Why this matters
Consistency across sites is usually treated as an execution question: did people follow the process. It is more accurately a documentation question: does the process, as written, still describe the work. Every site that has been operating for more than a short time has accumulated small departures from its own manual — a step done out of sequence because it is faster, a check added because a supervisor once got burned by skipping it, a workaround for a system limitation that was never updated in the guide. None of this is misconduct. It is simply what happens when documented instructions meet ongoing, variable, local conditions. The organisation experiences the result as inconsistency between sites. What is actually varying is the distance between the document and the work at each one.
This matters because the instinctive response — retrain, re-audit, reissue the standard — treats the document as the reference point and the site as the thing that needs correcting. Where the document itself has drifted furthest from reality, that response repeats the same gap it is trying to close.
A common misconception sits underneath most of this: the belief that a written standard, once issued, describes a fixed and stable thing until someone deliberately changes it. In practice a standard is closer to a moving average of everyone’s small adjustments to it, most of which are never recorded and most of which nobody involved would describe as a deviation. The person who resequences two steps because the second is faster to do first is not thinking of themselves as off-standard; they are thinking of themselves as competent. Multiply that by every site, every shift and several years, and “the standard” becomes a label attached to several related but distinct sets of actual behaviour.
How visible this is also depends heavily on context. In a manufacturing or logistics setting, where output is physical and countable, a gap between sites tends to surface quickly, in scrap rates or rework. In a professional-services or knowledge-work setting, the equivalent gap is harder to see, because the deliverable is judgement-based and a departure from the documented method can look identical to a good outcome for a long time before it is noticed. Neither context is exempt from the pattern; they simply differ in how long it stays hidden.
The standard on paper is not the standard in use
Consider the scenario “The Procedure Nobody Follows.” An audit finds a documented procedure that is not being followed by any of the staff responsible for it. The first assumption is a training or compliance failure. What complicates that reading: following the procedure as written takes roughly three times as long as the method staff actually use, and produces the same outcome. The written version was built for a system replaced four years earlier. Nobody updated the document when the system changed, because updating the document was never part of changing the system — the deeper observation being that the document and the work diverged the moment practice moved and the record did not move with it.
This is not a story about non-compliant staff. It describes a structural pattern that researchers studying safety and operations call the difference between work-as-imagined and work-as-done. Erik Hollnagel, Robert Wears and Jeffrey Braithwaite, writing across their work on this distinction, argue that things generally go well in organisations not because people simply follow the procedure as written, but because people make sensible, continuous adjustments to meet the actual demands of the situation in front of them. The written procedure is a snapshot of a moment, frozen at the point it was authored; the work keeps moving. Two sites diverge from each other, in other words, for exactly the same underlying reason that either one diverges from its own manual — because both are adapting to conditions the document does not capture, and they are not adapting in the same direction.
Framed this way, “inconsistency between sites” and “drift from the documented standard at a single site” are the same phenomenon observed at two different scales. A site audited in isolation looks like a compliance question. The same site compared against another site running the identical manual reveals that the manual was never the operating reality at either location — only the degree and direction of the departure differ.
This creates a genuine tension rather than a simple problem to be closed. Tightening the standard — adding detail, adding checkpoints, reducing the room for local judgement — reduces the space in which sites can quietly diverge, but it also reduces the space in which a site can make the kind of sensible, situation-specific adjustment that the Hollnagel, Wears and Braithwaite research treats as the reason work goes well at all. An organisation that eliminates local adaptation entirely does not eliminate the gap between the document and the demands of the situation; it simply removes the mechanism that was quietly closing that gap in an ad hoc way, and the pressure has to go somewhere. There is no version of a written standard, however detailed, that removes the need for judgement at the point of work — only versions that make clearer where judgement is expected and where it is not.
What makes one site diverge further than another
If every site adapts, the interesting question is not whether a gap exists but why it opens wider in some places than others. Some of that variation is a matter of information load rather than intent. Jay Galbraith’s work on organisational design describes how rising task uncertainty increases the amount of information a unit has to process to keep functioning, and how organisations cope through mechanisms such as self-contained tasks, vertical information systems and lateral relations between units. A single-site business processes most of that uncertainty through direct, face-to-face correction — someone notices a departure and adjusts it on the spot, informally, often within the same day. A multi-site business has to move the same corrective information across distance, through fewer and slower channels, so a departure that would be caught immediately at one site persists for months at another simply because nobody in a position to notice it happened to be looking.
A second source of divergence sits in how each site is measured. V. F. Ridgway’s early and still widely cited work on the dysfunctional consequences of performance measurement found that single, multiple and composite measures can all distort the behaviour they are meant to track — documenting, for instance, interviewers evaluated on volume who conducted faster interviews and placed fewer candidates. The same logic applies across sites operating under one shared target: wherever the measured outcome and the documented method can be separated, some sites will satisfy the measure by a route the manual never described, and the divergence will be invisible for as long as the number looks acceptable. In practice, organisations often find that the site closest to the written standard is not the best-performing one, and the best-performing site is quietly furthest from it.
Why the gap widens as the organisation grows
Left alone, this pattern does not stay still. Nelson Repenning and John Sterman’s research on what they call the capability trap describes why: under sustained pressure, managers systematically default to working harder rather than working on the underlying process, because the people who quietly prevent a problem from recurring rarely receive visible credit for the problem that never happened. A site that has developed a genuinely better local practice — an undocumented check, a resequenced step — gets no formal recognition for it, because officially the site is simply following the standard everyone else follows. When the person who built that adaptation moves on, the adaptation frequently leaves with them, and the site reverts toward whatever the document actually says, which may by then be the weaker version.
Growth compounds this in a second way, independent of pressure: it adds distance. Each additional site, region or reporting layer increases the number of steps between the place a standard is authored and the place it is executed, and each step is an opportunity for a small, reasonable, locally sensible adjustment. None of those adjustments look like a decision to deviate. Each one looks, from where it is made, like common sense. It is only in aggregate, compared across sites, that the pattern becomes visible — and by the time it is visible, it has usually been accumulating for years, not weeks.
The shape of the gap also changes as an organisation grows, not just its size. In a young multi-site business, most sites were opened recently, by people who were trained directly by the founder or an early team, so the sites tend to resemble each other even where they diverge from the document, because they diverged from the same source in similar ways. In an older, larger network, sites were opened at different points in the organisation’s history, under different versions of the standard, trained by different people who were themselves several generations removed from the original document. At that stage the variation is not simply “drift from the standard” in one direction; it is several distinct, internally consistent local versions of the standard, each defensible on its own terms, that happen to share a title page.
What the research shows
A multi-year study of more than 250 companies and several thousand managers, led by Donald Sull of MIT Sloan alongside Rebecca Homkes of London Business School and the LSE Centre for Economic Performance, found that only 9% of managers said they could rely on colleagues in other functions and units all of the time, with roughly half saying they could rely on them most of the time. The same research located the central execution problem in cross-unit coordination rather than in vertical alignment between levels of management — most firms, it found, already have reasonably sound processes for cascading a strategy downward; what breaks down is coordination sideways, between units operating in parallel. Sites are, in this sense, a specialised case of units: the same coordination gap that shows up between functions inside a single building shows up between locations that rarely interact at all.
Taken together, the research points in one direction rather than several. Galbraith’s account of rising information-processing demand explains why distance between sites widens the gap. Ridgway’s work on measurement explains why a shared metric can mask rather than close it. Repenning and Sterman’s capability trap explains why a site’s better practice does not travel and does not persist once its author leaves. And the Hollnagel, Wears and Braithwaite distinction between work-as-imagined and work-as-done explains why the document itself was never a complete description of the work at any single site, before a second site is even introduced. None of these bodies of work were produced to study multi-site organisations specifically, which is itself worth noting: the pattern is not an artefact of any one industry or business model. It is closer to a structural property of documented work carried out by more than one group of people.
What the research does not settle is how much local adaptation an organisation should tolerate before it stops being adaptation and starts being a genuine loss of consistency — that line appears to be organisation- and context-specific rather than a fixed threshold, and none of the cited work claims otherwise. What it does establish reasonably firmly is that the gap is not primarily a discipline problem. It is closer to an information and measurement problem: how quickly a useful local adjustment becomes visible to other sites, how quickly a harmful one does the same, and how much the measurement system in use can distinguish between the two.
The same pattern appears across many kinds of organisation; see Other Sectors for how it is read elsewhere.
Questions worth considering
- How would we know whether our documented standard is being followed as written, rather than assume it is because nobody has reported otherwise?
- What does the strongest-performing site currently do that has never been written down anywhere?
- When a site is measured against a shared target, is there a route to hitting that target that the manual does not describe?
- Who currently owns the gap between the written standard and the standard in use — and is that different from who owns the document itself?
This guide sits within Operations. A related guide: The Measurement That Outlived What It Measured.
Your team is facing similar challenges?
Subscribe to The Riddle Ledger
Perspectives are sent as they are published, with no promotional email, and you can unsubscribe at any time.