Why performance measurement drifts away from performance itself, why the drift is usually invisible while it is happening, and why the numbers often look their most reassuring at exactly the wrong moment.
Related advisory: organizational effectiveness consulting and organizational development consulting.
Organisational performance: why performance measurement drifts away from performance itself, and why the numbers look most reassuring at the wrong moment.
Performance is the area where leadership teams are most confident they have visibility, and it is one of the areas where that confidence is hardest to check. A financial control that fails announces itself. An operational failure announces itself. A performance measure that has quietly stopped describing performance keeps producing numbers on schedule, in the same format, in the same meeting, with no change in tone. An instrument that stops working is a nuisance. An instrument that keeps reading while it stops working is something else.
Measurement matters in growing organisations because it is what replaces direct observation. In a small organisation, senior people see most of the work. They know who is carrying which load and which client is unhappy before any system records it. That knowledge does not scale, and the reason it does not scale is structural rather than personal. Jay Galbraith, writing in Interfaces in 1974, set out an information-processing view of organisation design, arguing that task uncertainty is what drives organisational form: as uncertainty and scale rise, more information has to be moved through formal channels because informal ones cannot carry it. Performance measurement is one of those formal channels. It exists because looking is no longer possible.
Four patterns explored in this hub, in brief.
Almost every performance measure in use is a proxy. Response time stands in for service quality. Utilisation…
A measure is never only an observation. Once people know what is counted, the count becomes part of the…
Every performance measure encodes assumptions about the organisation that existed when it was written: how…
Two teams do work that looks the same. One reports consistently stronger numbers than the other, and has done…
Open any heading to read the full discussion.
Almost every performance measure in use is a proxy. Response time stands in for service quality. Utilisation stands in for productivity. Pipeline value stands in for future revenue. Attendance stands in for engagement. Cases closed stands in for the problem being solved. In each pair, the left-hand term is countable and the right-hand term is what the organisation actually wants, and the two are not the same thing.
This is not a design flaw. There is no version of managing at scale that avoids proxies, because the things organisations care about are mostly not directly countable, and waiting for them to become countable means waiting until it is too late to act. The proxy is a reasonable trade. The difficulty is what happens to the trade over time.
At the moment a measure is adopted, the proxy and the outcome are usually well correlated — that correlation is normally why the measure was chosen. Nobody selects a metric that does not track the thing they care about. The problem is that the correlation is a property of a particular moment: a particular volume of work, a particular customer base, a particular set of pressures. It is not a property of the measure. Nothing about the measure guarantees the correlation will hold, and nothing about the measure announces the moment it stops holding.
The most direct treatment of this in the research literature is also one of the oldest. In 1956, V. F. Ridgway published a short paper in Administrative Science Quarterly, the journal of the Johnson Graduate School of Management at Cornell University, examining what happens when organisational performance is reduced to quantitative measures. Ridgway looked at single measures, multiple measures and composite measures in turn, and found that all three could produce undesirable consequences for overall performance. His documented examples are worth stating precisely because they are so ordinary: public employment interviewers evaluated on the number of interviews conducted worked through interviews quickly and placed few applicants; law-enforcement investigators given a quota of eight cases a month selected the easy cases as the month drew to a close. In neither case is anyone acting against instructions. In both cases the instruction has become the work.
Ridgway’s own summary is more careful than most of what has been written since. Quantitative measures, he wrote, are tools and are undoubtedly useful, but indiscriminate use and undue confidence in them result from insufficient knowledge of their full effects, and “the cure is sometimes worse than the disease.”
Two things follow from that paper that are still routinely missed. The first is that adding measures is not a general solution. Ridgway found distortion in composite and multiple measures too, not only in single ones. An organisation that responds to a distorting metric by surrounding it with six more metrics has not removed the distortion; it has distributed it and made it harder to see. This is one reason dashboards so often disappoint. A dashboard changes the density of the display without changing the relationship between any measure and the thing it stands for, and a reader facing forty indicators has less capacity to interrogate any one of them than a reader facing four.
The second is that the drift toward weaker proxies is economic rather than accidental. Activity measures — calls made, tickets closed, interviews conducted, hours logged — are cheap to collect, arrive quickly, and attribute cleanly to one team. Outcome measures are expensive, arrive late, are contaminated by factors nobody controls, and are almost always jointly owned by several units at once. Faced with that asymmetry, organisations do not consciously choose activity over outcome. They choose the measure that can be produced reliably by the fifth working day, and over several such choices the reporting picture tilts toward activity without anyone deciding that it should.
Time makes this worse in a specific way. Effort and measurable results are separated by a lag, and the lag is longest for exactly the work that matters most: capability built, a process made more reliable, a client relationship repaired. Nelson Repenning and John Sterman, of the MIT Sloan School of Management, examined this dynamic in California Management Review in 2001, using system-dynamics modelling alongside more than a dozen case studies of process-improvement efforts. Their central observation is unusually well suited to measurement: nobody is credited for problems that never happened. Work whose success shows up as an absence is structurally disadvantaged in any reporting system, because reporting systems record events.
In practice, organisations often find that the honest response to a proxy that has drifted is not a better proxy. It is an acknowledgement that a particular question can no longer be answered from the reporting pack, and that answering it will require going and looking. That is an expensive conclusion, which is part of why it is reached late.
A measure is never only an observation. Once people know what is counted, the count becomes part of the environment they are working in, and behaviour adjusts to it. This is the second, and larger, half of the performance problem: not that measures describe imperfectly, but that measures act.
The clearest statement of it in the management literature is Steven Kerr’s 1975 paper in the Academy of Management Journal, later reissued as an Academy Classic. Kerr examined reward systems in a manufacturing company and an insurance firm, and set them alongside cases from politics, warfare, medicine and academia, in each of which the stated hope and the actual reward pointed in different directions. His formulation is blunt: “Most organisms seek information concerning what activities are rewarded, and then seek to do (or at least pretend to do) those things, often to the virtual exclusion of activities not rewarded.”
The parenthesis matters, but so does the main clause. There are two quite different behaviours that produce identical-looking numbers, and treating them as one thing is a persistent source of bad conclusions.
The first is gaming: deliberately manipulating a count in a way the person doing it knows is inconsistent with the intent. It is real, and it is generally the reading leadership teams reach for when a measure looks too good. It is also, in most organisations, the less common of the two.
The second is adaptation: a competent professional receiving a clear instruction about what matters and rearranging their work accordingly. Adaptation is not misconduct. It is compliance. When a support team becomes fast at acknowledgement and slow at resolution, the team has read the organisation’s stated priority correctly and acted on it. If the priority was wrong, that is not a fact about the team.
Distinguishing these matters because the two have opposite implications for where the problem sits, and because organisations that read adaptation as gaming reliably respond by adding scrutiny to people who were following instructions. That response has costs beyond the immediate unfairness: it teaches everyone watching that reporting an inconvenient reality is expensive.
There is a related observation from a different research tradition. Erik Hollnagel, Robert Wears and Jeffrey Braithwaite, in a 2015 white paper published jointly by the University of Southern Denmark, the University of Florida and Macquarie University, drew a distinction between work as it is formally imagined and work as it is actually done, arguing that the gap between the two is not a defect but the normal condition of operating systems. Their phrasing is direct: things go well, they wrote, not because people simply follow procedures and work as imagined, but “because people make sensible adjustments according to the demands of the situation.”
A performance measure is a description of work as imagined. It was written at a particular moment by people describing what they believed the work to be. The work as done diverges from that description immediately and continuously, and mostly for good reasons. Over time, a measure therefore tends to become an increasingly accurate description of the measurement system and an increasingly poor description of the work — while its precision, its formatting and its arrival date all stay exactly the same.
This is the point at which the most-repeated phrase in the field becomes actively misleading. “What gets measured gets managed” is generally deployed as an argument for more measurement. Its origin is Ridgway’s 1956 paper, and Ridgway was warning against over-reliance on measures, not recommending them. The popular usage inverts the source. That inversion is not a trivial piece of trivia; it describes how a caution became a mandate, and how an organisation can end up citing the literature in support of the practice that literature was written to question.
How strongly measurement shapes behaviour varies, and the variation is patterned rather than random. The effect is strongest where the measured individual has direct control over the count, where the measure is visibly tied to reward or job security, where the work is high-volume and repetitive enough that small adjustments accumulate, and where no competing professional standard pushes back. It is weaker where outcomes are visible to the person doing the work regardless of what is recorded, where a strong occupational norm supplies its own definition of good work, or where the measure is understood by everyone to be administrative rather than evaluative. Two organisations can run the same metric with entirely different behavioural consequences, and the difference will usually be found in the reward system rather than in the metric.
None of which leaves an easy position. Removing a distorting measure removes the distortion and the visibility together, and an organisation operating at scale without measurement is not operating with better information — it is operating with anecdote, which has its own well-documented biases and no audit trail at all. The trade-off is genuine and it does not resolve cleanly. Anyone claiming otherwise is selling something.
Every performance measure encodes assumptions about the organisation that existed when it was written: how large it was, how the work was divided, who the customers were, what the binding constraint was. Those assumptions are rarely written down alongside the measure, and the measure outlives them. This is the most under-noticed reason that performance reporting degrades. It is not that a measure was chosen badly. It is that it was chosen well, for a different organisation, which happens to have the same name.
Larry Greiner made a version of this argument about organisational structure in Harvard Business Review in 1972, in an article the journal later republished as a classic with new commentary from the author. His claim was that arrangements which resolve one period of an organisation’s development become the source of difficulty in the next, and that managers therefore “experience the irony of seeing a major solution in one period become a major problem in a later period.” Measurement systems follow the same trajectory, with an additional disadvantage: a structure that no longer fits produces visible friction, whereas a measure that no longer fits produces a number, on time, in the usual format.
Growth distorts measurement through several routes at once, and they interact.
Denominators move. A ratio built when the organisation had four teams and one product line describes something different once there are twenty teams and four product lines, even when the calculation is unchanged. Comparison across time silently becomes comparison across different organisations, and the chart offers no warning that this has happened.
Mix changes. Averages conceal composition, and composition is precisely what growth alters. An average handling time that is stable across a period during which the customer base has shifted toward more complex accounts is not evidence of stability. It is evidence of two changes cancelling, and the cancellation is invisible in the reported figure.
Measures accumulate. New measures are added when new questions arise; existing measures are almost never withdrawn, because withdrawal requires someone to argue that a thing currently being watched no longer needs watching, and that argument is unattractive to make and easy to lose. The reporting pack therefore grows monotonically, and attention — which does not grow — spreads thinner across it. Performance data multiplies considerably faster than insight into performance.
Distance grows. Michael Beer and Russell Eisenstat, then both at Harvard Business School, wrote in Sloan Management Review in 2000 about a set of mutually reinforcing barriers to strategy implementation and organisational learning, among them unclear strategy and conflicting priorities, and poor vertical communication. Their central point about these barriers is that they are usually invisible to the leadership team, precisely because the same conditions that create them also prevent them from being reported upward. A performance system that has aged is exactly the kind of thing that gets discussed candidly two levels down and never in the room where it could be changed.
There is also a rhythm problem that has nothing to do with measurement quality. Performance reviews and reporting cycles run on the calendar. Work runs on its own clock — a project that turned in month two, a client relationship that changed after one conversation, a capability that took eighteen months to build. When those two clocks are far apart, a review conducted properly and in good faith will still feel disconnected from the work it is reviewing, and both parties will usually attribute that feeling to the review process rather than to the mismatch.
The reason ageing is so rarely caught in progress is that nothing about it is anomalous. There is no error message. The measure produces plausible values within its historical range, and plausibility inside a familiar range is the strongest defence a broken instrument can have. In practice, organisations tend to discover that a measure stopped describing reality only when something outside the reporting system forces the comparison — a departure, a lost account, an external review, a complaint that arrives with detail the reports do not contain.
Two teams do work that looks the same. One reports consistently stronger numbers than the other, and has done for a year. The conclusion drawn is almost always about people: better manager, stronger team, weaker team, a hiring problem. It is the most available explanation, it is the one the organisation is structured to act on, and it is sometimes correct.
It is also only one of several explanations that fit the same evidence, and the others are harder to see.
The work may not be as similar as it looks. Two teams nominally doing the same thing frequently face different case mixes, different inherited backlogs, different quality of upstream input, and different levels of interruption. The similarity is often an artefact of the categories used for reporting rather than a property of the work.
The measurement may not be as similar as it looks. Where recording depends on human judgement — when a case counts as closed, when an issue counts as escalated, what gets logged as a separate item — two teams following the same written definition can produce systematically different numbers without either departing from it.
The dependencies may differ. Very little organisational work is self-contained, and a team’s reported performance incorporates the reliability of everything upstream of it. This is where the strongest recent evidence sits. Donald Sull, Rebecca Homkes and Charles Sull, drawing on a multi-year study of more than 250 companies and several thousand managers, reported in Harvard Business Review in 2015 that only 9 per cent of managers said they could rely on colleagues in other functions and units all of the time, with around half saying they could rely on them most of the time. Sull is a senior lecturer at the MIT Sloan School of Management; Homkes has been associated with London Business School and the LSE Centre for Economic Performance. Their broader finding was that the central execution difficulty in most firms is cross-unit coordination rather than vertical alignment, which was generally sound. If reliance across units is that uneven, then a substantial and variable share of any unit’s measured performance was determined somewhere other than that unit.
Capability may have accumulated differently. This is where the dynamic Repenning and Sterman described becomes relevant again. Under pressure, managers tend to default to working harder rather than working smarter, because the returns to working harder are immediate and the returns to improving capability arrive late and are attributed elsewhere. Two teams that started in the same position can diverge substantially over a year through nothing more than one of them having had slightly more slack at the outset — and the resulting gap will look like a difference in quality of people, because that is what it looks like from the outside.
What is reported may reflect willingness to report. Amy Edmondson’s research at Harvard Business School on psychological safety in work teams established that teams differ measurably in their readiness to raise problems, and that this difference predicts learning behaviour. Applied to measurement, the implication is uncomfortable: a team that surfaces problems will generate more recorded problems than an equivalent team that does not. Where recorded problems feed into performance assessment, the reporting system is partly measuring candour, and it is scoring it negatively.
Comparison outward carries its own version of the difficulty. Benchmarking against competitors imports not only their results but their measurement conventions, their definitions, their exclusions, and whatever incentive their published figures were produced under. A benchmark that is out of line may indicate a performance gap. It may equally indicate that two organisations count differently, and there is usually no way to establish which from the benchmark itself.
Attribution also has a slow, expensive symptom: people leaving. Departures from growing organisations often cluster among the people who were absorbing the difference between how the system was described and how it actually worked. That signal arrives long after the condition it reports, and it arrives in a form — an individual’s decision, with an individual’s stated reasons — that invites an individual explanation.
The honest position is that the person-or-system question is frequently not resolvable from the reporting alone, and that both the confident structural reading and the confident personal reading are usually running ahead of the evidence. Work published by Paul Rogers and Marcia Blenko, both partners at Bain & Company, in Harvard Business Review in 2006 mapped the recurring points at which decision-making stalls in large organisations — between global and local, between the centre and business units, between functions, and between an organisation and its outside partners — which is a useful reminder that a great deal of what surfaces as a performance problem originates at a boundary rather than inside any of the units being compared. That observation locates the question. It does not settle it, and this page will not settle it either.
The research on performance measurement is unusually consistent about one thing and unusually quiet about another, and the shape of that asymmetry is worth stating plainly.
Where it agrees: measurement changes what it measures. Ridgway established in Administrative Science Quarterly in 1956 that single, multiple and composite measures alike can degrade the overall performance they were installed to improve. Kerr established in the Academy of Management Journal in 1975 that where reward and stated intention diverge, behaviour follows reward. Hollnagel, Wears and Braithwaite, writing across the University of Southern Denmark, the University of Florida and Macquarie University, describe operational work as continuously and legitimately adjusted away from its formal description. Repenning and Sterman at MIT Sloan show why work whose payoff is a non-event is systematically under-recorded and under-resourced. These come from four separate decades and four unrelated traditions — public administration, reward psychology, safety science, system dynamics — and they converge. That convergence is the strongest claim this field supports.
Where it is quieter: none of this work establishes how far any particular measure in any particular organisation has drifted, or how quickly. The literature describes a mechanism, not a rate. Ridgway’s employment interviewers and Kerr’s reward cases show the mechanism operating; they do not license an estimate of how much of any given reporting pack has become self-referential. A body of research that explains why something happens is often read as though it also predicted how often it happens, and that inference is not available here.
Two related bodies of work sit alongside these. Greiner’s argument that the solution of one period becomes the problem of the next supplies the ageing mechanism. Sull, Homkes and Sull supply the cross-unit dependency that complicates attribution, and Galbraith’s information-processing account explains why formal measurement becomes load-bearing at scale in the first place. Beer and Eisenstat, and Edmondson, both address why the resulting problems are slow to travel upward — the first through barriers that are invisible to those best placed to remove them, the second through the differing willingness of teams to raise what they know.
The field also has a specific evidential problem of its own, and it is one worth being explicit about on a page like this. Large round failure statistics circulate widely in this territory and are rarely sourced. Mark Hughes of the University of Brighton examined the most repeated of them — the claim that 70 per cent of organisational change initiatives fail — in the Journal of Change Management in 2011, tracing five separate published instances of the figure and finding no valid or reliable empirical evidence supporting any of them. Hughes was writing about change initiatives rather than performance measurement specifically, and that limit should be respected rather than stretched. But the lesson generalises: a number that everyone repeats and nobody sources is not evidence, and a page arguing that organisations over-trust unexamined figures would be poorly placed to reproduce one. Comparable widely circulated statistics about performance-management failure and dashboard adoption are treated the same way here, which is why they do not appear.
What remains unresolved, and is not resolved by any of the above: whether a given performance gap is best explained by the people in a role or by the system around them. The research narrows that question considerably. It establishes that structural explanations are systematically under-weighted, that dependencies are more consequential than they look, and that reporting reflects candour as well as outcomes. It does not supply a means of settling any individual case, and the papers that come closest are careful not to claim otherwise.
“A measure that has stopped describing reality does not stop producing numbers. That is what makes it difficult.” — Zoom Riddle
Consider an organisation where a support team has met its response-time target every month for three years. The record is unbroken. Over the same three years, customer satisfaction for that same service has declined steadily, and the decline shows no sign of stopping.
The obvious reading is that one of these two things is wrong. Either the satisfaction data is unreliable — small samples, self-selecting respondents, a survey nobody trusts — or something outside the team’s control is driving it. Both are plausible, and both are the explanations most often reached for first, because both leave the operating picture intact.
What complicates that reading is the definition of the target. It measures time to first response. It does not measure time to resolution, and it never claimed to. Over three years the team has become steadily better at acknowledging quickly. It has also become slower at resolving, because acknowledgement is the thing that is counted, and because acknowledging quickly and resolving quickly draw on the same finite attention. Every month the team has done exactly what it was asked to do, and the record proves it.
The deeper observation is not that the team gamed the measure. Nothing here requires bad faith, and reading it as bad faith is the fastest way to reach the wrong conclusion. The measure and the outcome have separated. The team is performing precisely as instructed; the instruction is no longer aligned with the purpose it was written to serve. The organisation has three years of good data about a question it stopped caring about, and no data at all about the question it now has.
That raises a question worth sitting with. When a measure has been stable for years, is that evidence of consistency, or evidence of adaptation to the measure? Both produce the same flat line on the same chart. Nothing in the number itself distinguishes them.
This is an invented composite illustration, written to make an abstract observation concrete. It does not describe a real organisation.
These are open. They are not a checklist, and they are not arranged in an order.
Reflections from real work on organisations growing beyond their founding design.
GROWTH EXPOSES DESIGN
Building commercial infrastructure for an industrial manufacturer preparing for expansion shows that growth does…
EXPOSURE vs PERFORMANCE
Helping an established manufacturer open a secondary revenue channel shows the difference between a performance…
EXPERIENCE vs EVIDENCE
Introducing measurement into a professional-services firm that had grown for decades without needing it taught me…
More on organisational performance: when the measure and the reality separate.
Drift describes the widening gap between what a measure counts and what the organisation wanted it to represent. It happens without anyone changing the measure. The definition stays constant while the work, the customer base, the scale and the pressures around it all change, so the relationship between the count and the outcome weakens. Because the measure continues producing values in its familiar range, drift tends to be identified retrospectively — after something outside the reporting system contradicts the reported picture.
Because the reporting system is generally the last place a measurement problem shows up. If a measure has drifted, it will keep reporting acceptable values by construction. What surfaces instead is usually a second-order effect: unexplained turnover, a client relationship that ends without prior signal, a quarter that misses despite green indicators, or a persistent gap between what senior people believe about a team and what people close to that team describe.
It is ambiguous evidence. Long-term stability is consistent with genuine consistency and equally consistent with the organisation having adapted to the measure, and the two produce identical charts. Stability tells you the measured quantity has not moved. It says nothing on its own about whether that quantity still corresponds to the outcome it was chosen to represent.
Not reliably. Ridgway’s 1956 work in Administrative Science Quarterly found distortion arising from multiple and composite measures as well as from single ones, so adding indicators does not neutralise the underlying problem. Additional measures also spread a fixed amount of attention more thinly, which tends to reduce how closely any individual measure is interrogated. Volume of performance data and quality of insight into performance are largely independent of each other.
Usually not, and the distinction is consequential. Gaming means deliberately manipulating a count in a way the person knows is contrary to intent. Far more common is straightforward adaptation: people receive a clear signal about what is valued and reorganise their work around it, which is compliance rather than misconduct. Treating adaptation as gaming misplaces the problem and tends to discourage the honest reporting an organisation most needs.
Partly because reviews run on a calendar while the work runs on its own timescale. Judgements formed months earlier are discussed against a record compiled for administrative purposes, and the events that mattered most may fall between cycles or leave no trace in the reporting at all. The disconnect is frequently attributed to the review format, when its source is the mismatch between the two clocks.
It introduces its own difficulty. A benchmark imports the other organisation’s definitions, exclusions and incentives along with its numbers. A figure out of line with a benchmark may reflect a genuine performance difference or a difference in counting conventions, and the benchmark alone rarely allows a reader to tell which. External comparison is useful as a prompt to look more closely; it is weak as evidence on its own.
Ridgway, V. F. (1956). “Dysfunctional Consequences of Performance Measurements.” Administrative Science Quarterly, 1(2), 240–247. Published by Sage on behalf of the Johnson Graduate School of Management, Cornell University. DOI 10.2307/2390989.
Kerr, S. (1975). “On the Folly of Rewarding A, While Hoping for B.” Academy of Management Journal, 18(4), 769–783. Updated as an Academy Classic in Academy of Management Executive, 9(1), 1995, 7–14.
Greiner, L. E. (1972). “Evolution and Revolution as Organizations Grow.” Harvard Business Review, 50(4), 37–46. Republished as an HBR Classic, 76(3), May–June 1998, 55–68, with new commentary by the author.
Galbraith, J. R. (1974). “Organization Design: An Information Processing View.” Interfaces, 4(3), 28–36. DOI 10.1287/inte.4.3.28.
Edmondson, A. C. (1999). “Psychological Safety and Learning Behavior in Work Teams.” Administrative Science Quarterly, 44(2), 350–383. DOI 10.2307/2666999.
Beer, M., & Eisenstat, R. A. (2000). “The Silent Killers of Strategy Implementation and Learning.” Sloan Management Review (now MIT Sloan Management Review), 41(4), 29–40.
Repenning, N. P., & Sterman, J. D. (2001). “Nobody Ever Gets Credit for Fixing Problems that Never Happened: Creating and Sustaining Process Improvement.” California Management Review, 43(4), 64–88. DOI 10.2307/41166101.
Rogers, P., & Blenko, M. W. (2006). “Who Has the D? How Clear Decision Roles Enhance Organizational Performance.” Harvard Business Review, 84(1), 52–61.
Hughes, M. (2011). “Do 70 Per Cent of All Organizational Change Initiatives Really Fail?” Journal of Change Management. University of Brighton.
Hollnagel, E., Wears, R. L., & Braithwaite, J. (2015). From Safety-I to Safety-II: A White Paper. Published jointly by the University of Southern Denmark, the University of Florida and Macquarie University.
Sull, D., Homkes, R., & Sull, C. (2015). “Why Strategy Execution Unravels — and What to Do About It.” Harvard Business Review, 93(3), 57–66.
Services can be delivered remotely, and we accept requests from organisations wherever they are located.
