Principal-Agent Problems Inside the Growth Organization: Incentive Design for Sales, Agencies, and Attribution

TL;DR: A growth organization is not one decision-maker with a strategy. A growth organization is a chain of contracts: the CEO hires a CMO, the CMO pays a sales force, the company retains an agency, and every team answers to an attribution model that decides who gets credit. Each link is a principal-agent relationship, and each one fails the same three ways. Strong incentives load risk onto the agent and invite gaming (Holmström, 1979). When one task is measured and another is not, high-powered pay on the measured task starves the unmeasured one (Holmström and Milgrom, 1991). And any number good enough to pay against eventually stops measuring what it once measured (Baker, 1992; Goodhart, 1975; Kerr, 1975). The evidence is quantitative and consistent: enterprise-software sellers time and discount deals to game accelerators, at a cost to the firm on the order of several percent of revenue (Larkin, 2014); fiscal-year-end incentives bend a company's whole sales calendar (Oyer, 1998); a structural redesign of one sales plan raised revenue roughly 9 percent (Misra and Nair, 2011). The corrective is not a values poster. The corrective is contract design: lower-powered incentives for multitask roles, holdout-measured bonuses instead of last-click credit, clawbacks and vesting against timing games, and paying agencies for incremental results rather than for spend. Culture does not fix a comp plan. The comp plan is the strategy.

---

The Comp Plan Is the Strategy

A vice president of sales sits with a spreadsheet in the second week of January. The document is a compensation plan: base salaries, commission rates, quota levels, an accelerator that pays more per dollar once a rep clears the number, a decelerator that pays less on the far tail. The plan will govern the behavior of forty people for a year. Every one of them will read it more carefully than they read the strategy deck, because the plan, not the deck, decides what they take home in March.

The strategy deck says the company will move upmarket, sell multi-year contracts, and protect gross margin. The comp plan pays a flat percentage on booked first-year value, uncapped, settled at quarter-end. A reader who knows only the plan can predict the year: reps will chase logos over multi-year commitments, discount hard in the last two weeks of each quarter to pull deals forward, and route the margin-protection speeches straight to voicemail. The plan and the strategy are pointed in different directions, and the plan wins, because the plan pays.

Here is the uncomfortable version. A growth leader does not execute a strategy. A growth leader writes contracts with self-interested parties whose actions cannot be fully observed, and then lives with what those contracts pay people to do. The sales rep, the agency, the analytics team scoring channels, each is an agent acting on behalf of a principal who wants something slightly different and cannot watch every move. Economists have a name for the resulting friction, and a forty-year quantitative literature on exactly how it breaks.

The argument here has three claims. First, the growth organization is a stack of principal-agent relationships, not an org chart. Second, every layer of that stack exhibits the same three failures, and the failures are structural, not moral. Third, the evidence points to specific contract designs, so the fixes are choices a leader can make on a spreadsheet in January, not exhortations to try harder in July.

A Growth Organization Is a Stack of Contracts

Start with the unit of analysis. A principal delegates a task to an agent because the agent has time, skill, or information the principal lacks. Delegation is why organizations exist. Delegation is also why they leak, because the agent's interests and the principal's never coincide exactly, and the principal cannot see everything the agent does.

Jensen and Meckling (1976, Journal of Financial Economics 3(4), 305-360) gave the friction a budget. The total cost of delegation, in their accounting, is the sum of three things a principal pays for imperfect control:

$\text{Agency cost} \;=\; \underbrace{M}_{\text{monitoring}} \;+\; \underbrace{B}_{\text{bonding}} \;+\; \underbrace{R}_{\text{residual loss}}$

Monitoring is what the principal spends to watch the agent: dashboards, deal desks, media audits, holdout tests. Bonding is what the agent spends to prove good faith: references, case studies, service-level guarantees, a portion of pay put at risk. Residual loss is the gap that survives both, the value lost because no contract is perfect and no amount of watching is complete. A growth budget has all three, whether or not anyone lists them that way. The line item labeled "agency management" is monitoring. The stock a sales rep vests into is bonding. The revenue a team never earns because the plan pointed it at the wrong quarter is residual loss.

Now stack the relationships, because a growth organization is several of them nested inside each other. The board is principal to the CEO. The CEO is principal to the CMO. The CMO is principal to a sales leader, who is principal to a rep. The company is principal to an outside agency. And every operating team is, in a strange way, principal to an attribution model, which acts as an agent that reports back how well the others performed. Each arrow carries its own monitoring, bonding, and residual loss.

The reason this framing earns its keep is that it makes the failures predictable. Once we read the organization as a set of contracts rather than a chart of people, the problems in sales, agencies, and attribution stop looking like separate accidents to be fixed one at a time by hiring better. We are looking at three instances of one structure, and the structure fails the same three ways every time, unless we change the contracts themselves.

Three Failures That Recur at Every Layer

The theory of incentives is not a grab bag of findings. Three results, each from a specific paper, compose into a single diagnosis. Read them in order and the whole stack becomes legible.

The price of strong incentives

Holmström (1979, Bell Journal of Economics 10(1), 74-91) asked the founding question: how should a principal pay an agent whose effort is hidden and whose output is noisy? Pay a flat salary and the agent, bearing no consequence, supplies minimal effort. Pay pure commission and the agent bears all the risk of noise, luck, a bad territory, a recession, and demands a premium to accept it, because agents are risk-averse and randomness is expensive to hold.

The optimal contract splits the difference. Under the assumptions that later made the result tractable, pay is linear in output, w = α + β y, where y = e + ε is effort plus noise. The interesting quantity is β, the intensity, the share of each marginal dollar the agent keeps. Holmström and Milgrom (1987, Econometrica 55(2), 303-328) derived the intensity that balances incentive against insurance:

$\beta^{*} \;=\; \frac{1}{\,1 + r\,\sigma^{2}\,c''(e)\,}$

We can read the denominator as a list of reasons to pay softer. r is the agent's risk aversion, σ² the variance of the noise, and c''(e) how sharply effort gets more painful at the margin. As any of them rises, β* falls. The louder the noise in the output measure, the less of the outcome you should tie pay to, because you would mostly be paying the agent to hold randomness. The intuition is worth stating plainly, because it inverts a common instinct: a noisier metric argues for weaker incentives on it, not stronger ones.

Holmström's other contribution here is the rule for which signals belong in a contract at all.

Hold onto the informativeness principle, because it is the formal reason incrementality tests beat last-click credit later in this essay. A performance measure is worth paying against to the exact degree that it carries information about what the agent actually did, and no further.

The multitask trap

Real jobs are never one task. A sales rep closes deals and also qualifies pipeline honestly, coaches a junior, and protects the customer relationship past the signature. An agency buys media and also tells you the unwelcome truth about which channels are dead. Holmström and Milgrom (1991, Journal of Law, Economics, and Organization 7, 24-52) worked out what strong incentives do when effort splits across tasks and only some tasks are measured.

Let the agent choose efforts t₁ and t₂ against a convex cost C(t₁, t₂), paid at rates β₁ and β₂. The agent equates marginal pay to marginal cost on each task, and the response of the unmeasured task to a stronger incentive on the measured one falls out as:

$\frac{\partial t_2}{\partial \beta_1} \;=\; \frac{-\,C_{12}}{\,C_{11}C_{22}-C_{12}^{2}\,}$

The denominator is positive whenever the cost function is convex, so the sign of the whole expression is the sign of -C₁₂. When the two tasks compete for the agent's finite attention, so that doing more of one makes the other more costly (C₁₂ > 0), raising the incentive on the measured task 1 lowers effort on the unmeasured task 2. Pay harder for bookings and you get less honest forecasting, less coaching, less relationship care, not as a betrayal but as an arithmetic response to the prices you posted.

When the measure becomes the target

The third failure is the most famous and the least understood, and it is the one we most often mistake for bad luck. Baker (1992, Journal of Political Economy 100(3), 598-614) formalized it: the value of an incentive contract depends on how well the performance measure tracks the principal's true objective, and every real measure is a distorted proxy. Pay against the proxy and the agent moves the proxy, along whichever dimensions the proxy and the objective have quietly come apart. The larger the gap between the measure and the goal, the more a stronger incentive buys you motion in the wrong direction.

Two adjacent laws state the same mechanism in plainer language. Goodhart's law, from Charles Goodhart's 1975 work on monetary policy, holds that a measure used for control stops behaving like the measure it was. Campbell's law (Campbell, 1979, Evaluation and Program Planning 2(1), 67-90) says the more a quantitative indicator is used to make decisions, the more it corrupts the process it was meant to track. Both describe a metric collapsing under the weight of being paid against.

Steven Kerr saw the organizational version first. Kerr (1975, Academy of Management Journal 18(4), 769-783) catalogued institutions that reward one behavior while sincerely hoping for another, then profess bewilderment at the result.

We keep rewarding A while hoping for B, then act surprised when we get exactly the A we paid for.

, after Steven Kerr, Academy of Management Journal (1975), paraphrased

Put the three together and we have the whole diagnosis. Strong incentives load risk and invite gaming. Incentives on one measured task crowd out the unmeasured tasks. And the measure you chose degrades the moment it becomes the target. Gibbons (1998, Journal of Economic Perspectives 12(4), 115-132) and Prendergast (1999, Journal of Economic Literature 37(1), 7-63) survey the field, and their consensus is sobering for anyone hoping incentives are a tuning dial: incentives are powerful, which is precisely why they are dangerous, because they work on whatever you actually wrote down rather than on what you meant.

Before turning to the evidence, we should see the whole stack at once, because the three failures land in a fixed pattern at each layer. Each principal pays for a proxy, each agent maximizes the proxy, and the proxy decays as a measure the moment it becomes the target.

The Sales Force: Timing Games and What They Cost

Sales compensation is the layer where the theory was tested first and hardest, because the data are unusually clean: individual output, individual pay, sharp period boundaries, and steep nonlinearities. What the data show is that reps respond to the contract with a precision that would flatter a physicist.

Oyer (1998, Quarterly Journal of Economics 113(1), 149-185) documented the coarsest version. Firms whose incentives reset at the fiscal year end show a spike in sales in the final period, and the spike has no counterpart in underlying demand. The calendar of the comp plan, not the calendar of the customer, shapes when revenue lands.

Anyone who has run a quarterly quota knows the shape of it. Bookings limp along for ten weeks and then a wall of paper arrives in the last two, much of it discounted to close before the boundary. The pattern below is from an advisory engagement, not a public dataset, but it is the same curve Oyer, Larkin, and Steenburgh describe, and every sales operations leader recognizes it on sight.

Larkin (2014, Journal of Labor Economics 32(2), 199-227) put a price tag on the finer game. Studying an enterprise-software vendor with accelerating commission tiers, he found reps timing deals to land in periods where the marginal commission was highest, and discounting to move a deal across a boundary that mattered to the rep but not to the firm. A customer who would have paid more next month got a price cut this month, because the cut was worth more to the rep's accelerator than the discount cost the rep personally. The firm ate the difference.

Several percent of revenue is not a rounding error; for many software companies it is the entire annual margin-expansion target, handed away at the quarter boundary in exchange for personal commission timing. The high-powered plan did not fail to motivate. The plan motivated the wrong thing, expensively, exactly as Baker's distortion result predicts.

Now the necessary complication, because the story is not that all variable pay is waste. Steenburgh (2008, Quantitative Marketing and Economics 6(3), 235-256) asked whether lump-sum bonuses merely shift the timing of sales or actually call forth more effort. If the pure timing-game story were the whole truth, bonuses would only rearrange when deals close. He found the opposite: bonuses raised genuine effort, not just intertemporal reshuffling. Chung, Steenburgh and Sudhir (2014, Marketing Science 33(2), 165-187) reached a compatible conclusion with a dynamic structural model, showing that quota-bonus plans and overachievement commissions lift real productivity rather than only bending the calendar. Incentives are not futile. Incentives are sharp instruments that cut whatever you point them at, and pointing matters more than power.

The constructive end of this literature is the redesign result. Misra and Nair (2011, Quantitative Marketing and Economics 9(3), 211-257) estimated a structural model of salesperson behavior at a contact-lens manufacturer, then implemented the model's recommended plan in the field. The recommendation was, in part, to remove the quota ceiling that capped payout, because the cap gave top reps a reason to stop selling once they hit the number. Removing the cap and recalibrating the rates raised revenue by roughly 9 percent.

Notice what changed and what did not. Misra and Nair did not exhort anyone; they diagnosed a specific kink, the ceiling, that was paying people to stop, and removed it. The 9 percent came from geometry, not motivation. The shape of every real fix in this domain is the same: find the place where the contract pays for the wrong thing, and change the contract.

The Agency Relationship: Commission on Spend Is the Problem

Move one layer out. When a company hires an outside agency, it faces the oldest question in the theory of the firm: make or buy. Coase (1937, Economica 4(16), 386-405) asked why firms exist at all rather than buying every input on the market, and answered that markets have transaction costs, search, negotiation, and the hazard of holdup, that internal authority sometimes avoids. Williamson (1975; 1985) built this into a theory of when to integrate: the more relationship-specific the investment and the more room for opportunism, the more a firm should bring the activity in-house. In-housing a media team is not a fashion. In-housing is vertical integration, chosen when the agency contract cannot be written well enough to be worth its transaction costs.

The reason the agency contract is so hard to write well is that the standard form pays for the wrong variable. The traditional agency is paid a commission on media spend. Read that as a contract and the incentive is naked: the agency's revenue rises with your spend, not with your results, so the two of you are aligned on exactly one thing, spending more.

The problem is not a theoretical worry. An independent 2016 investigation of United States media buying, commissioned by the Association of National Advertisers and conducted by K2 Intelligence, documented non-transparent practices in which rebates, markups, and value-bank arrangements flowed from media suppliers to agencies rather than back to the advertisers whose money bought the media. The structure paid the agency to spend, so the agency found ways to be paid more for spending.

The remedy is to change the variable the contract pays against. Three common forms allocate risk very differently, and the choice among them is the whole game.

The fee-plus-incrementality form is the informativeness principle applied to a vendor. Pay a fee that covers the agency's real cost and holds its risk to a tolerable level, then put a bonus on a signal that actually carries information about the agency's contribution, incremental outcomes measured against a holdout, rather than on spend, which carries almost none. In one engagement, the difference was stark.

The pattern generalizes past this one case. Any time a vendor is paid on an input it controls, it will supply more of that input than you want, and the fix is always the same in structure: find a signal informative about the outcome you care about, make measuring it cheap enough to be worth doing, and pay against that.

Attribution Is a Contract

Here is the layer most growth teams miss, because it does not look like an employment relationship. An attribution model is a contract. The model decides which channel, which campaign, which team gets credit for revenue, and credit determines budget, headcount, and bonuses. The moment credit is allocated by a rule, everyone whose pay depends on the rule becomes an agent optimizing against it, and the rule becomes a performance measure subject to every failure Baker described.

Last-click attribution is the default, and read as a performance measure it is close to the worst possible one. Last-click hands full credit to whatever touch immediately preceded the conversion, which systematically overcredits channels that harvest existing intent, branded search, retargeting, the abandoned-cart email, and starves channels that create intent in the first place, upper-funnel display and social. Pay teams on last-click and they will pour budget into intent harvesting, because that is where the credit lands, until the demand those channels merely intercept quietly stops being generated upstream.

Last-click states the informativeness principle in reverse. The default signal carries very little information about which touch actually caused the sale, and a great deal of information about which touch happened to be nearest the finish line. Blake, Nosko and Tadelis (2015, Econometrica 83(1), 155-174) ran the decisive experiment: eBay switched off branded paid search and found the incremental return was near zero, because the users clicking those ads would have arrived anyway. The attribution model had been paying handsomely for traffic it was not causing. Gordon, Zettelmeyer, Bhargava and Chapsky (2019, Marketing Science 38(2), 193-225) generalized the lesson across large field experiments at Facebook, showing that common attribution methods diverge sharply from experimentally measured lift. Attribution, as usually practiced, is a distorted measure being paid against by everyone it scores.

The chart above is an advisory-engagement measurement, not a published dataset, but the direction is exactly the one Blake and Gordon document: the channels that win under last-click are the ones nearest the conversion, and the channels that generate the demand are the ones last-click cannot see. Incrementality tests, geo-holdouts, ghost bids, matched-market designs, are the informativeness principle made operational. A holdout answers the only question the contract should pay against: what would have happened without this spend?

Before adding any metric to a bonus, put Holmström's question to it directly. Does moving this number show that the agent caused the outcome, or only that the agent stood near it when it happened? A channel that would look identical whether or not it caused a single incremental sale is a spectator, not a cause, and paying against it funds spectating. Incrementality is slow and expensive, and it is also the only marketing measure that survives being paid against, because it is defined as causal contribution rather than correlational credit.

There is a political economy to all this, and it is where the abstract theory meets an actual conference room. Whoever owns the attribution model owns the budget, so the model is never a neutral instrument for long. The team scored by the model has every reason to lobby for the weights that flatter it.

Attribution is where the whole thesis closes. The company is principal to a model that acts as agent, reporting on the performance of all the other agents. When that reporting agent is captured by the parties it reports on, every contract above it in the stack is being scored by a corrupted instrument, and the three failures we have catalogued compound on top of each other.

Where the Theory Runs Out

Honesty about a framework requires marking its edges, and incentive theory has three that matter to an operator.

The first is measurement cost. The informativeness principle says to pay against the most informative available signal, but it is silent on the price of obtaining it. Incrementality tests cost real traffic, real time, and real statistical power; a geo-holdout large enough to detect a modest lift can idle a meaningful slice of a market for weeks. For a small program, the measurement can cost more than the misallocation it prevents. The right response is not to abandon the principle but to spend on measurement in proportion to the budget at stake: we run holdouts on the channels that move real money and accept cruder proxies where the dollars are small, only if the proxy is at least directionally informative.

The second is the unmeasurable, which the multitask result already flagged and which Baker, Gibbons and Murphy (1994, Quarterly Journal of Economics 109(4), 1125-1156) confronted directly. When the tasks that matter most cannot be captured in any objective number, the answer is not to invent a fake number and pay against it. The answer is subjective performance evaluation, a manager's considered judgment, backed by a relational contract in which the principal's promise to reward good work fairly is kept over time and therefore believed. Subjective measures reintroduce their own hazards, favoritism, leniency, the cost of trust once broken, but they are frequently less distorting than a precise measure of the wrong thing.

The third edge is fairness and retention, and it is where the mathematics goes quiet and management begins. A relative-performance scheme that filters out common luck, evaluating a rep against peers who faced the same market rather than against an absolute number, is statistically cleaner and is why Holmström's original work pointed toward it. But relative schemes pit colleagues against each other, discourage the coaching and cooperation that are themselves valuable unmeasured tasks, and in the extreme invite sabotage. A contract can be optimal on a whiteboard and corrosive in a room. The people under the plan have memories, compare notes, and leave, and a plan that treats them as effort-supplying machines will optimize itself into an empty office.

What to Build Instead

The purpose of the diagnosis is a different set of decisions in January. None of what follows is a tip. Each is a contract choice with a paper behind it.

Match incentive intensity to how noisy and how singular the role is. A single-task role with a clean, causal measure can carry a high β. A multitask role whose important work is unmeasurable should carry a low one, more salary and less commission, precisely because strong pay on the one countable task would starve the rest. Lowering the powder on a multitask role is not timidity. Lowering the powder is what Holmström and Milgrom's result prescribes.

Pay against incremental outcomes, not proxies. Replace last-click credit with holdout-measured contribution wherever the budget justifies the test. Replace agency commission on spend with a fee plus a bonus on incremental results. In both cases we swap a measure rich in luck and gaming for one that carries information about the agent's true contribution, which is the only kind of measure worth attaching money to.

Combine input metrics with output metrics rather than choosing between them. Output metrics resist gaming on quality but are slow, noisy, and hard to act on day to day; input metrics are fast and controllable but game easily once paid against. Paying on a small basket of both, and revising the basket before any single input hardens into a target, is the practical answer we keep returning to, and it is the bridge between strategy and daily action that the strategy-execution gap essay works through in detail.

Use the boundary treatments the sales literature validates. Accelerators past quota can lift real effort (Steenburgh, 2008; Chung, Steenburgh and Sudhir, 2014), so keep them, but pair them with clawbacks on churned or discounted-to-death deals and with vesting that pays commission as revenue is actually collected, not at signature. Clawbacks and vesting are commitment devices: they bind the agent's payoff to outcomes that arrive after the quarter closes, which is exactly when the timing games would otherwise pay off.

Design for truthful self-selection. A plan where reps sandbag their forecasts to earn a soft quota is a plan that punishes honesty; a plan where the rep does best by forecasting and working truthfully is incentive-compatible, and building that property in is an exercise in mechanism design, not in trust. Menus of quota-and-rate options that lead each rep to reveal their true expectation are a known instrument here.

And separate the scorekeeper from the players. The team paid on a number must not own the definition of the number. Put attribution, and any metric that allocates budget, in the hands of a function with no stake in the answer, and let experiments settle disputes. The sibling question of how two parties split the surplus once the contract's incentives are set, the bargaining rather than the moral hazard, is taken up in bargaining theory for enterprise deals; the two problems compose, because you first design the incentive and then negotiate the price of it.

Return with us to the January spreadsheet. The strategy deck and the comp plan were pointed in different directions, and the plan won. A leader who has internalized the three failures does not write a better speech about margin. The leader writes a plan that pays for multi-year value, vests commission as cash arrives, caps last-click's authority with a holdout, and pays the agency for lift instead of spend. The strategy and the contracts now point the same way, and the organization moves in that direction without anyone being asked to be more disciplined than their pay allows.

Key Takeaways

1. A growth organization is a stack of principal-agent contracts, CEO to CMO, CMO to sales, company to agency, team to attribution model, and each link carries Jensen and Meckling's (1976) three costs: monitoring, bonding, and residual loss. The comp plan is the strategy, because the plan is what actually pays people.

2. Strong incentives load risk onto risk-averse agents and invite gaming; the optimal intensity β* falls as the measure gets noisier (Holmström, 1979; Holmström and Milgrom, 1987). A noisier metric argues for weaker pay tied to it, not stronger.

3. In multitask roles, high-powered pay on the one measured task provably starves the unmeasured ones (Holmström and Milgrom, 1991). When the important work cannot be counted, the correct contract is deliberately low-powered.

4. Any measure paid against degrades as a measure (Baker, 1992; Goodhart, 1975; Campbell, 1979; Kerr, 1975). Sales reps time and discount deals to game accelerators at a cost of several percent of revenue (Larkin, 2014); fiscal-year-end incentives distort the sales calendar (Oyer, 1998).

5. Redesign beats exhortation. Removing a quota ceiling that paid top reps to stop selling raised revenue about 9 percent in a field implementation (Misra and Nair, 2011). Find the kink that pays for the wrong thing and change it.

6. Commission on media spend aligns an agency with your costs, not your growth (ANA/K2 Intelligence, 2016). A fee plus an incrementality bonus swaps a spend signal for a causal one; in one engagement it cut spend roughly 20 percent while raising incremental results about a quarter.

7. Attribution is a contract, and last-click is a distorted performance measure that overcredits intent harvesting and starves demand generation (Blake, Nosko and Tadelis, 2015; Gordon et al., 2019). Whoever owns the model owns the budget, so the scorekeeper must not be a player.

Further Reading

• The Strategy-Execution Gap in Growth Teams, how input metrics bridge strategy and daily action; the constructive companion to this essay's warning about metrics that degrade under targeting.
• Multi-Touch Attribution Is Broken, the measurement mechanics of why last-click misallocates budget, and the causal alternative.
• Incrementality Testing at Scale, the informativeness principle made operational: how to run the holdouts that make an incrementality bonus contractible.
• Unified Measurement Architecture, how MMM, attribution, and experiments reconcile into one scorekeeper the paid teams cannot quietly capture.
• Discount Engineering: When Promotions Destroy LTV, what gamed, timing-driven discounting does to the customer economics on the other side of the contract.