Where the Model Lies

Five lessons ago you were handed a tool that cuts almost anything: the principal–agent lens. It has since sliced through taxis, fund managers, insurance markets, CEO pay and mechanism design. A tool that sharp deserves a safety briefing before you leave the workshop — because this one has a caricature filed into its very edge, and if you don’t know it’s there, the tool will bite the hand holding it.

The caricature is the agent itself. All the way through, we drew the agent as a single, clean cartoon: a pure, self-interested money-maximiser — a creature that will shirk the instant no one is watching, lie whenever the lie pays, and respond to nothing but the size of the cheque. That cartoon is useful. It’s a worst-case stress test: design a contract that survives a pure money-maximiser and it’ll survive most real people. But it is a cartoon, and real agents are stranger and, mostly, better. They have professions and pride and a conscience and a reputation to protect. They sometimes do the right thing precisely when no one is watching — which is the one thing the cartoon swears is impossible.

This lesson is where we tell you where the model lies — the places its assumptions quietly fail, and the ways that forgetting so makes your contracts backfire. Then we’ll put the whole course back together in one breath. You’ve done the intro and lessons 1 through 5; keep the incentives course (intrinsic vs extrinsic motivation, crowding-out, the daycare fine) and mechanism-design within reach, because this is where they come collect their debts.

Money is not the only motive

The analogy. Ask a violinist why she practises six hours a day and she will not say “the marginal wage.” Ask a Wikipedia editor why they fixed a typo at 2 a.m. for zero pay and zero credit, and “incentives” isn’t the answer either. Some engines run on a fuel the model never put in the tank.

The naive model’s mistake. The pure-money-maximiser assumption says an agent supplies effort only up to the point where the marginal cheque justifies it — so unwatched, unpaid effort should be zero. Slack is the default; diligence has to be bought. Under that assumption, an unmonitored agent is a shirking agent, full stop.

The corrected view. Real agents run on intrinsic motivation — doing a thing because the doing itself is rewarding — alongside the extrinsic cheque. And they’re embedded in professional norms: internalised standards (“a real engineer doesn’t sign off on a bridge they wouldn’t drive their family over”) that bind behaviour even when no principal is watching and no bonus is attached. Craft, mission, identity, reputation among peers — these do real agency-work for free. The distinction (intrinsic vs extrinsic) is exactly the one the incentives course drew; here it rewrites your risk map.

A worked example. Line up three agents the cartoon says should shirk, and watch them not:

None of these is buyable by a bonus, and none needs a monitor. The nurse is not leashed by surveillance; she is held by who she takes herself to be.

The implication. Sometimes the standing incentive isn’t pulling against you at all — the agent’s own motives already point where you want to go. In those cases the entire agency apparatus (bonuses, monitors, tripwires) is not just unnecessary; as the next section shows, it can be actively destructive.

Crowding-out: paying more can get you LESS

The analogy. Imagine you cook dinner for a friend every week because you love them. Then they start leaving €20 on the counter each time “to be fair.” Notice what dies. The dinner was a gift; now it’s a transaction. Priced, it becomes worse — and if they ever stop paying, you may not go back to cooking for free, because the meaning has been drained out.

The naive model’s mistake. The cartoon treats money and motivation as additive: intrinsic motivation is whatever effort you’d supply for free, and every euro of extrinsic pay simply stacks on top. More pay, more effort, always, monotonically. Under that view a performance bonus can only help.

The corrected view. Motivation crowding theory (Frey, Deci–Ryan and others) says extrinsic and intrinsic motivation are not additive — they interact, and often destructively. Bolting a price onto intrinsically motivated work can crowd out the very motive that was doing the work, for two reasons: paying for something signals that it’s a transaction, not a duty (so the duty evaporates), and being paid to do something you chose freely can feel like a loss of self-determination (so the internal drive deflates). The net effect can be negative: pay goes up, performance goes down.

A worked example — three exhibits.

The daycare case is the cleanest, and it ties straight back to incentives: a fine is just a negative incentive, and here the incentive crowded out the norm it was meant to reinforce. Worse, morals proved sticky-when-removed — take the price away and the transaction framing stays, so you can’t simply undo the damage.

The implication. Pay-for-performance can lower performance whenever the pay signals distrust (“we’re bonusing you because we assume you’d slack otherwise”) or converts a relationship into a transaction (care work, teaching, public service, community). The money-maximiser model can’t even represent this failure — it has no variable for the motive that got crushed. That blind spot is exactly where well-meaning bonus schemes go to die.

Strong incentives amplify measurement error and invite gaming

The analogy. A powerful incentive is a loudspeaker wired to a microphone. Wire it to a clear signal and you amplify the signal. Wire it to a noisy, hackable proxy and you amplify the noise and the feedback squeal — deafeningly. The power of the incentive is neutral; what it amplifies is whatever you pointed it at.

The naive model’s mistake. The cartoon says a stronger incentive is a better incentive — high-powered pay-for-performance drives more of the desired outcome, so crank it up. This ignores two things lessons 4 and 5 hammered: your incentive is almost never attached to the true outcome, only to a proxy, and stronger incentives make agents work the proxy harder — including its errors and its exploits.

The corrected view. Two forces turn a high-powered incentive on a lossy proxy into a magnifier of everything wrong with the proxy:

• Goodhart’s law (again). When a measure becomes a target, it ceases to be a good measure. The harder you push a metric, the harder agents optimise the metric instead of the goal it stood for. A weak incentive on a bad proxy leaks; a strong one bursts.
• Multitasking. When a job has measured and unmeasured dimensions, a strong incentive on the measured part pulls effort away from the unmeasured part. High power doesn’t just fail to help the unmeasured dimension — it actively starves it.

So the strength of the incentive multiplies both the noise (random measurement error gets amplified into big, unfair payoff swings that demoralise and distort) and the manipulation (the return to gaming rises with the stakes). A high-powered incentive on a proxy is a magnifier held over every flaw in that proxy.

A worked example.

Wells Fargo is the textbook horror: leadership set a ferocious cross-sell target (“eight products per household”), pushed it with high-powered pressure, and the proxy — accounts opened — got maximised in the only way left, by fabricating accounts. The incentive worked perfectly on the metric and catastrophically on the goal.

The implication. Before you raise the power of an incentive, ask how good the proxy is. If the proxy is a clean, cheap, hard-to-game measure of the true goal, high power is your friend. If it’s lossy or hackable — and most interesting outcomes’ proxies are — high power is a mistake, and often a low-powered incentive (a salary, a satisficing target, professional judgement) does less damage. Strength is not a virtue on its own; it’s a multiplier, and multiplying a flawed proxy just gives you more flaw.

Monitoring and trust are not free

The analogy. You could stop your teenager from ever missing curfew by GPS-tracking their phone, reading their texts, and installing a camera in the car. It would work. It would also cost you a fortune in gadgets, a mountain of your own time reviewing footage, and — the expensive part — a child who now sees you as a warden and behaves like a prisoner: doing exactly the letter of the rules, nothing more, and looking for the blind spot. The surveillance “solved” shirking and bought you something worse.

The naive model’s mistake. Confronted with hidden action, the cartoon’s reflex fix is “just monitor them” — as if observation were free, complete, and side-effect-free. If you can watch, you can pay for the observed action, and the problem dissolves. Two lessons ago that was even the tidy escape route: buy observation, slide the delegation up out of the danger cell.

The corrected view. Monitoring is a cost, not a free lever, and it fails in three distinct ways:

• It’s expensive. Auditors, inspectors, software, dashboards, and above all your own attention — monitoring consumes the very resources you delegated to save. Past some point, the monitoring costs more than the shirking it prevents.
• It’s incomplete. You can watch what’s cheap to watch, which is rarely the thing that matters (you monitor hours logged, not judgement exercised). This is the multitasking trap wearing a surveillance hat: watched dimensions get done, unwatched ones get dropped.
• It’s corrosive. Surveillance signals distrust, which crowds out goodwill (the same crowding-out mechanism as the cash bonus) and provokes work-to-rule — the malicious-compliance move where agents do exactly what’s monitored and not one thing more, killing the discretionary extra effort that made them valuable.

A worked example. Two shops, same job:

The efficient amount of monitoring is frequently low. Trust — built not on hope but on selection (hire agents whose intrinsic motives and norms already align) and reputation (repeated dealings where a good name is an asset the agent won’t burn) — is often the cheaper technology. The cheapest monitor is an agent who monitors themselves because their profession or their reputation demands it.

The implication. “Just add oversight” is a beginning, never a free ending. Every monitor is itself a cost and a new agent (recall quis custodiet ipsos custodes). The right question isn’t “how do I watch them harder?” but “what mix of selection, reputation, light monitoring, and not-crushing-their-goodwill delivers the outcome cheapest?” Sometimes the answer is more monitoring. Surprisingly often, it’s less.

The over-cynicism trap

The analogy. Point a metal detector at a beach and it beeps at bottle caps, foil, and buried treasure alike — everything reads as metal. Point the agency lens at every human relationship and everyone reads as a shirking crook waiting to be caught. The instrument is real; the reading is a distortion of your own making.

The naive model’s mistake. The subtlest failure isn’t in any one contract — it’s in what the model does to the designer. Steeped in principal–agent thinking, you can start treating the pure-money-maximiser cartoon as literal truth about everyone, and use it as a licence to assume the worst: everyone is a crook, every relationship needs airtight incentives and heavy surveillance, trust is for suckers.

The corrected view. That stance is self-fulfilling. Design a system that treats every agent as a presumptive thief — bristling with monitors, tripwires, and distrustful bonus schemes — and you get three compounding harms: you crowd out the intrinsic motivation and goodwill of the good agents (who now feel insulted and disengage), you select for exactly the transactional, game-the-system agents you feared (the norm-driven ones leave for somewhere they’re trusted), and you signal that this is a low-trust, every-agent-for-themselves environment — which teaches everyone to behave that way. You manufacture the crooks you assumed. Meanwhile, high-trust equilibria are more productive: they run on cheap self-monitoring, discretionary effort, and low overhead — the whole expensive agency apparatus gets to stand down.

A worked example. Two companies hire from the same labour pool:

The implication — the mature use of the model. Hold two truths at once: incentives matter (never design as if they don’t — the cartoon’s stress test is real and adverse selection and moral hazard are real), and many agents are partly intrinsically motivated and norm-bound (never design as if they aren’t). The craft is to build systems that guard against the bad motives without crushing the good ones — verify enough to keep the honest honest and the free-riders uncomfortable, but not so much that you insult your best people into leaving or your relationships into transactions. The model is a scalpel, not a worldview. Diagnose with it; don’t let it diagnose your soul.

The whole model, in one breath

Here’s the entire course, exhaled at once. Delegation creates a principal (who wants an outcome) and an agent (who acts for them), and between them stands a wall of two bricks: misaligned interests and information asymmetry — remove either and the wall falls. That asymmetry splits by timing into hidden action (behaviour after the deal → moral hazard) and hidden information (type before the deal → adverse selection), and the friction they generate — the shirking, the lemons, the distortions, plus everything you spend fighting them — is the agency cost you’re trying to minimise. Against it you wield a toolkit: align interests (skin in the game, ownership, performance pay), pierce the information gap (monitoring, screening, signalling, warranties), and design the rules so honest behaviour is the agent’s own best move (mechanism design). But there’s no perfect contract — the risk–incentive trade-off means you can’t load an agent with outcome-risk they don’t control without overpaying for it, Goodhart and multitasking mean every proxy leaks, and the residual gap never closes to zero. And finally, the model lies about the agent: real agents have intrinsic motives you can crowd out by paying, norms and trust that do work no incentive can, and they live in systems where monitoring is costly and cynicism is self-fulfilling — so the mature designer guards against the bad motives without crushing the good ones. That’s the whole machine.

Recap

• The agency model has a caricature at its core — the agent as a pure, self-interested money-maximiser. It’s a great stress test and a poor portrait. Mistaking one for the other makes contracts backfire.
• Money is not the only motive. Intrinsic motivation, professional norms, craft, mission and identity make many agents do the right thing unwatched. Sometimes the standing incentive isn’t pulling against you at all.
• Crowding-out means paying more can get you less: pricing intrinsically motivated work can destroy the motive (the daycare fine, paid blood donation). Pay-for-performance lowers performance when it signals distrust or turns a relationship into a transaction.
• Strong incentives amplify whatever the proxy gets wrong — Goodhart plus multitasking turn a high-powered incentive on a lossy metric into a magnifier of noise and gaming (Wells Fargo’s fake accounts, test-score scandals).
• Monitoring and trust aren’t free. Surveillance is costly, incomplete, and corrosive — it crowds out goodwill and provokes work-to-rule. Often the efficient monitoring level is low, and trust built on selection + reputation is cheaper.
• The over-cynicism trap: using the model as a licence to assume everyone’s a crook builds low-trust systems that perform worse (self-fulfilling). The mature use holds both truths — incentives matter and many agents are partly intrinsically motivated — and guards the bad motives without crushing the good.
• In one breath: principal/agent gap → hidden action + hidden information → agency costs → toolkit → no perfect contract → these human limits.

That’s the whole model, caveats and all. One thing remains — the Final Exam: a single graded run across everything you’ve learned. No Back, no retry. Take a breath, then begin.