The Head-On Collisions

The objections in Part I were local. Each chapter met one opponent, on one rule, and mostly the rule survived in a smaller form.

This chapter is for the five arguments that are not local. Each of them says the creed is wrong as a whole: not that a particular rule needs a clause, but that the entire method is aimed at the wrong thing, or rests on a premise that is false, or works in a world that is smaller than it thinks. Any one of them, if correct in its strong form, takes the book down.

I have tried to write each of them the way its best advocate would write it. Where I answer, the answer is at the end, after the case has been made rather than woven through it, because an objection interrupted is an objection defanged.

In 1951 Eric Trist and Ken Bamforth published a study of coal mining in Durham that is still, seventy-five years later, the single most instructive thing I know about what goes wrong when work gets reorganized around a better method.

The old way of getting coal was called hand-got. A small team worked a short face together, and every member did every part of the job: cutting, loading, moving supports. The team was self-selected, it was paid as a group, it worked in close proximity, and it held the whole task. If something went wrong the people who could fix it were the people standing there.

Mechanization produced the longwall method, and on every technical dimension it was superior. A long face, worked by a much larger workforce, split across three shifts, each shift performing one specialized part of the cycle. More coal per man, better use of the machinery, a cleaner process. It is exactly what this book recommends: decompose the task, standardize the parts, specialize the roles, sequence the work.

It went badly in a way the technical analysis had no vocabulary for. The whole task was no longer visible to anyone; no person or team could see the cycle end to end. Work that had been done by a cohesive group was now spread across shifts that did not meet. When one shift fell behind, the next inherited the problem with no relationship to the people who had caused it, and the result was blame passed along the chain in the only direction it could go. Absenteeism rose. Supervision had to grow, because supervisors could no longer see the work and so had to be told about it, which required a layer of people whose job was telling.

The output did not deliver what the method promised, and the reason was not resistance to change or poor management. The reason was that the design had optimized a technical system while destroying a social one, and the social one had been carrying load nobody had put in the drawing.

The principle that came out of that work is joint optimization: a work system has a technical dimension and a social dimension, they interact, and optimizing either alone reliably underperforms optimizing both together.

Now read this creed against it. Twelve rules. Every one of them is a technical-system rule. Define the variables. Standardize the method. Convert decisions into formulas. Move knowledge out of the head. Push what matters. Minimize what a person supplies. There is not one term in the entire creed for the social system, and not because I considered it and decided it was covered. Because it did not occur to me.

The objection in its strongest form is therefore not that I am wrong about any rule. It is that I have written a sophisticated, carefully qualified, evidence-checked account of one half of the problem, and that the half I left out is the half that sank the longwall. Every chapter in Part I that discovered an unmeasured loss was rediscovering the social dimension one symptom at a time, from inside a framework that has no name for it. The picture that stopped being formed. The judgment that stopped being exercised. The capability that decayed.

Dave Snowden and Mary Boone's framework sorts situations by whether cause and effect can be known. Where they can, analysis works and rules work, and the creed is at home. Where cause and effect cohere only in retrospect, the sequence has to be probe, sense, respond, and their warning is that leaders who try to impose order in a complex context will fail. Not underperform. Fail.

The creed is a sense-categorize-respond instrument from end to end. Every rule in it presumes that the situation can be identified, that identification determines the response, and that the mapping holds still long enough to be worth writing down.

Charles Perrow supplies the structural version of the same objection and it is harsher, because it applies to systems the creed builds. Systems that are both interactively complex and tightly coupled produce accidents that are nobody's fault and nobody's error. And two of his findings go directly at this method. Adding procedures and controls increases interactive complexity, because each addition is a component that can fail and, worse, interact with other components in ways nobody designed. And tight coupling requires centralized control while complex interaction requires decentralized control, and those requirements contradict, and Perrow holds they cannot be readily resolved.

Standardization is the centralizing move. Automation is the coupling move. This creed is a manual for building systems that are more tightly coupled and more interactively complex than what they replaced, and its central claim is that this makes things less complicated.

James Scott extends it to what happens when the simplification meets the world. His foresters replaced the actual tree, with its vast number of possible uses, with an abstract tree representing a volume of lumber. The abstraction made the forest calculable and manageable and produced real gains and then produced ecological collapse a century later, because everything the abstraction dropped turned out to have been doing something. His term for what the abstraction cannot hold is metis: practical knowledge, particular to its setting, adaptive to shifting circumstances, useless when written down and complete when possessed.

The strong form of this objection is that the creed is a small-scale instance of a well-documented catastrophic pattern. Scott names four conditions for large planning disasters, and the first two are administrative ordering of nature and society and a muscular confidence in technical progress. This book is both of those, applied to a company rather than a country, and the only thing preventing the outcome is the absence of his third and fourth conditions, which are coercive power and a society too weak to resist. Which would make the creed safe not because it is right but because its author is not powerful enough to do the damage.

Gary Klein and his colleagues watched twenty-six fireground commanders, averaging twenty-three years of experience, across a hundred and fifty-six real decision points. In fewer than twelve percent was there any evidence of two or more options being compared. In over eighty percent, the commander recognized the situation as a type, recognized an action as the one that goes with that type, mentally simulated it to check that it would work, and acted.

That is not a shortcut and it is not intuition in the vague sense. It is a different architecture, and it is the architecture that produces good decisions under time pressure from people who have seen ten thousand cases.

The creed has no representation for it. One cannot act upon that which has not been defined is a direct denial of what those commanders do. Reduce it to a quantifiable formula asks for the enumeration of conditions that recognition does not use and could not produce. If it lives in your head it should live in the system is an instruction to convert exactly the thing that cannot be converted.

And the objection has a second edge that hurts more than the first. Formalizing expert judgment does not merely fail to capture it. It degrades it. A commander required to classify before acting is slower, which on a fireground is itself a failure mode. A checklist derived from an expert's practice captures a fraction of what the practice used, because most of what the practice used was never propositional, and the fraction then becomes the standard, and the standard trains the next generation, who learn the fraction and never build the rest.

The strong form: this creed is a method for organizing the work of people who do not know what they are doing, generalized into a universal principle by somebody whose career has been spent building systems rather than exercising a craft. Where real expertise exists, applying it destroys value, and the destruction is invisible because what is lost was never measured and the replacement is documented.

Michael Polanyi's argument is not that some knowledge is difficult to articulate. It is that tacit knowing is a dimension of all knowing. Explicit knowledge is a surface resting on something that cannot be made explicit, and that something is doing work in every act of understanding, including the act of reading an explicit instruction and knowing what it means here, now, in this case, and when it does not apply.

We can know more than we can tell. Not sometimes. Structurally.

If that is right, then the central operation of this book does not do what it claims. Every rule in the creed is an instruction to move something from a person into a structure. Polanyi's claim is that what arrives in the structure is never the thing, because the thing has a tacit component that does not survive the transfer, and what you have moved is a representation that functions only in a reader who already possesses the component you failed to move.

Run that through the creed and everything inverts. A definition works on somebody who already understands the domain. A procedure produces correct action in somebody who already knows when it does not apply. A formula's conditions are observable only to somebody who already knows what counts as observing them. A recorded demonstration teaches somebody who already knows what to attend to. In every case the artifact is a prompt for existing competence rather than a carrier of it, and the organization believes it has banked the knowledge when it has banked a reminder.

Which means the creed's apparent successes are all cases where the tacit component happened to be widely held, and its failures are cases where it was not, and it has no way to tell which situation it is in before it acts.

David Autor's version of this is the empirical footnote: the tasks resisting automation are precisely the ones governed by knowledge nobody can articulate, and the extent of machine substitution for human labor has been overestimated, because the hard part was never the computing. It was that the rule cannot be written down by the person who runs it.

This one does not come from a scholarly tradition. It comes from looking at where the method is being applied most successfully in the world right now, and noticing that it is not being applied by people like me.

Every rule in this book describes a conversion: a decision that a person was making becomes a rule that runs without them. The claim is that this returns something scarce to the person. The observation is that the largest, best-funded, most technically sophisticated implementations of exactly this method are aimed in the opposite direction, converting a person's deliberate attention into an automatic response that serves whoever built the system.

The mechanism is identical. Define the variables, which here means the signals that predict engagement. Standardize and reduce ambiguity, which means a consistent interface that produces a consistent response. Reduce the decision to a quantifiable formula: when X is observed, present Y. Push rather than pull, so the person never has to go looking. Minimize user input while maximizing automated output, so that nothing requires deliberation. Free the mind from the simple task of choosing what to look at.

That is my creed, executed at a scale and level of instrumentation I will never approach, and the outcome is the precise inverse of what the creed promises. There is nothing in my twelve rules that rules it out. The rules describe a technique, and a technique has no preferences.

The strong form of the objection is that this is not an unfortunate misuse to be deplored in a footnote. It is evidence about what the method is. A method whose most effective application in history runs against the interests of the people it acts on is a method whose stated purpose is decoration, and the author's intentions do not enter into it. I have written twelve rules for building systems that decide on people's behalf, and I have written them from the chair of the person who builds rather than the person decided about, and the entire creed reads differently from the other chair.

And the objection is not answered by pointing out that I build for willing colleagues. The longwall was built for willing colleagues. The question is not consent to the project. It is whether the person governed by a rule has any standing with respect to it, and eleven chapters of this book do not raise the question once.

Five objections. Here is what I actually think, after writing eleven chapters against them.

The creed has no social term and it needs one. Not as a caveat. As a peer to the technical rules, because Trist's finding is not that the social dimension is a nice thing to consider; it is that a design optimizing one dimension and damaging the other underperforms a design that optimizes both, on the technical measures.

That reframes every unmeasured loss in Part I as one phenomenon. The classification automation that removed the reason to read a request whole. The dispatcher who stopped building a map of her crews. The operator whose model never formed. Those are not four separate accounting failures. They are the social system absorbing damage from technical optimizations that had no term for it, and the reason each one felt like a surprise is that the framework I was using could not have predicted any of them.

So the creed gains a rule it did not have, and I would put it at the same level as the others: before changing how work is done, name what the current arrangement does for the people doing it, and count that as part of the design rather than as resistance to it. The longwall is what happens when you skip it. Nobody in that story was wrong about coal.

The restriction is real and Chapter 5 accepts it: the creed's methods are ordered-domain methods and they suppress the variation that is the only instrument available where cause and effect cohere only in retrospect. I do not have a reliable test for which domain I am in, which Chapter 18 supplies and which remains the creed's most expensive gap.

Perrow's relocation argument I also accept: standardization does not reduce complexity, it moves it, and the exception class grows as the process class shrinks. The design requirement that follows is that every standard needs an explicit path for the case that does not fit, and the rate of use of that path is the measurement that matters.

What I contest is Scott's frame applied at this scale, and the contest is narrow. His four conditions include coercive power and an incapacitated society, and he includes them because they are load-bearing rather than decorative. A standard imposed on a work group that can complain, leave, ignore it, or route around it is a fundamentally different object from a village resettlement scheme. The feedback that was absent in his cases is present in mine, and feedback is the whole difference between an experiment and a catastrophe.

That is a real distinction and it is also exactly the kind of thing the author of a method would say, so I will put the falsifiable version on the record: if the people governed by a rule cannot override it, cannot report that it is wrong, and cannot be heard when they do, then Scott's analysis applies to that rule at whatever scale it operates, and the creed has produced something it should not have.

Before the concession, one thing the objection gets wrong about its own evidence, because it matters for how far it reaches. Klein's commanders were operating in a domain with unusually fast, unusually unambiguous feedback. A building either behaves as expected or it does not, within minutes, in front of you, repeatedly, for twenty-three years. That is close to ideal conditions for building accurate recognition, and it is why their expertise is real in a way that survives measurement.

Most organizational judgment does not have those conditions. The feedback is slow, ambiguous, confounded by everything else that happened, and frequently absent because nobody follows up. Somebody who has classified requests for twenty years in that environment has twenty years of practice and not necessarily twenty years of learning, because the correction signal that turns practice into skill was never delivered. Klein's finding licenses trusting recognition where recognition had a chance to be trained. It says nothing about recognition trained on noise, and the person holding it cannot tell the difference from inside, which is exactly what Chapter 15's hundred and thirty-six studies found.

Klein's commanders are right and Chapter 2 grants it: inside genuine expertise, under time pressure, on a problem the expert has seen, forcing verbal definition makes the decision worse. The answer offered there is that recognition is an operational definition, with a test and a criterion and a yes-or-no decision, implemented in tissue. What it lacks is transferability, inspectability, auditability, and replaceability, and those four are what an organization runs on.

Where the objection overreaches is in its generalization. The creed's subject is coordinated work among people of varying experience, most of whom are not the twenty-three-year veteran, on problems that recur, where the output has to be consistent across people. That is most work. Klein's finding is about a real and important minority of decisions and it does not license leaving the majority to judgment, which is what it gets used for, and Chapter 15's evidence is that the majority left to judgment loses to a simple formula in about half of the comparisons anyone has run and ties in nearly all of the rest.

The observation is correct. The mechanism is indifferent, the largest implementations run the other way, and my rules do not distinguish them. Chapter 14 offered three questions as a first cut, and they hold: can the person the rule runs on read the rule; does the released capacity return to them or get converted into engagement with the system; can they turn it off today at no cost.

Those three are not a philosophy and they are not nothing. Run them against my morning sequence and all three come out clean: I wrote it, I can read it, it hands the time back, and I can delete it before breakfast. Run them against a feed and all three fail. That is a sharper separation than I expected from three questions anybody can answer in a minute, and the sharpness is itself evidence that the distinction is real rather than a distinction I want to exist.

What I will not claim is that the three questions are a defense. They tell you which side a system is on. They do nothing about the systems on the wrong side, and the person subject to one of those cannot usually apply the test, because the first question already fails: they cannot read the rule.

So the creed gains a second rule it did not have, and like the social-system rule it is a peer rather than a caveat: a rule that governs a person should be readable by that person, and reversible by them, or it is not a formula they installed but a condition they live in. That is a constraint on what this method is allowed to build, stated from the chair of the person decided about, and its absence from Part I is the clearest thing this chapter has shown me.

Polanyi's tacit dimension is right. The system does not hold knowledge. It holds artifacts, artifacts are inert without an interpreter, and every apparent transfer is a case where the interpreter was already present.

What follows from that is less than it appears. The alternative to an inert artifact is not a system that holds knowledge. It is nothing, and a successor rebuilding from zero. Judged against transfer, every knowledge-management effort in history has failed. Judged against the cost of rebuilding, they routinely succeed. The creed has been claiming the first and delivering the second, and the honest version claims the second: what is in your head should leave a trace that lowers the cost for the next person to build their own version.

That is a smaller promise and it survives the objection completely, because it never required the tacit component to move.

All five share a shape, and it took five opponents to see it.

Each objection is about something the creed cannot see rather than something it gets wrong. The social system. The complex domain. The recognition that does not decompose. The tacit dimension underneath the explicit. In every case the method produces a correct answer within its frame and the frame excludes the thing that mattered, and because the thing excluded was never measured, the method's own instruments report success.

That is not a reason to abandon the method. Every method has a frame. It is a reason to hold the frame loosely and to keep somebody looking at what falls outside it, which is the same conclusion the internal contradictions reached from a different direction, and which is now the most heavily supported claim in this book: the creed works, and it requires a permanent, funded, deliberately unautomated human attention to the things it cannot represent.

Which is a stranger conclusion than it sounds, and worth saying plainly before Chapter 18 turns it into a procedure. A method for removing people from decisions ends, after five serious attempts to knock it down, requiring people in the decisions it removed them from. Not the same people doing the same work at the same volume. A small, deliberate, funded remainder, whose job is to be in contact with what the system cannot see.

That is not a contradiction and it took me a while to stop reading it as one. The conversion is still worth making; the returns are large and real and Part I documents them. What changes is the shape of the bill. The creed as I have stated it for years reads as though automation is a purchase: pay once, receive forever. Every serious objection in this chapter says it is a lease, and the payments are in the one currency the automation was supposed to save.

Chapter 18 tries to turn all of this into one instrument. Chapter 19 restates the isms as they stand after Parts I and II, and the creed leaves this chapter with two rules more than it arrived with: the one out of the Durham coalfield, and the one about who a rule is allowed to govern without being readable.