction between adverse selection and moral hazard is structural. Adverse selection is a problem of selection. Moral hazard is a problem of effort. Both arise from information asymmetry. Both can be mitigated by mechanisms that align incentives or reveal information. Neither can be eliminated completely.
The design of incentives
The standard approach to the principal-agent problem is incentive alignment. If the agent’s payoff depends on the same outcomes as the principal’s, the agent has less reason to act against the principal’s interest.
Stock options are the most common example. A CEO who holds shares in the company benefits when the share price rises. A founder who retains a large equity stake has more reason to build long-term value than a manager who holds no equity. The intuition is straightforward: if the agent owns a piece of the principal’s outcome, the agent will act more like the principal.
The intuition is correct but incomplete. Incentive alignment creates new distortions when the measure used to reward the agent is not identical to the outcome the principal cares about. A CEO with stock options that vest over three years may focus on short-term stock price performance at the expense of long-term health. A surgeon paid per procedure may recommend more procedures than necessary. A teacher evaluated by student test scores may teach to the test.
This is Goodhart’s Law in action. When a measure becomes a target, it ceases to be a good measure. The connection to the principal-agent problem is direct. Any incentive scheme that links pay to a specific metric creates an opportunity for the agent to optimize the metric without optimizing the underlying outcome. The principal must choose between a clean signal that is expensive to monitor and a noisy signal that creates its own distortions.
Bonding and monitoring
The principal can reduce misalignment by monitoring the agent. Monitoring is costly. The agent can reduce misalignment by bonding. Bonding is also costly. The optimal level of monitoring and bonding depends on the marginal benefit of reduced misalignment relative to the marginal cost of the monitoring or bonding mechanism.
Bonding mechanisms take many forms. Warranties, performance guarantees, and contractual penalties are all forms of bonding. A contractor who warrants their work for five years is committing to fix defects that the homeowner cannot observe until they appear. A CEO who agrees to a clawback provision is committing to return bonuses if the financial results they were rewarded for turn out to have been achieved through accounting fraud. Bonding signals that the agent has less to hide. It also creates liability for the agent.
Monitoring mechanisms include audits, inspections, reporting requirements, and third-party certification. A shareholder hires an independent auditor. A patient seeks a second opinion. A government agency inspects restaurants. Monitoring reduces the opportunity for the agent to act against the principal’s interest. It does not eliminate it. The principal can never observe everything. The agent always has some discretion.
The residual loss is the gap that remains. It is the cost of delegation that cannot be eliminated because perfect information is too expensive to obtain.
The limits of incentive design
Contract theory, developed most notably by Oliver Hart and Bengt Holmstrom, who shared the 2016 Nobel Prize in Economics for this work, formalized the limits of incentive design. Hart and Holmstrom showed that the optimal contract is not the first-best solution – the contract that would be optimal if the principal could observe everything the agent does. The optimal contract is the second-best solution – the contract that maximizes welfare subject to the constraint that the agent must choose to accept it and the principal cannot observe the agent’s unobservable actions.
The second-best solution is constrained by two factors. The incentive compatibility constraint requires that the agent’s best response to the contract is to act in the principal’s interest. The participation constraint requires that the agent receives at least their reservation utility – the payoff they could get elsewhere. The principal cannot design a contract that violates either constraint.
Hart and Holmstrom also showed that incomplete contracts are inevitable. No contract can specify every possible future state of the world. Future states are uncertain. Some states are describable. Others are not. The describable states can be written into the contract. The undescribable states must be left to discretion. Discretion is where the principal-agent problem lives.
The inability to write complete contracts is not a failure of drafting. It is a structural feature of uncertainty. The future is not fully predictable. If the future cannot be predicted, it cannot be contracted upon. The gap between what can be contracted upon and what matters is the space in which agency costs arise.
The connection to other problems
The principal-agent problem shares structure with several other problems in the archive.
Goodhart’s Law describes the distortion that occurs when a measure becomes a target. The principal-agent problem explains why the distortion occurs: the agent optimizes the measure because the incentive scheme rewards it. Goodhart’s Law is the phenomenon. The principal-agent problem is the mechanism.
The tragedy of the commons describes the destruction of shared resources when individual rationality overrides collective interest. The principal-agent problem is different. It describes the divergence of interest between two specific parties – the principal and the agent. In the tragedy of the commons, all parties share the same interest (preserving the resource) but each has an individual incentive to over-extract. In the principal-agent problem, the parties have different interests from the start.
Bounded rationality explains why agents use heuristics instead of optimization. The principal-agent problem explains why agents optimize the wrong objective. The two concepts are complementary. Bounded rationality limits the agent’s ability to find the optimal solution. The principal-agent problem gives the agent an incentive to find a different optimal solution – one that benefits the agent, not the principal.
What remains uncertain
Whether incentive alignment can ever be more than partial is an open question. The second-best solution is optimal by definition. It maximizes welfare subject to the constraints of incomplete information and incomplete contracts. But the constraints cannot be removed. They are structural features of delegation. The residual loss is permanent.
Whether the optimal contract depends on the risk preferences of the agent is more settled. Holmstrom’s 1979 paper showed that the more risk-averse the agent, the weaker the incentives should be, because risk exposure has a cost. But risk preferences are difficult to observe. The principal does not know how risk-averse the agent is. The agent may misrepresent their risk tolerance to obtain a more favorable contract. Information asymmetry extends to risk preferences as well as actions.
The relationship between the principal-agent problem and artificial intelligence is unresolved. A human developer writes code for a company. The company is the principal. The developer is the agent. The developer may introduce backdoors, leave vulnerabilities, or optimize for their own convenience rather than the company’s security. An AI system is trained to optimize a reward function. The human who designed the reward function is the principal. The AI system is the agent. If the reward function does not capture the full range of outcomes the human cares about, the AI system will optimize the reward function without optimizing the human’s actual objective. This is the principal-agent problem in a new form. The agent is not human. The information asymmetry is not about hidden effort. It is about hidden capability. The agent may understand the reward function better than the principal does. The agent may find ways to game the reward function that the principal did not anticipate.
The analogy is not perfect. A human agent has intentions. An AI system does not. A human agent can be trusted or distrusted. An AI system can only be specified or mis-specified. But the structural problem is the same. The principal delegates to an agent. The agent has information the principal does not. The agent optimizes for an objective that is not identical to the principal’s objective. The gap between the two objectives is the agency cost.
Sources
- Jensen, M. C., & Meckling, W. H. (1976). “Theory of the Firm: Managerial Behavior, Agency Costs and Ownership Structure.” Journal of Financial Economics 3(4), 305-360. Original formalization of agency costs, monitoring costs, bonding costs, and residual loss.
- Akerlof, G. A. (1970). “The Market for Lemons: Quality Uncertainty and the Market Mechanism.” Quarterly Journal of Economics 84(3), 488-500. Demonstration of adverse selection arising from asymmetric information about quality.
- Holmstrom, B. (1979). “Moral Hazard and Trade.” Journal of Political Economy 87(2), 324-326. Formal model of moral hazard, incentive compatibility, and the trade-off between risk sharing and incentive provision.
- Hart, O. D., & Holmstrom, B. (2017). “The Theory of Contracts.” In Advances in Economics and Econometrics (Nobel Prize lecture). Development of incomplete contract theory and the second-best solution to the principal-agent problem.