A KPI encourages the wrong behaviour when the figure can rise without the result you are after rising. People are not necessarily “cheating”: they do exactly what the indicator asks of them, and the indicator asks for only part of the work. The article explains the mechanism (Goodhart’s law and its relatives), shows for eight roles what can go wrong and which paired indicator prevents it, calculates an example in which the pair changes the ranking and gives you a five-question test for any KPI before you tie it to money.
Quick answer. An indicator tied to a stake (a bonus, a ranking, a sanction) becomes a target, and people optimise the figure, not the result. Prevent the effect with a paired indicator that would fall if the volume were achieved badly, with data from sources the person does not control, with spot checks and with continuous thresholds instead of abrupt steps. Test every KPI with the question: “How could the figure rise without the result rising?”
Definitions: Goodhart’s law, Campbell’s law, the perverse incentive
Goodhart’s law is the observation that a statistical regularity tends to collapse once it is used for control. The economist Charles Goodhart formulated it in 1975 for monetary policy. The form known today, “when a measure becomes a target, it ceases to be a good measure”, belongs to the anthropologist Marilyn Strathern (1997), in an article on the assessment of British universities.
Campbell’s law says that the more a quantitative social indicator is used for decision-making, the more it is subject to corruption pressures and the more it distorts the processes it is meant to monitor. Donald T. Campbell published the formulation in Evaluation and Program Planning (1979). His classic example is teaching to the test: marks go up, competence does not.
The perverse incentive is a reward that produces the opposite of the intended effect. Steven Kerr described the organisational version in 1975, in “On the folly of rewarding A, while hoping for B”: companies hope for one behaviour (B), but reward another (A), then wonder why they get A.
The cobra effect names the case in which the incentive makes worse the very problem it was meant to solve. The anecdote about bounties for cobras in colonial Delhi is popular, but its historicity is disputed. The case documented in the archives is that of the bounties for rat tails in Hanoi in 1902: rats without tails appeared, left alive to breed.
Remember. The problem is not measurement, but the stake placed on an incomplete measure. An indicator followed only as information has fewer reasons to go wrong. The same indicator, tied to a bonus or to a public ranking, starts to be “managed”.
Why does a good indicator go wrong when it becomes a target?
Four mechanisms explain almost every case. You can use them as a checklist for any new KPI.
1. The indicator covers only part of the result
The number of calls taken measures volume, not resolution. Sales measure what was collected, not the returns or complaints that follow. When only the measured part carries a stake, effort moves there. Ordóñez, Schweitzer, Galinsky and Bazerman call the first side effect of targets “a narrow focus that neglects non-goal areas”; their list also includes a rise in unethical behaviour and in risk-taking.
2. The person controls the data
If the person being evaluated reports the figure themselves (visits made, tasks closed, hours worked on a project), the figure also depends on how strictly they report. No bad faith is needed: a task that is “almost done” more easily becomes “done” at the end of the month, when it counts.
3. The threshold has an abrupt step
A bonus that is triggered at 100% of plan creates a specific behaviour around the threshold. At 97% on the 28th of the month, any effort for another 3% is worth it, including moving sales from the following month. At 130%, it is worth putting sales off until the following month, especially if there is a cap. The spread of results shows the effect: many values just above the threshold, few just below it.
4. The moment of recording can be chosen
Orders, invoices, ticket resolutions or hires have a date that someone can move. Monthly indicators with a high stake encourage moving things between periods. The annual result does not change, but the monthly evaluation becomes a measure of skill at choosing the date.
A documented case: accounts opened without consent
In September 2016, the US Consumer Financial Protection Bureau (CFPB) fined the bank Wells Fargo 100 million dollars. The CFPB press release says that employees, “spurred by sales targets and compensation incentives”, opened accounts without customers’ knowledge. The bank’s own analysis identified more than two million deposit and card accounts that may not have been authorised.
The case shows the first two mechanisms, amplified by the stake: selling additional products to existing customers (cross-selling) did not measure whether the customer wanted the product, the accounts were opened by the very people being evaluated, and the targets were tied to pay. The CFPB notes explicitly that the bank did not monitor closely enough how the incentive programmes were being applied. The lesson for a small company is not the size, but the lack of a paired indicator and of checking.
Examples by role: what can go wrong and which indicator balances it
The table below is illustrative, built from the mechanisms above; it does not describe any particular company. The third column shows what can happen, not what will certainly happen.
| Role | KPI with a stake | Possible wrong behaviour | Paired indicator or countermeasure |
|---|---|---|---|
| Shop sales consultant | Sales value | Pressure on customers, unsuitable products, returns | Returns rate, service assessment (mystery shopping) |
| Call centre operator | Average call length | Calls closed before the problem is solved | First-contact resolution, repeat calls within 7 days |
| Hotel receptionist | Additional sales (upsell) | Pushy offers, slow check-in | Check-in time, review score, billing errors |
| Estate agent | Number of viewings | Viewings with unqualified customers | Viewing → offer conversion, exclusive agreements signed |
| Recruiter | Time to fill | Rushed hires, poor fit | Retention at 90 days, the manager’s assessment after probation |
| IT support | Tickets closed | Tickets closed without being solved, split into several | Reopened tickets, user confirmation |
| Warehouse operator | Lines picked per hour | Picking errors, incomplete orders | Errors found at checking or reported by the customer |
| Unit coordinator | Staff cost | Understaffed shifts, undeclared overtime | Customer waiting time, team turnover |
For pharmacies, the rule also has an ethical and legal side: an incentive on selling certain products can conflict with professional regulations. The details are in the article on KPIs for pharmacy employees.
Worked example: how the pair changes the ranking
Hypothetical example. Fictitious data. Two call centre operators, the same month. Ana has an average call length of 4 minutes and solves 55% of problems at first contact. Bogdan has 5.5 minutes and 85%. Targets: 5 minutes (less is better) and 80% first-contact resolution. Each part is capped at 120.
| Calculation | Ana | Bogdan |
|---|---|---|
| Length: target ÷ achieved × 100 | 5 ÷ 4 × 100 = 125 → cap 120 | 5 ÷ 5.5 × 100 = 90.91 |
| Resolution: achieved ÷ target × 100 | 55 ÷ 80 × 100 = 68.75 | 85 ÷ 80 × 100 = 106.25 |
| Score on length only | 120 | 90.91 |
| Score with the pair, 50% / 50% | (120 + 68.75) ÷ 2 = 94.38 | (90.91 + 106.25) ÷ 2 = 98.58 |
| Customers who come back, per 1,000 problems | 450 | 150 |
With a single indicator, Ana looks the best. With the pair, Bogdan moves ahead, and the last row shows why: out of 1,000 problems, 450 of Ana’s customers have to come back, against 150 for Bogdan. The weight matters: if length had 80% and resolution 20%, Ana would have 0.8 × 120 + 0.2 × 68.75 = 109.75, and Bogdan 0.8 × 90.91 + 0.2 × 106.25 = 93.98. A pair with a token weight balances nothing.
Signs that an indicator is being “managed”
- Bunching at the threshold: many more results between 100% and 102% than between 97% and 99%.
- A jump at the end of the period: the last days of the month bring a disproportionate share of the result, and the first days of the following month are weak.
- Divergence: the indicator rises, the result it stands for does not (sales rise, margin or customer retention fall).
- Suspicious uniformity: everyone hits the target exactly, month after month, with smaller variations than real work would allow.
- New complaints: customers or colleagues in another department complain about exactly the thing the indicator does not measure.
Countermeasures that work
Pairs of indicators
Each volume indicator gets a quality one that would fall if the volume were achieved badly. The guide on how to choose KPIs sets out the rule; here the weight is what matters: the pair must carry enough weight to change the level, as the example above shows.
Data from neutral sources and spot checks
Prefer figures from systems (billing, the complaints register, the telephone system) to those reported by the person being evaluated. Where self-reporting cannot be avoided, check a small sample every month and announce in advance that the check exists. An announced check can discourage a deviation before it happens; a hidden one only discovers it afterwards.
Continuous thresholds and caps
A payment proportional to the result, possibly capped, reduces the pressure at a single point. Thresholds remain useful, but avoid one extra percentage point changing the amount a lot. Building stepped bonuses, with quality gates, is covered in the article on thresholds for performance bonuses.
Longer periods for indicators that are easy to move
An indicator accumulated over three months, or the average of the last three months, makes moving things between months pointless. Keep monthly what really happens monthly; move to a quarter what can be put off with a signature.
Indicators for information only
Not everything you follow needs a weight. An indicator visible in the report but carrying no stake stays cleaner and can signal in time that another indicator is being managed.
The five-question test for any KPI with a stake
Put each indicator through the steps below before you tie it to a bonus or a ranking. It takes an hour for a whole scorecard.
- What result do we actually want? Write it in one sentence without figures: “the customer leaves with the problem solved”.
- How could the figure rise without the result rising? Note at least three ways. Ask two people who do the work for help: they are the first to know them.
- Which indicator would fall on each of those ways? That is the candidate for the pair. If there is no measurable one, plan a spot check.
- Who produces the data? If it is the person being evaluated, look for a neutral source or a confirmation from someone else.
- What happens around the threshold? Simulate on past months how the payment changes at 98%, 100% and 102%. For the simulation, see testing a KPI scorecard.
After launch, review the spread of results at three months, looking for the signs above.
Caution. Romania’s Labour Code (Codul muncii, Law no. 53/2003) gives the employer the right to set individual performance objectives and the evaluation criteria (art. 40(1)(f)). The criteria and procedures for professional evaluation are part of the minimum content of the internal regulations (art. 242(i)). When you change a KPI tied to pay, check with the company’s lawyer where the rule is written down (contract, internal regulations, collective agreement) and how it is brought to people’s attention.
Common mistakes
- Sanctioning people for a badly designed indicator. If the whole team “games” the indicator, the problem is the indicator. Correct the rule first, then discuss the individual cases.
- Confusing optimisation with fraud. Putting a sale off by a day is optimisation; opening an account without consent is fraud. The first calls for redesign, the second for the disciplinary procedure, with the prior investigation required by law.
- A pair with a token weight. A quality indicator at 5% does not balance a volume indicator at 95%.
- Too many paired indicators. Ten countermeasures make the scorecard unreadable. One good pair for each volume indicator is enough.
- Irrelevant checks. An audit that checks forms, not the real result, becomes a managed indicator itself.
- Changing the indicator after you have seen the results. Make the change for the following periods and announce it beforehand.
Frequently asked questions
What does Goodhart’s law say, in short?
Goodhart’s law says that a measure used as a target loses its value as a measure. People learn to move the figure by the shortest route, and that route does not always pass through the result the figure was meant to show. The wording used today belongs to Marilyn Strathern (1997), starting from the economist Charles Goodhart’s observation of 1975.
Does this mean we should not tie bonuses to KPIs?
No. It means that a KPI with a stake needs protections that an information indicator does not: a paired quality indicator, data from neutral sources, spot checks and a threshold without an abrupt step. Without these protections, the bonus pays for the figure, not the result, and the difference usually shows only after a few months.
How do I recognise that an indicator is being manipulated?
Look at the spread and the rhythm: many results just above the threshold and few just below it, jumps in the last days of the month, the indicator rising while the real result stagnates. Then compare with an indicator the person does not control, for example returns, complaints or repeat calls from the same customers.
What is a paired indicator?
A paired indicator is a quality indicator set beside a volume one, chosen so that it falls if the volume is achieved badly. For sales, the returns rate; for tickets closed, reopened tickets; for call length, first-contact resolution. The pair works only if it carries enough weight to change the score’s final level.
Is the cobra effect a true story?
The anecdote of the cobras in colonial Delhi has no reliable documentary confirmation, and its historicity is disputed. A case documented in the French colonial archives is that of the rat tails in Hanoi in 1902: the bounty per tail led to rats being left alive and even to rat farms. The idea still holds even if the original anecdote does not.
Can I use a team indicator to avoid individual gaming?
A team indicator reduces internal competition and the pressure on the personal figure, but it does not eliminate gaming: the whole team can move sales between months. Besides, a team indicator with a large weight says little about the person. Use it with a small weight, alongside individual indicators balanced by pairs.
How often should a KPI with a stake be reviewed?
Review the spread of results three months after launch, then at least once a year. Review immediately if new complaints appear in exactly the area the indicator does not measure. A change takes effect for the following periods and is announced beforehand, with a worked example.
What to do next
Take the scorecard for a single role and put each indicator through the five-question test. Wherever you find a way to raise the figure without the result, add the pair or the check and simulate the effect on past months.
- For the difference between activity and result indicators: activity and result indicators.
- For how to communicate the rules so that people trust the score: scores people trust.
- For roles where figures say little: the performance of roles that are hard to measure.