The standard you cannot see is not a standard

A site starts logging a particular class of fault as priority two rather than priority one. The reasoning is sound. Their territory is spread out, a priority one clock cannot be met once travel time is counted, and a breached target reads worse to central than a lower grade does. So the threshold moves by a notch, and the work gets done at the same speed it always did.

Nobody hides this and nobody announces it. Three years on, every site has its own threshold, arrived at honestly and never written down. Priority one volumes look stable. Response-time compliance looks healthy. The staffing model built on those numbers is wrong in a different direction at every site, and the error is invisible precisely because every figure in it is accurate.

This is not a quirk of any one sector. It turns up wherever central owns the definitions and the field owns the execution: retail chains, bank branch networks, hospital groups, franchise operators, engineering service organisations. The goods differ. The gap between the written process and the performed one behaves the same way.

Researchers who study this behaviour call it a workaround, and they define it more narrowly than the everyday use of the word: a goal-directed deviation from a standard process, undertaken to get the work done. The qualifier does real work. Mistakes are not workarounds, because mistakes are not aimed at anything, and neither is fraud, which has a different goal entirely. What is left is a category of deviation that carries information. Someone with local knowledge decided the prescribed route would not produce the required outcome and acted on that judgement. They may have been wrong. The point is that a judgement was made, by someone in a position to make it, and the organisation has no record of it.

A study in Business and Information Systems Engineering traced thirteen such workarounds across four organisations. One of them involved a procurement process whose approval sequence ran longer than the production schedule could absorb, so staff ordered directly from suppliers they already knew and skipped the offer stage. The process documentation described a governed sequence for orders above a threshold. The event log recorded something else. Production kept to schedule, which was the point.

Forbidding this does not remove it. The conditions that produced the deviation are still there after the policy is issued, so the deviation continues, minus the record. An organisation in that position has swapped a variation it could observe for one it cannot, and the compliance rate improves either way.

Where the evidence usually dies

The priority example is one instance of something more general, and the general version is worth naming because it is quiet.

Any organisation running the same work across a large field has to classify things: faults, assets, accounts, cases, customers. Whatever the unit of work is, some scheme sorts it, and the sorting drives how it is handled, staffed and priced. Edge cases need local judgement, which is why the scheme has people applying it rather than a rule engine. Judgement then does what judgement does, and settles into a convention.

Nothing about this appears in reporting, because reporting aggregates. Central receives counts by category, and counts by category are accurate. What central does not receive is the reasoning that put each item into its category, and that is precisely where the divergence lives. The classification step is the last point at which the difference is visible, and it happens below the line at which anyone looks.

Which is how a divergence survives for years without concealment, without dishonesty, and without any control failing in a way a dashboard could show. Every figure is correct. The definitions underneath them stopped being shared some time ago.

An answer from outside management

Elinor Ostrom spent her career studying how communities govern shared resources: irrigation networks, inshore fisheries, forests, groundwater basins. She was awarded the Nobel prize in economics in 2009 for the work. None of it concerns companies, which is part of why it is useful, because the findings were not shaped by anyone’s view of how a corporate operating model ought to look.

From field cases across several continents she drew a set of design principles separating arrangements that endure from arrangements that collapse. Rules have to fit local conditions, because a rule written around one set of physical circumstances does not survive transplanting into different ones. The people a rule binds should have a hand in writing it, a pattern she found in the systems that lasted generations and found missing in those that failed. And authority should sit in nested layers rather than a single centre, an arrangement she called polycentric. The same instinct sits behind context-driven frameworks in our Insights library, which sort decisions by the kind of problem rather than applying one method everywhere. Public administration has a companion idea, subsidiarity, in which a decision defaults to the lowest level capable of taking it and rises only when that level has been shown to be inadequate.

Taken together those change what the standardisation question is. Asking how much variation to tolerate produces an argument that nobody wins, because both sides are describing real costs. Asking where the authority to write this particular rule belongs produces an answer, and different rules get different answers.

Making exceptions legal

The practical version is unglamorous, which is some of why it gets skipped.

Hold central to outcomes, definitions and the interfaces between sites, and leave method to the people who can see the building. Central can defend a definition across two hundred locations. It cannot defend a physical method across two hundred locations it has never visited, and the attempt is what generates the invisible deviation in the first place.

Then give exceptions a legal route. An exception acquires an owner, a written reason and an expiry date, and it goes into a register that somebody actually reads. Most of them will be reasonable. When the same exception turns up in five places, that is not five breaches to close out, it is evidence that the standard is wrong and that five sites have already worked out how. The same logic applies to gates that never reject anything, which we covered in when stage-gate stops killing things.

The measure that tells you whether any of this is working is not the compliance rate. It is the spread between your strongest and weakest sites on the same process, and whether that spread is closing. Compliance can sit at ninety-five percent while the spread widens, which usually means the good sites improved without help and the standard contributed nothing. That number is harder to look at, and it is the one that answers the question.

None of this is free. It requires central to build the mechanism by which it will be told, regularly and from several levels down, that it got something wrong. A compliance rate never says that, which is most of its appeal.