Agency Costs & the Alignment Toolkit

We’ve spent two lessons diagnosing the disease. Moral hazard — the agent takes a hidden action you can’t see (the contractor who cuts corners once the cheque clears). Adverse selection — the agent holds hidden information you can’t verify (the “totally reliable” used car whose transmission is a time bomb). Both grow from the same two bricks: the principal and the agent want different things, and the principal can’t fully see what the agent does or knows.

Diagnosis is satisfying. It is also useless on its own. Knowing why your CEO empire-builds doesn’t stop the empire. So this is the constructive lesson: first we put a number on the problem — what does misalignment actually cost? — and then we open the toolbox and pick up every instrument for fighting it. Each tool works. Each tool also backfires in a specific, predictable way, and knowing the backfire is the difference between an engineer and a cargo-cultist.

Agency costs

You can’t manage what you can’t measure, so start by measuring the damage. In 1976 Michael Jensen and William Meckling gave us the cleanest accounting of it ever written. They defined agency costs as the total economic loss the relationship suffers because the agent isn’t the principal — and, crucially, they split it into three buckets that sum to the whole.

• Monitoring costs — what the principal spends watching the agent. Audits, boards of directors, dashboards, KPI reviews, the manager who walks the floor. Every euro spent verifying the agent’s behaviour lands here.
• Bonding costs — what the agent spends credibly committing to behave. This one surprises people: the agent pays too. Publishing audited accounts, offering a warranty, accepting a clawback clause, agreeing to a non-compete. The agent burns real resources to prove they’ll behave, because a promise nobody can verify is worth nothing.
• Residual loss — the value still lost after all that monitoring and bonding, because alignment is never perfect. Even with a great contract, the agent’s choices won’t be exactly what a perfectly-informed principal would have picked. The gap between “what the agent actually did” and “what the principal would have wanted if they could see everything” — measured in euros — is the residual loss.

Agency cost = monitoring + bonding + residual loss. That’s the whole bill.

Worked example: you versus your contractor

You hire a builder to renovate a kitchen for a fixed €40,000. Left completely unwatched, the builder would rationally use cheaper materials and rush the finish, delivering work worth (to you) about €31,000 — a €9,000 gap between the ideal job and the coasting job. Here’s how you fight it, and what each move costs.

Note what happened. The raw misalignment was €9,000. By spending €3,500 on monitoring and bonding, you clawed most of it back — the residual loss fell to €2,000. Total damage dropped from €9,000 to €5,500. That’s a win.

But look hard at that last row. The residual loss is still there. You spent real money to shrink the problem and you did shrink it — from €9,000 to €5,500 — but you did not, and could not, drive it to zero. Push monitoring harder and the surveyor’s bill climbs faster than the corners it catches. There is an optimum amount to spend, and past it you’re burning euros to save cents.

The alignment toolkit

Agency costs are the disease’s price tag. Now the treatments. Every tool below attacks one of the two bricks — it either makes the agent want what you want (fixing the goal conflict) or lets you see what the agent does or knows (fixing the information gap). And every tool has a matching failure mode. Here they are, one per section.

Pay-for-performance & equity

What it does. Stop paying for time and start paying for outcomes. Tie the agent’s pay to the thing you actually care about, or hand them equity so they own a slice of the upside and feel your gains as their gains. This is the purest form of skin in the game: make the agent a mini-principal.

Worked example. A SaaS company pays its account executives a €50,000 base plus 8% of net-new recurring revenue they close. An AE who lands €600,000 of new contracts earns €50,000 + €48,000 = €98,000. Suddenly the AE wants the same thing the company wants — signed, retained revenue — with no manager standing over them. Founders get equity for the same reason: give them 15% of the company and a €10M exit is €1.5M in their pocket, so they grind like it’s theirs, because it is.

How it backfires. You are now paying for a proxy, and — as the incentives lesson hammered — the agent optimises the proxy, not your true goal. Goodhart’s law: when a measure becomes a target, it stops measuring what you meant. Pay for closed revenue and the AE stuffs the pipeline with deals that churn in three months. Pay a teacher on test scores and you get teaching-to-the-test, not learning. Worse, equity and stock options are convex: options pay off big if the stock soars and are merely worthless (not negative) if it craters. That one-sided payoff rewards recklessness — the agent is tempted to bet the company on a moonshot, because heads they’re rich and tails they’re just… unemployed, same as before. High-powered pay doesn’t just risk gaming; it risks the agent taking swings you would have vetoed.

When to reach for it. When output is measurable and hard to game, when you want the agent bearing real risk (they’re less risk-averse than you, or closest to the information), and when the upside you share is genuinely the upside you want. The more gameable the metric, the lower-powered you should keep the pay.

Monitoring & audits

What it does. Shrink the hidden in “hidden action.” Require reporting, run audits, seat a board, send inspectors, install the dashboard. If you can observe more of what the agent does, the goal conflict matters less — because now shirking gets caught.

Worked example. A restaurant chain can’t watch every line cook, so it sends mystery diners four times a quarter and audits food-cost ratios monthly. A branch quietly over-portioning to hit its own bonus shows up as an anomalous cost ratio; a branch cutting hygiene corners gets flagged by the mystery diner. The possibility of being seen changes behaviour even on the un-inspected days.

How it backfires. Monitoring is costly and never complete — it’s a monitoring cost in Jensen–Meckling for a reason, and no amount of it catches everything (the mystery diner visits four times out of ninety). And there’s a subtler poison: over-monitoring signals distrust and can crowd out goodwill. Bolt a keystroke logger onto a salaried professional and you may convert someone who was working hard from intrinsic pride into someone who does the literal minimum the logger measures — you’ve replaced “I care” with “I’m being watched,” and the second is weaker. (We devote lesson 6 to this crowding-out effect; it’s real enough to invert the whole tool.)

When to reach for it. When actions are observable at reasonable cost, when the stakes justify the surveillance, and when you can monitor outputs or outcomes rather than micromanaging effort — watching results bruises trust far less than watching keystrokes.

Screening & signalling

What it does. These two attack the other brick — hidden information (adverse selection) — by making the different agent types sort themselves, exactly as previewed in lesson 3.

• Screening is the uninformed party (principal) designing choices so the agent’s pick reveals their type. Offer a menu; watch which one they grab.
• Signalling is the informed party (agent) taking a costly action a bad type couldn’t profitably fake, to prove they’re a good type.

Worked example. Screening: an insurer offers Plan A (low premium, €1,000 deductible) and Plan B (high premium, €100 deductible). Careful, low-risk drivers self-select into A — they rarely claim, so a fat deductible barely stings — while accident-prone drivers reveal themselves by reaching for B. The deductible screens the pool without the insurer reading a single mind. A probation period is the same trick for hiring: bad fits quit or wash out before you’re committed. Signalling: a degree, a professional certification, or a real warranty works because it’s cheap for the good type and painfully expensive for the bad type — a manufacturer who knows the product is junk can’t afford a 10-year warranty, so offering one credibly signals quality.

How it backfires. The costs are real and often wasted. A signal only works if it’s expensive, which means society burns resources on it — credential inflation is the poster child: when everyone signals with a degree, the bar creeps to a master’s, and we’ve spent years of tuition to convey the same one bit of information. And screens can be gamed: a savvy bad type learns which choice the good types make and mimics it, quietly poisoning the pool the screen was meant to purify.

When to reach for it. When the problem is hidden information, not hidden action, when there exists an action whose cost genuinely differs by type (that’s what makes a signal separate), and when the sorting is worth more than the resources the signal burns.

Reputation & repeated dealing

What it does. Stop treating the relationship as one-shot. When you’ll deal again, the shadow of the future disciplines the agent all by itself: cheat today and you forfeit every profitable tomorrow. This is precisely the evolution-of-cooperation result — in a repeated game, cooperation can be a best response, sustained not by kindness but by the credible threat of losing the relationship. Reputation is that threat made portable.

Worked example. A self-employed plumber lives and dies by referrals. Overcharging one customer nets an extra €200 today but poisons the word-of-mouth that brings ten future jobs worth €400 each in profit — €4,000 forfeited to grab €200. The arithmetic disciplines them where no contract could. Brands scale the same logic: a hotel chain posts an enormous, sunk reputation bond in its name. One filthy room risked across thousands of properties threatens the whole brand’s value, so headquarters wants to police quality — the reputation bond aligns them without you monitoring a thing.

How it backfires. Reputation runs on the future, so anything that shortens the future collapses it — the dreaded end-game effect. The employee who’s retiring next month, the founder about to sell, the firm circling bankruptcy, the last period of any finite game: when there’s no tomorrow to protect, the discipline evaporates and the agent cashes in. And reputation is asymmetric in time — slow and expensive to build, catastrophically fast to spend. A thirty-year name can be liquidated in one scandal, and a sufficiently patient bad actor will build trust for years precisely to exploit it once at the end.

When to reach for it. When the relationship genuinely repeats with no clear last round, when the future’s value to the agent outweighs the one-shot temptation, and when reputations are observable and transferable (referrals, reviews, a brand) so cheating actually gets punished by the market.

Efficiency wages

What it does. Pay the agent above the market rate — deliberately more than they could get elsewhere — so the job itself becomes valuable to keep. Now the threat of being fired has teeth: get caught shirking, lose the job, and fall to a worse-paid alternative. This is the Shapiro–Stiglitz insight — a wage premium buys effort by making dismissal costly to the worker.

Worked example. In 1914 Ford doubled pay to the famous \$5 day — roughly double the going rate. Turnover, which had been brutal, cratered; workers guarded jobs that now paid far more than anything down the road, and effort and quality rose. The premium wasn’t charity; it was an incentive device. A job worth losing is a job worth doing well.

How it backfires. It’s expensive by design — you’re voluntarily overpaying every worker, including the ones who’d have worked hard for less. Paid economy-wide, above-market wages create involuntary unemployment: if everyone pays a premium, the market can’t clear and a queue of willing workers forms outside (that’s the mechanism, not a bug). And here’s the sting — efficiency wages only bite if you can detect shirking well enough to fire for it. If the hidden action stays hidden, the threat of dismissal is empty and you’re just… overpaying. It complements monitoring; it doesn’t replace it.

When to reach for it. When effort is hard to specify in a contract but shirking is at least occasionally detectable, when turnover or training costs are high (so retention pays for the premium), and when the threat of job loss is a meaningful stick — i.e. the outside option really is worse.

No single tool is enough

Notice that every treatment above cured one failure and opened another. That’s not bad luck — it’s structural. So real designs never lean on one tool. They stack them, letting each cover the others’ backfires:

• A base salary to insure a risk-averse agent against pure luck (so you don’t have to pay a giant risk premium)…
• …plus some performance pay for effort (enough to pull effort, not so much it invites gaming)…
• …plus monitoring of a few outcomes that are cheap to observe…
• …plus the discipline of reputation and repeat business…
• …and maybe a wage premium so the whole package is worth keeping.

A real executive contract is exactly this cocktail: base salary + bonus on measured targets + restricted stock (with vesting and clawbacks to blunt the convexity) + a board that monitors + the reputational stakes of the labour market. But — and this is the catch that runs through the whole lesson — each tool you add has its own cost. More performance pay adds gaming and risk premium. More monitoring adds surveillance cost and corrodes trust. So you can’t max out every dial; you trade them off. The optimal design lives at an interior optimum — not zero of anything, not max of anything, but a balanced mix where the marginal benefit of pushing any one dial harder equals its marginal cost. Let’s see that shape directly.

Effort versus distortion: find the interior optimum

The island below is the whole lesson made tactile. It plots the principal’s net payoff across every possible pay mix, from pure salary on the left to pure commission on the right. Drag the top slider to change the mix. Watch the tension: slide left and the agent coasts (low effort, low payoff); slide right and effort rises — but so do gaming, risk-taking, and the risk premium a risk-averse agent charges to bear pay that swings with luck. Those convex costs eventually overwhelm the concave gains in effort, and net payoff falls back down — often below what a flat salary would have paid you. The peak is in the interior: neither extreme wins.

Now drag the second slider — the agent’s risk aversion — and watch the sweet spot move. A more risk-averse agent demands a bigger premium to accept luck-swung pay, so the optimum slides toward a flatter, lower-powered contract: less performance pay, more salary. That’s the risk/incentive trade-off in one gesture — the more of your firm’s risk you dump on a risk-averse agent, the more you have to pay them to bear it, and past some point that premium isn’t worth the extra effort you buy. It is a genuine trade-off, not a free lunch: you can’t have full effort and cheap risk-bearing at once. (We’ve only shown it here as a shape; the formal proof that the optimal contract steepens as the agent gets less risk-averse — and flattens as noise rises — is lesson 5.)

Recap

We put a price on misalignment (Jensen–Meckling’s monitoring + bonding + residual loss), learned that the residual loss never reaches zero, and then walked the whole toolkit — pay-for-performance and equity, monitoring and audits, screening and signalling, reputation and repeated dealing, and efficiency wages — cataloguing what each fixes and exactly how each one bites back. The punchline: no single tool is enough, real designs stack them, and because every dial has its own rising cost, the best contract lives at an interior optimum, not at any extreme.

Every tool we picked up left a residual loss behind — and we kept insisting it can never be driven to zero. Why never? Next up, lesson 5, “The Impossible Perfect Contract”, proves it: there is no contract that gets full effort, cheap risk-bearing, and perfect honesty all at once.