Everything until now has been a tour of how incentives go wrong: behavior follows the reward (lesson 1), people sincerely rationalize whatever pays them (lesson 2), agents act on their own payoff, not yours (lesson 3), rewarding a proxy manufactures the gaming you feared (lesson 4), and money can crowd out the very motive doing the work (lesson 5). That’s a lot of cautionary tales. This lesson is the one where we stop diagnosing wreckage and start building — turning all of it into a small set of design rules you can apply on purpose, before a reward ever ships. By the end you’ll be able to take a broken incentive and repair it, not just name what’s broken.
The good news: there’s really one master move, and the rest are tools for getting it right. Let’s lock in the shape before we unpack it.
Before you read — take a guess
You want a sales team to build healthy, lasting customer relationships, but you can only realistically pay them a bonus on something measurable. Which design is least likely to backfire?
Reward the outcome, not the proxy
This is the master principle, and you already met its silhouette in lesson 4 — the slider’s fix, the toggle that made the whole rat-catching population stay honest no matter how high the bounty climbed. Here it is stated plainly. Reward the outcome, not the proxy. Aim the reward at the real goal so that the self-interested path and the desired path are the same path. When they coincide, you no longer have to fight human nature, hope for virtue, or police every loophole — chasing the reward is achieving the goal. The cobra effect happens because the cheap way to win the reward differs from the goal; this principle closes that gap by construction.
The analogy. A proxy reward is like steering a car by watching the speedometer instead of the road. The needle correlates with progress — until you optimize the needle directly, at which point you’ll happily floor it straight into a wall and call it a personal best. Reward the destination reached, and suddenly the driver wants to watch the road.
Worked example — the FedEx night shift. This is the clean positive case, the mirror image of the cobra story. FedEx’s overnight operation depends on a sorting crew getting every package routed before the planes leave. For a long time the crew was paid by the hour — and the sort kept finishing late. Why? Look at the payoff. Hourly pay rewards time spent, not work done. A worker who sorted fast and finished early simply earned less; lingering paid better. The incentive was quietly pulling against the goal. The fix was almost embarrassingly simple: pay per shift instead — finish the sort correctly, and you go home (paid in full). Suddenly finishing fast was the worker’s own goal, not a sacrifice. The chronic lateness reportedly vanished. Nobody’s character changed; the reward stopped pointing at “hours logged” and started pointing at “the sort is done.”
| Scheme | What it rewards | What the worker wants | Result |
|---|---|---|---|
| Pay by the hour | Time spent on the clock | Stretch the job out — speed cuts my pay | Sort finishes late |
| Pay per completed shift | The outcome (sort done, planes loaded) | Finish correctly and leave | Sort finishes on time |
Same people, same packages — one column flips the whole result, because the outcome is now what pays.
The pitfall. The trap that produces proxy rewards is seductive and specific: you reward the convenient, measurable proxy because the real outcome is harder to measure. “Hours” is trivial to count; “the sort is reliably done well” takes more thought to define and verify. Defaulting to the easy-to-measure number feels responsible — you’re being data-driven! — but you’ve just handed people a target that isn’t the goal. The discipline is to spend the extra effort defining and measuring something closer to the outcome, even when a lazier proxy is sitting right there.
The proxy temptation, named
The real goal is usually harder to measure than its proxy — that’s why perverse incentives are so common. Convenience pulls you toward rewarding the easy number (hours, tickets, accounts, tails). Resist it. A rougher measure of the right thing beats a precise measure of the wrong thing every time.
When to use it
Always — it’s the first question for any reward you design or critique: what outcome do I actually want, and does this reward land only when that outcome happens? It’s most urgent when the reward is large (a big reward on a proxy drags everyone toward the gap fast) and when the proxy is far from the goal (counting activity — calls, hours, lines of code — rather than results). If you can restate the reward as “you get paid when [the real goal] occurs,” you’ve passed.
Align the agent with the principal — skin in the game
Recall the principal–agent problem from lesson 3: someone acts on your behalf — a broker, a contractor, an employee, a fund manager — but their payoff doesn’t match your goal, so their effort and advice bend toward their reward, not your result. The design fix has a name borrowed from Nassim Taleb: give the agent skin in the game — make them share in your outcome, so that when you win they win and when you lose they lose too.
The analogy. You want the chef to care about the meal as much as you do? Have the chef eat the same dinner. “Eat your own cooking” isn’t a folksy nicety — it’s incentive design. The instant the agent’s plate is the same as yours, you stop needing to inspect every dish.
How alignment is built. The concrete tools all do the same thing — they couple the agent’s payoff to the principal’s outcome:
| Misaligned (agent’s payoff ≠ your goal) | Aligned (shared outcome) |
|---|---|
| Salesperson paid on contracts signed | Paid on revenue retained (the relationship must last) |
| Surgeon/clinic paid per procedure performed | Paid on patient outcomes — “value-based care” |
| Fund manager paid a flat % of assets, win or lose | Manager invests their own money in the same fund |
| Contractor paid more to replace than to repair | Fixed-price bid, or a stake in the building’s long-run cost |
Worked example — the fund manager. A manager paid a flat 2% of assets under management earns the same whether your money grows or shrinks; their incentive is to gather assets, not to grow yours. Now require the manager to hold a large chunk of their own net worth in the very fund they run. Every bad bet now costs them personally, in the same direction it costs you. Their sincere judgment about risk — recall incentive-caused bias from lesson 2 — quietly realigns, because now understanding the risk protects their own money, not just yours.
The pitfall — asymmetric risk. The dangerous half-measure is letting the agent keep the upside while offloading the downside onto the principal. A trader who pockets a fat bonus on winning bets but loses only their job (not their savings) on a blowup has skin in the game on one side only — so they’re incentivized to take wild risks with your money. That’s “heads I win, tails you lose,” and it caused more than one financial crisis. Real alignment means the agent shares the loss, not just the gain.
When to use it
Whenever someone acts on your behalf and you can’t watch every move (which is almost always). The bigger the information gap between you and the agent — you can’t evaluate the surgery, the trade, the repair — the more you must rely on alignment rather than supervision. Ask: does this person lose when I lose? If the honest answer is “no, they get paid either way,” you have a principal–agent gap to close.
Pre-mortem the gaming — and keep it simple
You can’t always reward the perfect outcome, and even aligned agents find loopholes. So before any reward ships, run it through the pre-mortem — the design-time stress test you met in lesson 4. Ask, out loud, in writing: “How would a clever, lazy, slightly dishonest person maximize this reward without doing the real work?” Then assume someone will actually do it. The pre-mortem catches perverse incentives on the drawing board, which is the cheapest possible place to find them — infinitely cheaper than discovering them after you’ve trained a thousand people to game you.
The analogy. It’s the same move security engineers make: don’t ship the lock and wait to get robbed — hire someone to try to pick it first. You play the attacker against your own incentive before reality does.
Worked example — a pre-mortem catching a bug on paper. Say a hospital wants to reward surgeons for good results and proposes paying a bonus on each surgeon’s risk-adjusted survival rate. Sounds rigorous. Now run the pre-mortem: how would a clever, lazy, slightly dishonest surgeon maximize this number without saving more lives? Answer in five seconds: refuse to operate on the sickest patients. Turn away the hard cases and your survival rate climbs while your actual contribution falls — and the patients you reject may die elsewhere, invisibly. You’ve found the cobra on paper, before a single patient was turned away. The fix (a guardrail counting patients declined, or rewarding lives saved against expected baseline) is cheap to add now and impossible to retrofit after the gaming is entrenched.
Keep the formula simple. A second, underrated rule: keep incentive formulas simple and transparent. A complex, multi-factor bonus formula — weight this metric 30%, that one 25%, adjust for a quality modifier and a stretch multiplier — is a bigger attack surface, not a safer one. Every factor is a new dial to game, and the interactions between factors produce loopholes nobody foresaw (and nobody can even see, because the formula is too tangled to reason about). Complexity doesn’t make a reward gaming-proof; it just hides the gaming.
Why complexity backfires
People treat a baroque bonus formula as a puzzle to solve, and they will solve it — finding the cheapest combination of inputs that maxes the payout, goal be damned. A reward you can’t explain in one sentence is a reward you can’t reason about, which means you can’t predict how it’ll be gamed. Simple beats clever.
A team designs an employee bonus with eight weighted metrics, a quality modifier, and a stretch multiplier, reasoning that 'with so many factors, no single one can be gamed.' Why is this reasoning backwards?
Pair metrics with guardrails — and mind crowding-out
The pre-mortem will usually tell you the same hard truth lesson 4 proved: no single metric is gaming-proof. Every metric is a lossy projection of a rich goal onto one cheap number (Goodhart’s law), and whatever the number ignores is a free lunch for anyone optimizing it. So the practical defense isn’t “find the one perfect metric” — it doesn’t exist — it’s to pair a metric with a guardrail: a counter-metric that gets gamed in the opposite direction, so the cheap exploit on one shows up as damage on the other.
The analogy. A quantity metric alone is a gas pedal with no brake; flooring it just means crashing faster. The guardrail is the brake — paired so that you can’t slam one without the other screaming.
How to pair. Match each quantity/speed metric with a quality/counter metric that the obvious gaming would wreck:
| Quantity / speed metric | Obvious way to game it | Paired guardrail that catches it |
|---|---|---|
| Sales volume (deals closed) | Sign anyone, including bad-fit churners | Customer retention / revenue retained |
| Support tickets closed | Close fast without solving | Reopened-ticket rate; satisfaction (CSAT) |
| Calls handled per hour | Rush or hang up on customers | Post-call satisfaction score |
| Code shipped / story points | Ship fast, leave bugs | Defect / rollback rate |
| Production speed | Cut corners on quality | Error / defect rate |
Worked example — support, paired. Reward “tickets closed per day” alone and reps close tickets fast without solving anything (lesson 4’s Goodhart in action). Add a guardrail — a penalty for reopened tickets plus a satisfaction score — and the cheapest exploit (slam tickets shut) now hurts the rep’s number, because every unsolved ticket bounces back and every annoyed customer rates them down. The pair makes pure speed-gaming visible and costly. Back it with audits and human judgment that can say “you hit the number but you obviously cheated” — because no metric replaces a manager with discretion.
The pitfall — and a hard limit from lesson 5. Two traps here. First, don’t just pile on more targets: each new target is a new gaming surface, so a wall of fifteen KPIs is more gameable, not less. Add the minimum guardrails that close the cheapest exploits, then stop and lean on judgment. Second — and this is where lesson 5 returns — sometimes the right reward is not money at all. For work driven by intrinsic or moral motivation (care, craft, mission, helping), a clumsy cash bonus can crowd out the very motive that was doing the work; a bonus for “compassionate care” can cheapen compassion into a transaction. For that kind of work, recognition, autonomy, and mastery often beat cash, and the best “incentive design” may be to protect the intrinsic motive rather than bolt a price onto it.
Match the reward to the motive. For dull, rote, extrinsically-motivated work nobody loves (the FedEx sort, piecework, routine quotas), money aimed at the outcome is exactly right and won’t crowd out a motivation that was never there. For intrinsically or morally motivated work (caregiving, creative work, volunteering, mission-driven roles), lead with recognition, autonomy, and meaning; if you use money at all, make it large enough to stand on its own and frame it as appreciation, not a per-unit price — or you risk the daycare-fine backfire from lesson 5, where pricing the act crowded out the norm that was holding it up. The reward has to fit what’s already driving the person.
The incentive-design checklist
Before you ship any reward, run all five:
- Reward the outcome, not the proxy — does the payday land only when the real goal happens? Make the self-interested path the desired path.
- Align the agent with the principal — does this person lose when I lose? Give them skin in the game, both upside and downside.
- Pre-mortem the gaming — how would a clever, lazy, slightly dishonest person max this without doing the real work? Assume they will.
- Keep it simple — can you explain the reward in one sentence? If not, simplify; complexity hides gaming.
- Pair metrics + mind crowding-out — add a guardrail gamed in the opposite direction; audit; keep human judgment — and if the motive is intrinsic/moral, consider recognition over cash.
When to use it
Pairing and guardrails apply the moment you’re forced to reward a metric rather than a pure outcome (most real jobs). The crowding-out check applies whenever the behavior you want is already running on intrinsic or moral fuel — before you “improve” it with a bonus, ask whether the bonus might replace a stronger motive with a weaker one.
Match each broken incentive to the design fix that repairs it.
Pick a term, then click its definition.
Sort each incentive design into whether it's well-aimed (the self-interested path IS the goal) or will be gamed (the cheap path skips the goal).
Place each item in the right group.
- Bonus on contracts signed, as fast as possible, with no retention check
- Pair 'tickets closed' with a reopened-ticket penalty and CSAT
- Pay support reps purely on tickets closed per day
- Pay salespeople on revenue still retained a year later
- Pay the night crew per completed sort, then they go home
- An eight-factor weighted bonus formula 'too complex to game'
- Require the fund manager to hold their own savings in the fund
- Pay surgeons a bonus on raw survival rate, no guardrail
Fill in the master principle and its FedEx illustration:
Pick the right option for each blank, then check.
The master rule of incentive design is to reward the you actually want, so the self-interested path and the desired path become the path. FedEx fixed its late night sort by switching from paying — which rewarded time spent — to paying , which rewarded the completed sort.
A government wants to reduce hospital readmissions and proposes a bonus for hospitals with the lowest 30-day readmission rate. Run the pre-mortem: what's the most likely gaming, and the right design response?
Recap
You now have the constructive half of the whole subject — the rules that turn a dangerous lever into a useful one:
- Reward the outcome, not the proxy. Aim the reward at the real goal so the self-interested path is the desired path. The FedEx night crew finished on time the moment pay-per-shift rewarded the completed sort instead of hours logged. Resist the convenient measurable proxy.
- Align the agent with the principal — skin in the game. Make the agent share your outcome (retained revenue, patient outcomes, the manager’s own money in the fund), and beware asymmetric risk where they keep the upside but dump the downside on you.
- Pre-mortem the gaming, and keep it simple. Before shipping, ask how would a clever, lazy, slightly dishonest person max this without doing the real work? — then assume they will. Keep formulas simple; complexity hides gaming, it doesn’t prevent it.
- Pair metrics with guardrails, and mind crowding-out. No single metric is gaming-proof (Goodhart), so pair a quantity metric with a counter-metric gamed the opposite way, plus audits and human judgment. Don’t pile on targets (each is a new gaming surface). And for intrinsic or moral work, recognition and autonomy can beat cash — a clumsy bonus crowds the motive out.
- The fix is never “find more virtuous people.” It’s aim the reward at the thing you actually want.
Big picture
Designing good incentives — the whole machine
- Make the self-interested path the desired path
- Reward the outcome, not the proxy
- FedEx: pay per shift (sort done) beats pay by the hour (time spent)
- Pitfall: rewarding the convenient measurable proxy because the goal is harder to measure
- Align agent with principal
- Skin in the game: retained revenue, patient outcomes, manager invests own money
- Pitfall: asymmetric risk — keep the upside, offload the downside onto you
- Pre-mortem + keep it simple
- Ask: how would a clever, lazy, slightly dishonest person game this? Assume they will
- Simple, transparent formulas; complexity hides gaming, never prevents it
- Pair metrics + mind crowding-out
- No metric is gaming-proof (Goodhart): pair quantity with a quality guardrail + audits + judgment
- For intrinsic/moral work, recognition & autonomy beat cash — a bonus can crowd the motive out
- Reward the outcome, not the proxy
Check yourself: designing good incentives
What is the single master principle of designing good incentives?
Check your answer to continue.
Where this goes next
That’s the whole arc, end to end: you can now read a reward and predict the behavior it will produce, spot when a proxy will be gamed, recognize when money will crowd out the motive — and, as of this lesson, design a reward that aims at the outcome, gives the agent skin in the game, survives a pre-mortem, and carries a guardrail. There’s no more theory to add. What’s left is to prove you can run all of it, fast, on questions you haven’t seen.
Next is the Final Exam — graded, one question at a time, and one-way: once you submit an answer it locks for good, with no Back button, no retries, and the score shown only at the end. You’ll need 70% to pass. Carry the one sentence in with you: show me the incentive and I will show you the outcome — and when you design one, make the self-interested path the path you actually want.