
The indicator after it becomes a target
When a measure becomes a target, it can change the work and the data describing it. Understanding an improvement requires examining how it was achieved and what was left out.
Between 1994 and 1995, Safelite changed its auto glass installers’ pay, moving from hourly wages to a scheme tied to units installed. In Edward Lazear’s study published in 2000 (opens in a new tab), the change is associated with a 44% increase in output per worker. When the same workers are compared across the two regimes and other factors controlled for, the estimate is approximately 22%. The aggregate result also included changes in workforce composition; it did not mean every person had increased output by 44%.
There were conditions that made that form of incentive particularly relevant: output could be counted and defects detected relatively quickly. The study also describes mechanisms for dealing with rework. The quantity being rewarded was not entirely separate from quality. The value of the case is not in transferring its pay scheme to any company, but in recognizing something a hasty critique of metrics may forget: changing behavior through a measure can produce a real improvement. Often that is precisely what we seek.
When we turn an indicator into a target, we expect people to pay it more attention and organize part of their work to improve it. Measurement no longer merely describes an activity: it intervenes in the reasons for doing it one way rather than another. That can direct effort toward a neglected need. It can also change which cases receive attention, which tasks are postponed and what ultimately enters the record. To interpret the subsequent figure, we need to understand that response, not assume the data are still being produced under the same conditions.
The same improvement can come from different work
Other responses are possible. The team may learn to record what it already did more accurately and correct an earlier underestimate. It may change a request’s classification, the point at which the clock starts or the criterion for considering it resolved. Some modifications improve measurement; others allow the target to be met without solving the problem better. A written definition that stays the same does not guarantee comparability either: the population of cases may have changed even if the formula retains its terms.
The problem, then, is not exhausted by discovering who manipulated the data. In Bengt Holmström and Paul Milgrom’s multitask model (opens in a new tab), incentives affect not only how much effort is made but also its distribution across tasks or dimensions of work. Strongly rewarding what is easy to measure may draw attention away from less observable contributions. This is a theoretical relationship, not a prediction that every target will impair performance. It helps explain why someone may honestly follow the priority they were given and, in doing so, leave something the organization also needs undone.
In our example, spending time understanding a complex request may worsen the individual’s result for the period and prevent several later requests. If that contribution has no recognizable place in the assessment, the design makes it more costly to undertake. Adding an exhortation to quality does not, on its own, resolve that difference. Whoever organizes the work needs to decide how to reconcile the speed demanded with the time an adequate response requires, rather than expecting each person to absorb the conflict silently.

The relationship with the outcome also needs review
A measure may have been a good signal before consequences were attached to it and become less informative afterward. Pascal Courty and Gerald Marschke (opens in a new tab) examined this problem in work published in 2008 and applied to a US job training program. Their proposal analyzes whether the association between a performance measure and the organization’s objective changes when the measure is introduced into the incentive system. They found evidence consistent with distortions: a rewarded measure may lose part of its relationship with what it was intended to represent. Establishing that relationship before activating the incentive is therefore not enough.
The resulting management question does not require reproducing that statistical analysis in every team. It does require retaining evidence about the outcome we cared about and examining how the indicator is now produced. For requests, we would also need to look at cases still open, their difficulty and the recurrence of the problem after closure. Those observations do not guarantee a causal explanation, but they allow us to distinguish questions the average conceals: whether we resolve cases sooner, resolve a different mix of cases or merely changed what we record as resolution.
Turning each of those observations into another target may reproduce the problem rather than clarify it. Some signals need to serve investigation, comparison or learning without automatically determining a reward or penalty. That does not make them immune to error or grant them neutrality. It keeps open a source of understanding that need not be subject to exactly the same pressure as the main indicator. The way forward is not to design a dashboard capable of rewarding everything, but to retain ways of examining what its rewards leave out.
Conversation with those doing the work also has an evidentiary function. They can show which decisions changed when the target was introduced, which cases became less attractive and which information stopped being recorded. Their explanations require checking, just as the report does. But if communicating a difficulty can only be interpreted as an excuse or a threat to compliance, the organization reduces its chances of recognizing how the design it introduced is working.
Before making a target more demanding, it is worth articulating the improvement we expect to produce: what should be done differently and which conditions we are unwilling to worsen to achieve it. We will then need to compare that expectation with the work observed. If the case mix changed, the response may be to review the comparison. If a necessary contribution emerged that the target penalizes, the assignment or incentive may need to change. If there was deliberate falsification, it should be treated as such. These situations do not call for the same intervention and are not resolved simply by demanding a higher number.
Anyone setting a target is choosing a way to change the work, even if all they wrote was a number. The review does not end with checking whether it was reached. It must be able to answer a more uncomfortable question: what did we learn to do in order to meet it, and what did we stop doing along the way?





