DAY 14 · ORGANIZATIONS AND INSTITUTIONS

Organizations & Institutions

The most counterintuitive line here: a department is often created not because it is useful, but because everyone else has one

2026-09-07 · BigCat

The process a company follows most diligently may be one almost nobody believes in — and removing it still costs more than keeping it.

If you have changed employers, the first week at the second one comes with an odd familiarity: the performance-review form looks much the same, quarterly goals are written much the same way, even the number of "values" on the wall is about the same. The two firms are in completely different businesses and differ several times over in headcount. Who mandated any of this? Nobody did.

This page takes apart three things: why organizations doing completely different work grow to look alike; why what is written in the documents and what people actually do stay out of alignment for years, and why that is not necessarily bad; and how rules turn from means into ends.

// 01

You could read the new company's org chart on your first day

Start with a piece of real history. Between 1880 and 1935, American cities progressively adopted civil service systems — hiring by competitive examination, protection of tenure — replacing the old practice where whoever won arranged jobs for their own people. In 1983 Pamela Tolbert and Lynne Zucker laid out the dates on which each city adopted and asked a plain question: what predicts when a given city adopts?

The answer splits in two. Among the early adopters, a city's own characteristics predict it: how large it is, how many immigrants it has, how complicated its administration has become — back then it went in because it was genuinely needed. The later the adoption, the less those characteristics predict anything: whether a city adopted came loose from its own size and its own troubles.

So why did the late ones adopt? Because by then "a proper city government has a civil service" had become common sense that needed no argument. Early adoption tracks need; late adoption tracks respectability.

The term for it: organizational sociology calls this "organizations in the same circle growing more alike" institutional isomorphism, and the circle is a field — not only competitors, but regulators, big customers, certifying bodies, the people who hire you and the people you hire.

Paul DiMaggio and Walter Powell's 1983 paper "The Iron Cage Revisited" separates three routes into convergence:

Forced: law, audit, a large client's compliance demands make you grow the matching department and paperwork.
Copied: when goals are vague, when nobody can say what counts as doing it right, when success is hard to attribute, copying a peer who looks successful is the cheapest move available.
Carried in: people trained in the same profession move between firms carrying the same sense of how things are done. Business schools, trade associations and certification exams all do this.

The second route deserves the most attention, because what triggers it is uncertainty, not effectiveness. You don't know which practice is right, so you copy; having copied, you and the firm you copied together become the evidence that everyone does it this way. That is how an entire industry can march in step toward a practice nobody ever tested.

The criticism belongs here too. The complaint this account most often draws is that it is too accommodating: convergence is legitimacy pressure, divergence means the field is immature, and it is hard to refute. It also explains similarity better than it explains why some organizations refuse to follow. Evidence: the Tolbert and Zucker study is a correlational analysis of historical archives, not an experiment — though the pattern of early adoption driven by need and late adoption driven by respectability recurs across many later diffusion studies.

// 02

What is written in the documents and what people actually do were never one thing

In 1977 John Meyer and Brian Rowan made a claim that still stings: much of an organization's formal structure is not there to get the work done, it is there to be seen from outside — it demonstrates that this institution is a proper one. Their term is myth and ceremony; the relationship between formal structure and actual activity, hanging together but not meshing, is called decoupling.

Around the same time Karl Weick used schools to explain loose coupling: the timetable, teaching credentials and evaluation procedures are all in place, yet once the classroom door closes, how a teacher actually teaches and whether the students actually learn is only slackly tied to any of it — most of the formal instruments in a principal's hands do not reach inside the classroom.

A case you can check. In 2001 James Westphal and Edward Zajac went through the stock repurchase programs announced by large American firms in the late 1980s and early 1990s and found that a substantial share never actually bought the shares back — while the market gave a positive reaction to the announcement itself. Formal adoption happens first; substantive implementation is a separate matter, and sometimes it simply does not happen. Evidence: correlational analysis of large-sample archival data; do not read it as "firms are all lying".

The upper layer is for outside eyes; the lower layer is where the work actually gets done. The links between them are dashed — not severed, just slack: swap the upper layer and the lower one often carries on unchanged. The pressure presses on the formal layer, and it can expand without the lower layer's consent.

The easiest misreading here is the moral one. Decoupling can of course be hypocrisy, but it also does work: external demands frequently contradict each other and frequently have nothing to do with the actual task, and hanging them on the formal layer shields the people doing the work. Loose coupling also keeps a local failure from dragging the whole chain down with it.

The cost sits on the other side: the formal layer grows on its own. It has dedicated staff, a budget and a motive to prove it should exist, so procedure only accumulates. Patricia Bromley and Walter Powell pointed in 2012 to a more awkward version: what is common today is not policy failing to land but means decoupled from ends — a great deal of activity is performed in earnest, and nobody goes back to check whether it produced the result it was meant to. Unimplemented policy is at least visible; activity that lands but has nothing to do with the outcome looks like everything is fine.

// 03

The rule started out as a means and turned into the end

Robert Merton supplied the term in 1940 and it is still in use: goal displacement. For the sake of reliability and fairness an organization asks its members to follow rules strictly; follow them long enough and compliance itself becomes the thing to be defended, while the original goal recedes into the background. The clerk at the counter who turns you away saying "those are the rules" is telling the truth — for them, that really is the job. To be clear about status: this is an argument, not an empirical study.

Now a real piece of fieldwork. From the late 1940s into the early 1950s, Alvin Gouldner and his students carried out long observation at a gypsum mine and its processing plant in the northern United States, publishing in 1954. What he saw was not how strict the rules were, but that how far rules are enforced is itself an exchange: the old manager looked the other way — lateness, taking a bit of company material home — the rules were there in black and white but nobody pressed them, and workers repaid it with cooperation. Gouldner called this the indulgency pattern. Then a new manager arrived and went by the book. The result was not higher output but a wildcat strike. Rules that go unenforced are not waste paper; they are held in reserve as bargaining chips. Evidence: a single-plant ethnography, not a general law.

The contemporary version is more familiar. In 2007 Wendy Espeland and Michael Sauder reported several years of interviews with deans and admissions staff at American law schools, looking at how one public ranking remade the schools. Schools began managing themselves backwards from the ranking's algorithm: scholarships shifted from financial need toward test scores, because scores enter the ranking; when graduate employment numbers looked bad, schools funded temporary positions themselves to place people first. They call this reactivity — once a measure is publicly used to evaluate people, the people being evaluated change their behavior to meet it, and what the measure measures changes. Evidence: multi-year qualitative research from interviews and archives.

So is the conclusion "bureaucracy is bad"? No. The Weber and rationalization topic covered the iron cage; here is the other side: take the rules away and what usually returns is not freedom but personal dependence and arbitrary discretion — whoever is closest to the boss decides — which is precisely the disease civil service reform was invented to treat. The iron law of oligarchy that Robert Michels drew in 1911 from the German Social Democratic Party points the same way: even an organization with the most democratic program in its charter grows a leadership that is hard to replace once scale requires full-time staff (a strong claim from one case, much qualified by later work).

So the real question is not whether to have rules, but which matters to hand over to them. Rules are good at things that repeat, whose consequences are clear, and that must be treated alike; they are bad at things that require judgement on the spot — and an organization's default drift is to stuff the latter into the former, because the former can be written into a report.

🎒 In practice · BigCat

① Engineering and system design — at an architecture review someone proposes bringing in a component or a process, and the reasons boil down to "the big companies all do it this way". Running here is the mechanism from "You could read the new company's org chart on your first day": the greater the uncertainty, the more copying becomes the cheapest source of legitimacy, and what you copy arrives carrying somebody else's constraints. What to change: require the proposer to write down at what data volume, what failure rate and what team size the copied firm introduced it — if two of those three numbers are an order of magnitude away from yours, demote it to "later". Stop using "industry standard" as an argument.
② Leading a team and organizing work — once the quarterly goals are written, what actually decides what people do is the handful of requests that arrive mid-quarter; the document and the reality are two separate layers, and nobody dares skip writing the document. Running here is the mechanism from "What is written in the documents and what people actually do were never one thing": decoupling has a function, and forcing the two layers into one usually will not budge. So the thing to change is not making the document match reality, it is keeping the formal layer from eating more resources: turn quarterly goals from "fill the page" into "at most three, each with a number checkable by quarter end", log the mid-quarter arrivals in a separate column, and at quarter end count how many person-days they took. Watch that number, not the completion rate.

🌀 Crossing over · connections beyond the field

Distributed systems and AI · hierarchy buys down communication, and pays for it with compression at every hop: for a set of machines to agree on one thing, the number of direct channels grows with the square of the machine count, which is why real systems almost always elect a coordinator and squash the square into a line. An organization growing layers is the same arithmetic: fifty people aligning pairwise is a disaster, split into squads each with a lead it is not. Running here is the mechanism from "The rule started out as a means and turned into the end": every hop restates what it passes upward in the vocabulary its own layer is measured by. From which follows a conclusion you would not otherwise see: distortion does not add across layers, it multiplies — lose only a fifth at each layer and under a third of the original reaches the top after six. So flattening genuinely reduces distortion, but it pushes the fan-in onto the single person at the apex; the two cannot be optimized together, and the shape of that trade-off can be computed rather than argued about. In observable terms: to find out how distorted a reporting chain is, don't ask people — compare what one event looks like in the raw record at the bottom and on the dashboard at the top. Where it breaks down: in distributed systems messages are copied bit for bit, the loss happens in aggregation, and the noise is roughly unbiased; in an organization every hop's incentive points the same way — bad news filtered on the way up, progress rounded up. So the error-correcting move of "redundant paths plus a vote" fails here: three reporting lines bent by the same incentive, and the vote only confirms the bias a second time.
Buddhism · an organization whose output cannot be verified can only be trusted through procedure: in the monastic code of the sangha, the Vinaya, the largest share of text is not about what counts as awakening but about procedure: how the uposatha recitation of the rules is conducted, how many are needed for a saṅghakamma decision, who is qualified to attend, which process settles a dispute. An organization that has lasted more than two millennia has almost all of its rules grown around "how a decision is made" rather than "what counts as having succeeded". Running here is the mechanism from "What is written in the documents and what people actually do were never one thing", used in reverse: when the core output cannot be verified from outside, credibility can only be hung on formal structure. From which follows a claim you can go and check: how elaborate a rulebook is can be predicted from how hard the output is to inspect — religion, education, consulting, peer review, places where nothing can be checked on the spot, grow their rules on procedure; conversely, an organization whose output is plainly measurable yet whose procedures are minutely specified is most likely handling not efficiency but the assignment of blame. In observable terms: open an organization's handbook and measure how much space goes to "how decisions are made" versus "what counts as done". Where it breaks down: decoupling theory assumes the formal layer is mainly performed for an outside audience, whereas the Vinaya's first audience is the sangha itself, a breach has concrete consequences, and form and practice are tightly coupled. More fundamentally, the rite is not a shell that proves the practice — it is part of the practice, so the form/substance distinction loses its purchase here.

Going deeper

If convergence is copying, how do you tell it apart from "everyone independently found the same optimum"?

Two traces you can look for: whether the correlation between adoption timing and internal need decays over time (the test Tolbert and Zucker used), and whether performance shows an inflection after adoption. Neither showing up makes imitation the cheaper explanation. The difficulty is that both tests are hard to run cleanly — performance often has no comparable control, and "internal need" is itself defined by the same discourse.

Once a measure is made public, is it bound to be gamed?

Not bound to, but the conditions for resisting are demanding: it measures a final outcome rather than an intermediate action; the data does not originate with the people being evaluated; several mutually independent sources can be reconciled against each other. Drop any one and manipulating the number costs less than doing the work. So the real design question is not "is the measure accurate" but "is gaming it more expensive than meeting it".

Why can an organization remember what people cannot? And what does it take for one to forget something?

Institutions are external memory: procedures, archives, posts — the person leaves, the post remains, and the new arrival grows back into the shape of the previous one by following the job description. So for an organization to forget something you usually have to remove the post or the form that carries it; swap people while leaving the structure alone and it will not be forgotten. This also explains why abolishing a useless process is so much harder than adding one.