You hire a recruiter and pay them per role filled. Three months later the seats are full and the people are wrong. You didn't get cheated. You got exactly what you measured, speed of fill, and not the thing you actually wanted, which was good hires who stay. That gap has a name, and a fifty-year research literature behind it.

The quick version

  • A principal-agent problem appears whenever one party (the principal) relies on another (the agent) to act for them, the two want slightly different things, and the principal can't fully see what the agent is doing.
  • You can't fix it by trusting harder. You fix it by changing three dials: incentives (what gets rewarded), information (what you can actually observe), and job design (what you ask one person to do at once).
  • Reward only what's easy to measure and people will quietly starve everything that isn't. The recruiter fills seats fast and stops caring whether the hire lasts.
  • The goal isn't perfect control, that's too expensive. It's getting the agent to want a bit more of what you want, at a cost worth paying.

The idea in depth: why delegation leaks

The formal version comes from a 1976 paper by Michael Jensen and William Meckling, "Theory of the Firm: Managerial Behavior, Agency Costs and Ownership Structure" (Journal of Financial Economics). They defined an agency relationship as any contract where a principal engages an agent to perform a service on their behalf, and pointed out that if both parties are trying to maximise their own welfare, "there is good reason to believe that the agent will not always act in the best interests of the principal."

Their lasting contribution was to put a price on that. Agency costs come in three parts: monitoring costs the principal pays to watch the agent (dashboards, reviews, audits); bonding costs the agent incurs to prove they're trustworthy (reporting, guarantees, sign-offs); and the residual loss, the value still lost because no contract is ever tight enough to close the gap completely. The useful idea here isn't that delegation is dangerous. It's that supervision is never free, and the residual loss never quite hits zero. So the move is to stop asking "how do I make sure they do exactly what I'd do?" and start asking "what's the cheapest combination of incentive and oversight that gets me close enough?"

flowchart TD
    P("Principal: wants good hires who stay") --> D("Delegates and pays per role filled")
    D --> A("Agent: wants to maximise their own payout")
    A --> M("Hidden action: you can't watch every choice")
    M --> G(["Gap: seats filled fast, quality quietly ignored"])
    G --> C("Agency cost = monitoring + bonding + residual loss")
The agency relationship and where value leaks out of it. Leaders Loop

Information is the lever: measure the right things

If the core problem is that you can't see everything the agent does, the natural question is: which things are worth seeing? Bengt Holmström answered it precisely in "Moral Hazard and Observability" (Bell Journal of Economics, 1979). His informativeness principle says a performance measure is worth building into someone's incentives only to the extent it tells you something new about the effort or choices you actually care about. A signal that's pure noise adds risk without adding alignment. (Holmström later shared the 2016 Nobel Memorial Prize in Economics, with Oliver Hart, for this strand of contract theory.)

So audit your metrics by what they reveal, not by how easy they are to pull. "Roles filled" is cheap to count but a noisy signal of "good hiring," because it says nothing about fit or retention. A better incentive ties part of the reward to something closer to the real outcome, say, hires still in the role and rated as performing at six months. It's harder to measure and it lands later, which is exactly why it carries information the fast metric doesn't.

"Reward A, while hoping for B.", Steven Kerr's 1975 line for the most common incentive mistake there is.

Steven Kerr named the failure mode in 1975 in the Academy of Management Journal: "On the Folly of Rewarding A, While Hoping for B." Leaders constantly install a metric as a convenient proxy, then act surprised when people optimise the proxy and ignore the goal it was standing in for. This connects directly to marginal thinking: an agent doesn't weigh your mission, they weigh the next unit of effort against the next unit of reward, and they push effort toward whatever pays at the margin.

Don't over-reward one task, the multitask trap

There's a sharper version of the proxy problem, and it's the one that bites real teams hardest. Holmström and Paul Milgrom set it out in "Multitask Principal-Agent Analyses" (Journal of Law, Economics, & Organization, 1991). When a job has several tasks but only some can be measured well, attaching a strong incentive to the measurable ones doesn't just boost those, it actively pulls effort away from the unmeasured ones, because the agent's time is finite.

Their conclusion is counter-intuitive and worth holding onto: when tasks compete for attention and some are hard to measure, the right strength of incentive on the measurable task can be low, sometimes a flat salary. A teacher paid hard on test scores teaches to the test and drops everything a test can't capture; a support agent paid on tickets-closed closes fast and helps badly. When a role has a hard-to-measure quality dimension, the answer is to weaken the high-powered incentive rather than chase an ever-cleverer metric, and lean on selection, professional norms and judgement-based review for the parts a number can't hold.

flowchart LR
    I("Strong bonus on the one measurable task") --> Q("Agent shifts finite effort toward it")
    Q --> W(["Unmeasured tasks get starved: quality, mentoring, care"])
    W --> R("Net result can be worse than a flat salary")
The multitask trap: a powerful incentive on a partial measure quietly defunds everything it doesn't count. Leaders Loop

One honest limitation. Agency theory models people as fairly cold calculators of self-interest, and Kathleen Eisenhardt's review, "Agency Theory: An Assessment and Review" (Academy of Management Review, 1989), is candid that this is a partial picture, it captures the incentive and information side of organisations but underweights trust, identity, intrinsic motivation and culture. Treat it as one lens, not the whole eye. People sometimes do the right thing because it's the right thing; a leader who designs as if no one ever will can corrode the very motivation that was working for free. Use the model to find where incentives are fighting you, not as a theory of all human behaviour.

A worked example: the support team that got faster and worse

A support team of twelve is missing its speed targets. The manager introduces a bonus: a meaningful payout (illustrative figures) for anyone who keeps average handle time under six minutes and closes more than forty tickets a day. Within a month the dashboard is green. Handle time drops, closes climb, the bonus pays out.

Two months later, customer satisfaction has fallen and repeat-contact rates are up. People are closing tickets fast by closing them badly, marking issues "resolved" that aren't, skipping the careful explanation that stops a customer calling back. The team optimised exactly the two things that were rewarded and quietly abandoned the one that wasn't: actually solving the problem. This is the multitask trap in miniature, and Kerr's "rewarding A while hoping for B" made flesh.

The fix isn't a better speed metric. Following the informativeness principle, the manager adds a signal that carries real information about quality, a customer-rated resolution score and the seven-day repeat-contact rate, and rebalances so the reward sits mostly on resolved-and-stayed-resolved, with speed as a guardrail rather than the prize. Per the multitask logic, they also dial the bonus down: a smaller incentive on a more honest, broader measure beats a big incentive on a narrow one. Speed recovers more slowly; the work gets genuinely better. The lesson generalises, when behaviour goes sideways, look at the scoreboard before you question the people.

Frequently asked questions

Isn't this just "people are selfish"?

No, and that misreading causes bad fixes. The problem isn't bad character, it's a structural gap between two parties' interests plus the principal's limited view. Honest, committed people still drift toward whatever the system rewards, because effort is finite and incentives are loud. You're designing around a structure, not policing a moral failing.

Why not just monitor everything more closely?

Because monitoring is a cost, not a free good, that's the heart of Jensen and Meckling's framing. Tighter surveillance buys some alignment but burns money, slows the work, and signals distrust that can erode the goodwill you relied on. Past a point, another dashboard costs more than the leakage it catches. Aim for good-enough alignment at a sane price, not control.

Do incentives have to mean money?

No. Recognition, autonomy, interesting work, promotion paths and ownership stakes are all incentives, and they often align better with hard-to-measure quality than cash bonuses do, partly because they're less easy to game on a single number. The informativeness principle applies whatever the currency: reward against signals that actually track the outcome you want.

How do I know if I have a principal-agent problem or just a performance problem?

Ask whether the unwanted behaviour is the rational response to your current rewards and measures. If a reasonable person chasing your scoreboard would do exactly what your team is doing, it's an alignment problem, change the design. If they'd have to ignore the scoreboard to behave that way, it's closer to a genuine performance or capability problem.

What's the single highest-leverage move?

Write down what you actually want, then write down what your incentives and metrics currently reward, side by side. The gap between the two columns is your principal-agent problem, made visible. Most teams have never put those two lists next to each other, and the fix is usually obvious the moment they do. It's a small example of choosing a reversible experiment over a grand redesign.

Related in the Toolkit

Where to go next