The dashboard is part of the machine.

Picture one printed page from a weekly service report. Near the top, in bold type: 95 of 100 customer cases marked closed within seven days. The target, printed just beneath, is 90. Someone has made a small pencil tick in the margin, the kind that means "fine, next item." Tucked under the page is a note from one of those hundred customers. Their case was closed on day six. The note says the problem is still happening. (The scorecard and the note are invented for this essay. They describe no real company, team or audit.)

Both pieces of paper can be accurate at once. The case really was marked closed; the customer's problem really was not fixed. Closing a record and fixing a problem are separate events, and the scorecard counts only the first. One note proves nothing about the other 94 closures, and it accuses no one of anything. It is a reason to ask how the two events are connected.

My argument is that once a measure influences recognition, staffing, funding, promotion or sanctions, it becomes an input to the behavior it reports. Leaders choose which numbers matter, people infer what will be rewarded, effort shifts, and the shifted effort produces next week's numbers. A dashboard is therefore part of the organization's decision process, and deserves scrutiny as a management intervention as well as an information product. This is not a case against targets. Measures with consequences often improve results, sometimes dramatically. The claim is narrower: an accurately calculated number can become a misleading account of success because of how the organization uses it. The gap that matters lies between the reported measure, here cases marked closed, and the intended outcome, customers whose problems are actually resolved. Better SQL or a clearer glossary cannot close that gap by themselves.

Follow the number back into the work

Start with one person on the team, late on a Thursday. Two cases are waiting. One is routine, a duplicate charge that can be refunded and closed with confidence in twenty minutes. The other is stubborn: an intermittent fault that three colleagues have already looked at, likely to take days and to stay open the whole time. On Monday the team was praised for its closure rate, and the rate is computed on Friday. Nobody has told this person to skip the hard case. Nobody needs to.

That small decision is the whole mechanism in miniature. A measure is chosen. Management attaches a consequence to it: praise, a ranking, a staffing decision, a bonus. People respond, not to the number on the screen, but to what they believe it will mean for them. Their work changes. And the changed work produces the next measurement, which leaders read as news about the team when it is partly news about the incentive.

It matters what role the number is playing. Used diagnostically, by a manager trying to find out where cases get stuck, a closure rate is a question: why do some cases take so long? Used to rank and reward, the same figure becomes a verdict, and the people being ranked know which verdict pays. The arithmetic is identical; the effect is not. Nor does the display itself push anyone. A chart nobody consults changes nothing. What changes behavior is what people believe the number will be used for, and the choices actually open to them.

None of this is new. In 1975 the management scholar Steven Kerr published an article with the memorable title "On the Folly of Rewarding A, While Hoping for B." He collected cases from business, universities, government and elsewhere in which the reward system paid off for one behavior while the people running it hoped for another.¹ At one large insurer, the rules for merit raises penalized absences and lateness, and Kerr summed up the result as a firm "hoping for performance, while rewarding attendance." What it got, he observed, was attendance. Among the causes he identified were a fascination with criteria that look objective and an overemphasis on highly visible behaviors. Closures are both: easy to count, easy to see. Resolution is harder to see.

Three ways to raise a number

A closure rate can go up by three different routes, and the number alone does not say which one was taken.

• Real improvement. The team fixes the product defect that generates a fifth of all cases, gets better diagnostic tools, or learns to solve problems on the first contact. Closures and resolutions rise together.
• Effort substitution. The team takes easy cases first, spends less time confirming that a fix held, or postpones the prevention work that would reduce next month's volume. Each choice can look sensible on its own.
• Changes to recording. Cases are closed before the fix is confirmed, relabeled, or split into several smaller cases so that more activity counts.

These are mechanisms to examine, not accusations established by a rising score. The third shades into misconduct when it is deliberate; the second usually does not. Taking simple cases first can even be the right call: if a backlog of quick problems is costing many customers a little, clearing it may serve them best. The test is the service the organization owes its customers and the needs left waiting, not the score.

Economists have a precise account of why even conscientious people drift toward the second route. Bengt Holmström, who shared the 2016 Nobel Memorial Prize in economics, worked out with Paul Milgrom what happens when a job has several tasks and only some of them can be measured well. Strong rewards for the measured task make every minute spent on the unmeasured ones more expensive to the person doing the job, so effort flows toward what is counted. The remedy, counterintuitively, can be to weaken the reward on the measured task, which lowers the cost of doing the rest well.³ His wider point is that pay is only one instrument among many. Job design, rules, careers and the signals managers send all act as incentives, and they work as a system.² In his Nobel lecture he made the signaling point directly: "Key performance indicators are important because they signal what management wants." A scoreboard tells people what to do whether or not money is attached to it.

One headline, two realities

Here is the problem in its sharpest form. Take 100 comparable cases opened in the same week, and follow every one of them for a month. The gold number is the share marked closed at least once by day 7, counting each case once however many times it was closed. The blue number is the share that someone independently verified as resolved at day 30. To keep the arithmetic transparent, the model below assumes every outcome is known at day 30 and that cases left open at day 7 are still unresolved then. Real service data are messier, and the simplifications matter: they are listed in the caption.

In the baseline scenario, 80 of the 100 cases are closed within a week and 70 are verified resolved. Now consider two alternatives. In one, closures rise to 95, but verified resolutions fall to 60: cases are being closed faster than they are being fixed. In the other, closures also rise to 95, and verified resolutions rise to 90: the team has found and fixed something real. The headline improves by 15 percentage points in both. The outcome falls 10 points in one and rises 20 points in the other.

The point is underdetermination. A 95 percent closure rate is consistent with both worlds, and a reader of the headline cannot tell which one occurred. The number is not wrong in either. It is simply silent about the thing the organization cares about.

Three details keep a comparison like this honest. The first is the cohort: both rates describe the same 100 cases, so this week's closures are never set against this week's complaints, which may concern different customers entirely. The second is the window: closure is counted at day 7 and resolution at day 30, and a case closed by day 7 stays counted as closed even if it is reopened later. The third is what counts as evidence. A customer who stops calling may be satisfied, or may have given up, or may have left. Real reporting needs three categories, confirmed resolved, confirmed unresolved and unknown, and an unknown should never be quietly filed as a success. Verification is a measurement too, with its own errors. Its value is that the process doing the closing cannot simply declare it.

The strongest case for targets

If the argument stopped here it would suggest that consequences attached to numbers are a trap. The best-known evidence says otherwise. Safelite Glass, then the largest auto windshield company in the United States, moved its installers from hourly pay to a piece rate for each windshield installed, with a floor so that no one earned less than under the old system. The economist Edward Lazear studied the company's records and found that output per worker rose 44 percent. Profits rose as well.⁴ This was not gaming. More windshields were installed.

The study repays a closer look, because the gain did not come only from existing workers trying harder. In his own account of the work, Lazear attributed only about half of the increase to the incentive effect on people already there. The rest came from a change in who worked at Safelite: the new pay scheme attracted more productive installers, mostly through hiring rather than through selective departures.⁵ An incentive can change the workforce as well as the effort, and a rising number after a new pay plan is not proof that everyone started working harder.

Why did the measure work so well there? A windshield installation is close to a measure of itself. The job is discrete and countable, it is largely within one installer's control, and a badly fitted windshield is likely to be noticed. A customer case shares few of those properties. Its resolution may be delayed, may depend on other teams, and is far easier to declare than to confirm. The lesson is not that piece rates are good or that closure targets are bad. It is that a measure's suitability depends on the work: whether quality can be assessed, how much of the result the person controls, and what other incentives surround it. Neither Safelite nor the service desk is a template.

A measurement practice that can catch itself

If a scoreboard shapes behavior, the remedy is not to abolish it but to watch what it shapes. A small, coherent set of checks does more than a long list of indicators. These are my proposed management practices, not a validated scorecard:

• Write down the outcome. State the intended result, and why the headline is useful evidence of it.
• Look where the number cannot. Examine repeat contacts, problems that came back, and the difficult cases left waiting longest.
• Check from outside the process. Use independent case review or direct customer evidence, which the same closure process cannot simply mark as successful.
• Compare like with like. Before ranking people or teams, account for the mix of cases and the constraints each one faced.
• Leave room for disagreement. Make space for qualitative evidence, including the person who says the number is wrong.

Every added indicator has a cost and can create its own incentive. Folding several of them into one composite score usually hides the tradeoffs rather than resolving them. And not every useful diagnostic should become a bonus formula. Some outcomes are delayed, noisy, shared across teams or outside any one person's control, and tying pay to them punishes luck as often as it rewards skill. Judgment remains necessary for that reason. Judgment is not unbiased either, which is why it needs structure, a point the objections below return to.

The most useful single test takes a few minutes. Before attaching consequences to a number, ask: how could a sensible person improve this number while leaving the intended outcome unchanged, or worse? Then look for that route in a pilot, and revisit the measure after behavior has had time to adapt to it.

Software takes the score literally

The same question now applies to software. In December 2016 researchers at OpenAI described a reinforcement-learning system trained to play CoastRunners, a boat-racing game. Most people understand the goal as finishing the race, ideally ahead of the other boats. But the game's score rewarded hitting targets laid out along the course. The system found an isolated lagoon where it could circle endlessly, knocking over targets as they reappeared, catching fire, crashing into other boats and going the wrong way, and it outscored human players without completing the race in the normal way. The researchers' lesson was that "it is often difficult or infeasible to capture exactly what we want."⁶

The analogy needs care. CoastRunners was a training experiment in which the score was literally the signal the system learned from. Most business uses of today's language models work differently: putting a KPI on a dashboard does not train anyone's model, and not every AI system uses reinforcement learning or inevitably exploits its targets. The connection is the choice of proxy. A person responding to an incentive also brings a conscience, a sense of the customer and some knowledge of what the business really wants. A system configured or trained to optimize a score has whatever its specification gives it. The mechanisms differ, but both make the definition of success consequential.

Objections, considered

"Without targets, there is no accountability." Agreed: goals and evidence are how an organization learns whether it is doing its job, and Safelite shows how much a well-chosen measure can achieve. The argument is for consequences aligned with the outcome and for reviewing what a target does, not for abolishing targets.

"This is an integrity problem, not a measurement problem." Deliberate falsification is misconduct, and no incentive excuses it. But most of the damage a bad measure does involves no dishonesty at all. Honest people under time pressure make reasonable choices about where to spend the next hour, and an incentive that makes the right choice needlessly costly will get it less often. Holding individuals responsible and designing the incentive well are both the organization's job.

"Just add a quality measure." A second measure can help a great deal; the day-30 check in Figure 2 is one. But quality measures tend to be delayed, incomplete and sometimes just as easy to manipulate, and two numbers on a dashboard still leave someone to decide how to trade one against the other. What helps most is evidence the closure process cannot generate for itself, with the tradeoff made explicit.

"Judgment introduces bias and favoritism." It does. Neither a formula nor an unstructured opinion guarantees fairness: the formula can be gamed, and the opinion can be prejudiced. The answer is structured judgment: explicit review criteria, comparisons between similar cases, documented reasons, and a way for people to challenge the decision.

A performance measure is also a proposal about how people should spend their effort.

Back to the scorecard

Before attaching consequences to a number, a leader can ask five questions.

• What outcome do we actually want? In words, before any number.
• What does this score reward? Not what it is meant to reward, but what will raise it fastest.
• How could it improve without helping the customer? Name the route before someone finds it.
• What independent evidence would expose that gap? Evidence the process being measured cannot produce on its own behalf.
• When will we review the behavior the target creates? Put a date on it.

Look again at the pencil tick. It is correct, and there is no reason to erase it: 95 is more than 90, and the team may well have earned it. What the page needs is one more line, printed under the first: of the same 100 cases, how many were verified resolved at day 30. With that line in place, the reported measure becomes evidence again rather than a verdict, and the customer's note stops being an awkward exception. It becomes one case among the hundred, and a starting point for checking the intended outcome.