Goodhart’s law describes a measurement problem: when people optimize a metric as a target, the number can become less reliable as evidence of the outcome it was meant to represent. A support team can close more tickets while leaving more problems unresolved. A transformation program can complete every training session while everyday working practices remain unchanged.
The practical question is whether improving the reported number still improves the underlying result. This guide explains the mechanism, compares workplace examples, and provides a KPI-review worksheet you can use before a target becomes part of a scorecard, bonus, or delivery commitment.
What is Goodhart’s law?
A proxy is an observable measure used to stand in for something harder to assess. Ticket closures may indicate service throughput. Training completion may indicate exposure to new practices. Neither directly establishes that customer problems were resolved or that people can use a new capability.
The principle is named after economist Charles Goodhart. Its origins concern the instability of statistical relationships when they are used for control. The familiar broader formulation is attributed to anthropologist Marilyn Strathern:
“When a measure becomes a target, it ceases to be a good measure.”
This wording and its attribution are documented in Michael McIntyre’s University of Cambridge note on Goodhart’s law, which references Strathern’s 1997 article “Improving Ratings”: Audit in the British University System.
In management, treat the principle as a warning to test measurement validity. It does not establish that every target fails, every employee manipulates data, or every KPI should be abandoned. Targets can focus effort. The problem appears when success against the measure becomes an inadequate substitute for success against the objective.
How a useful metric becomes a misleading target
Consider a team whose objective is to resolve customer problems effectively. Managers count closed tickets because closure appears to signal completed work. They then reward higher closure counts. Staff may begin prioritizing easy tickets, closing cases before confirmation, or dividing work differently. The same label—“closed”—can now represent a different quality of service.
- Define the objective: customers receive durable resolutions.
- Select a proxy: count closed tickets.
- Add pressure: link the count to performance judgments.
- Observe adaptation: work selection or closure practices change.
- Check divergence: the number rises without equivalent improvement in resolution.
Metric failure does not always involve cheating
David Manheim and Scott Garrabrant’s Categorizing Variants of Goodhart’s Law distinguishes several mechanisms of proxy failure. These include selecting for noise, pushing beyond conditions where a relationship holds, confusing a proxy with a causal lever, and strategic exploitation. Their taxonomy helps explain why “stop people gaming the system” is an incomplete response.
For a manager, the diagnostic consequence is simple: investigate the measurement system before assigning motives. A team may follow the rules exactly and still improve the wrong thing. Ask what the target encourages people to prioritize, neglect, or redefine.
Goodhart’s law examples at work
The following scenarios are illustrative. They describe risks to investigate, not inevitable consequences of using these metrics.
| Target metric | Possible distortion | Balancing evidence |
|---|---|---|
| Tickets closed | Premature closure or selection of easier cases raises throughput figures. | Repeat contacts, age of unresolved cases, and sampled resolution quality. |
| Sprint velocity | Story-point estimates increase without a corresponding increase in useful delivery. | Delivered outcomes, cycle time, defects, and changes in estimation practice. |
| Training completion | Attendance meets the target while practical capability remains weak. | Observed use, task demonstrations, and support needed after training. |
| Milestones completed | Acceptance conditions are weakened or unfinished work is moved beyond the milestone. | Operational readiness, outstanding dependencies, and acceptance evidence. |
| Customer satisfaction | Survey invitations disproportionately reach customers expected to respond positively. | Invitation coverage, response patterns, complaints, and qualitative feedback. |
A balancing measure needs scrutiny too. A repeat-contact metric can miss customers who give up instead of returning. A defect count can fall because reporting became harder. The purpose is to assemble evidence that can challenge the headline KPI, rather than create another unquestionable number.
Goodhart’s law versus other KPI problems
| Problem | Main concern | Diagnostic question |
|---|---|---|
| Goodhart effect | Optimizing the measure weakens its usefulness as evidence of the goal. | Does improving this number still improve the outcome? |
| KPI overload | Too many indicators consume attention and blur priorities. | Which measures actually inform a decision? |
| Vanity metric | An impressive number provides little evidence of meaningful progress. | What decision or outcome does this number support? |
| Perverse incentive | A reward or penalty encourages behavior that undermines the intended objective. | What behavior becomes rational under these consequences? |
A vanity metric can be weak before anyone targets it. KPI overload can exist even when individual measures are sound. Perverse incentives can accelerate Goodhart effects, but measurement can also fail without deliberate manipulation. For the broader reward-design problem, see The Incentive Trap. For dashboard proliferation, see why organizations create too many KPIs.
Worked example: more tickets closed, more unresolved problems
Hypothetical example: the numbers below illustrate a review method. They are not a reported client case or research finding. Assume the same team and comparable incoming case mix across two four-week periods; verify those assumptions in a real review.
A support manager introduces an individual closure target. The next report shows a higher closure count. Before declaring success, the team checks whether customers return about the same issue.
| Measure | Before target pressure | After target pressure |
|---|---|---|
| Tickets closed | 400 | 500 |
| Closed tickets followed by a same-issue repeat contact within seven days | 40 | 100 |
| Repeat-contact rate | 40 ÷ 400 = 10% | 100 ÷ 500 = 20% |
Count each closed ticket once if it generates at least one qualifying repeat contact. Give every ticket a full seven-day follow-up, including tickets closed at the end of each period. Otherwise, the latest cohort may appear artificially better simply because customers have had less time to return.
Closures increased by 25%, while the repeat-contact rate rose by 10 percentage points. That combination challenges the claim that higher throughput means better service. It does not prove the target caused the deterioration. A product incident, staffing change, or more difficult work could also affect the results.
What should the manager do next?
Sample cases from both periods, including cases with and without repeat contacts. Check closure reasons, customer confirmation, assignment practices, staffing, and issue complexity. Ask the team whether the target changes which work gets attention. Inspect the oldest unresolved tickets as well as completed work.
If the review finds premature closure encouraged by individual quotas, revise the incentive and closure criteria. Retain throughput as operational context, but assess it alongside resolution quality and neglected work. Pilot the revised approach and review the next fully matured cohort before claiming improvement.
This is the kind of evidence challenge described in delivery assurance: a status indicator should trigger examination of commitments and outcomes, rather than replace that examination.
How to reduce Goodhart effects in KPI design
1. State the outcome before choosing the metric
Write a sentence describing the change that matters to the customer or organization. “Customers can complete their task after receiving help” is clearer than “increase closures.” Identify whose outcome counts and the time horizon over which it should hold.
2. Explain why the proxy should work
Document the expected relationship between the number and the outcome. Specify definitions, exclusions, denominators, and data sources. List conditions that could break the relationship. For example, training attendance becomes a weaker capability proxy when completion requires no practice or demonstration.
3. Ask how the target could be met without progress
Invite the people doing the work to identify shortcuts, neglected cases, and reclassification opportunities. Use this as a design exercise, not an accusation. Could a deadline be met by reducing scope invisibly? Could satisfaction rise by changing who receives a survey? These answers reveal where additional evidence is needed.
4. Review the consequences attached to the number
A metric used for team learning creates different pressures from one used for rankings, pay, or public escalation. Decide whether those consequences are proportionate to the metric’s validity and to how much control people have over it. Avoid treating a noisy proxy as a precise judgment of individual contribution.
5. Add focused balancing evidence
Choose evidence that could disconfirm the favorable story: outcome samples, customer observations, unresolved dependencies, or a relevant countermeasure. For transformation management office KPIs, faster decisions deserve examination alongside decision readiness and subsequent rework. Adding every available indicator would make the review harder to use.
6. Check whether the rules are drifting
Review changes in definitions, exclusions, acceptance criteria, and exception handling. An improving trend may partly reflect a changed counting rule. Repeated departures accepted because nothing visibly went wrong can also become normalization of deviance. Record such changes so reviewers can interpret the trend honestly.
7. Connect a warning signal to a decision
Name the owner, review date, and possible responses. Depending on the evidence, the response might be to investigate data quality, revise incentives, restore acceptance criteria, or retire a misleading target. Set triggers using the service’s baseline and consequences; there is no universal percentage that proves a Goodhart effect.
Copyable KPI-review worksheet
Copy this worksheet into your review notes and complete it for one consequential KPI. Start with a measure linked to funding, promotion, bonuses, or delivery approval. Leave unknowns visible and assign someone to investigate them.
Short completed example
Metric: tickets closed per four-week period. Outcome: durable resolution of customer issues. Failure route: premature closure raises the count. Balancing evidence: seven-day same-issue repeat-contact rate, sampled case quality, and aging unresolved work. Trigger: a material departure from the service’s baseline prompts investigation after checking cohort maturity and case mix. Owner: the support manager, with authority to revise closure criteria and recommend incentive changes.
The worksheet is complete when the team can explain what evidence would change its decision. A form filled with reassuring language but no possible corrective action adds reporting effort without improving control.
Frequently asked questions about Goodhart’s law
What is Goodhart’s law in simple terms?
When people focus on improving a measurement, they can make the number look better without improving the thing it was meant to measure. Check the outcome behind the score.
What is a simple workplace example?
A support team closes more tickets to meet a quota, but customers return because their problems remain unresolved. The closure count improves while the intended service outcome may deteriorate.
Who said “When a measure becomes a target…”?
The familiar wording is attributed to Marilyn Strathern. The underlying principle is named after Charles Goodhart. The Cambridge reference below distinguishes the two formulations.
Does Goodhart’s law mean targets are useless?
No. A target can coordinate effort when the measure remains meaningful and the consequences are well designed. Continue testing whether meeting the target produces the intended result.
How does Goodhart’s law apply to Agile velocity?
If increasing story points becomes a performance target, changing estimates can raise reported velocity without increasing useful delivery. Use a team’s velocity cautiously as planning context and assess delivery through outcomes and operational evidence.
Can multiple metrics prevent Goodhart’s law?
They can expose trade-offs, but they cannot guarantee prevention. Several proxies may share the same blind spot or become targets themselves. Independent outcome checks and periodic review remain necessary.
How can I tell whether my KPI is being gamed?
Look for changes in selection, exclusions, definitions, or acceptance standards alongside divergence from customer outcomes. Investigate alternative explanations before attributing intent. A suspicious pattern is a starting point for review, not proof of misconduct.
Look beyond the dashboard. Measures shape attention, decisions, and incentives across an organization. Explore the System Shaping book for a broader perspective on the structures that keep organizational problems recurring.
Sources and further reading
- McIntyre, Michael E. Goodhart’s Law. University of Cambridge. Attribution note referencing Goodhart’s formulation and Strathern’s 1997 article.
- Manheim, David, and Scott Garrabrant. Categorizing Variants of Goodhart’s Law. Submitted 2018; revised 2019. Research preprint on distinct mechanisms of proxy failure.
Method note: the workplace scenarios, numerical support-team case, checklist, and worksheet are original illustrative teaching material. They are not empirical findings or validated assessment instruments. The practical recommendations apply the measurement problem to management reviews and should be adapted to the context.