If you haven’t read any of these yet, the gist is that I’m writing a book about mental models and writing these notes up as I go. You can find links at the bottom to the other frameworks I’ve written. If you haven’t already, please subscribe to the email and share these posts with anyone you think might enjoy them. I really appreciate it.
The vast majority of the models I’ve written about were ones that I discovered at one time or another and have adopted for my own knowledge portfolio. The Variance Spectrum, on the other hand, I came up with. Its origin was in trying to answer a question about why there wasn’t a centralized “system of record” for marketing in the same way you would find one in finance (ERP) or sales (CRM). My best answer was that the output of marketing made it particularly difficult to design a system that could satisfy the needs of all its users. Specifically, I felt as though the variance of marketing’s output, the fact that each campaign and piece of content is meant to be different than the one that came before it, made for an environment that at first seemed opposed to the basics of systemization that the rest of a company had come to accept.
To illustrate the idea I plotted a spectrum. The left side represented zero variance, the realm of manufacturing and Six Sigma, and the right was 100 percent variance, where R&D and innovation reign supreme.
While the poles of the spectrum help explain it, it’s what you place in the middle that makes it powerful. For example, we could plot the rest of the departments in a company by the average variance of their output (finance is particularly low since so much of the department’s output is “governed” — quite literally the government sets GAAP accounting standards and mandates specific tax forms). Sales is somewhere in the middle: A pretty good mix of process and methodology plus the “art of the deal”. Marketing, meanwhile, sits off to the right, just behind R&D.
But that’s just the first layer. Like so many parts of an organization (and as described in my essays on both The Parable of Two Watchmakers and Conway’s Law), companies are hierarchical and at any point in the spectrum you can drill in and find a whole new spectrum of activities that range from low variance to high variance. That is, while finance may be “low variance” on average thanks to government standards, forecasting and modeling is most certainly a high variance function: Something that must be imagined in original ways depending on a number of variables include the company, and its products and markets (to name a few). Zooming in on marketing we find a whole new set of processes that can themselves be plotted based on the variance of their output, with governance far to the low variance side and creative development clearly on the other pole. Another way to articulate these differences is that the low variance side represents the routine processes and the right the creative.
While I haven’t seen anyone else plot things quite this way, this idea, that there are fundamentally different kinds of tasks within a company, is not new. Organizational theorists Richard Cyert, Herbert Simon, and Donald Trow, also noted this duality in paper from 1956 called “Observation of a Business Decision“:1
At one extreme we have repetitive, well-defined problems (e.g., quality control or production lot-size problems) involving tangible considerations, to which the economic models that call for finding the best among a set of pre-established alternatives can be applied rather literally. In contrast to these highly programmed and usually rather detailed decisions are problems of a non-repetitive sort, often involving basic long-range questions about the whole strategy of the firm or some part of it, arising initially in a highly unstructured form and requiring a great deal of the kinds of search processes listed above. In this whole continuum, from great specificity and repetition to extreme vagueness and uniqueness, we will call decisions that lie toward the former extreme programmed, and those lying toward the latter end non-programmed. This simple dichotomy is just a shorthand for the range of possibilities we have indicated.
This also introduces an interesting additional way to think about the spectrum: The left side is representative of those ideas where you have the most clarity about the final goal (in manufacturing you know exactly what you want the output to look like when it’s done) and the right the most ambiguity (the goal of R&D is to make something new). For that reason, high variance tasks should also fail far more often than their low variance counterparts: Nine out of ten new product ideas might be a good batting average, but if you are throwing away 90 percent of your manufactured output you’ve massively failed.
It is difficult to deal with the uncertainty of the future, as one must to relate an organization to others in the industry and to events in the economy that may affect it. One must look ahead to determine what forces are at work and to examine the ways in which they will affect the organization. These activities are less structured and more ambiguous than dealing with concrete problems and, therefore, the CEO may have trouble focusing on them. Many experiments show that structured activity drives out unstructured. For example, it is much easier to answer one’s mail than to develop a plan to change the culture of the organization. The implications of change are uncertain and the planning is unstructured. One tends to avoid uncertainty and to concentrate on structured problems for which one can correctly predict the solutions and implications.2
Going a level deeper, another way to cut the left and right sides of the spectrum is based on the most appropriate way to solve the problem. For the routine tasks you want to have a single way of doing things in an attempt to push down the variance of the output while on the high variance side you have much more freedom to try different approaches. In software terms this can be expressed as automation and collaboration respectively.
While this is primarily a framework for thinking about process, there’s a more personal way to think about the variance spectrum as it relates to giving feedback to others. It’s a common occurrence that employees over-or-misinterpret the feedback of more senior members of the team. I experienced this many times myself in my role as CEO. Because words are often taken literally from the leader of a company, an aside about something like color choice in a design comp can be easily misconstrued as an order to change when it wasn’t meant that way. The variance spectrum in that context can be used to make explicit where the feedback falls: Is it a low variance order you expect to be acted on or a high variance comment that is simply your two cents? I found this could help avoid ambiguity and also make it more clear I respected their expertise.
This paper is kind of amazing to read. It feels revolutionary to actually look at how specific decisions come to be made within a company.↑
There’s a whole other really interesting area to explore here that I’m mostly skipping over about using the variance spectrum to help decide types of problems and the mix of work. Although I don’t have a specific model (hence why this is a footnote), the idea that you should decide on your portfolio of activities based on having a good diversity of work across the spectrum is fascinating and seems like a good idea. It’s also in line with a point Herbert Simon makes at the very beginning of his book Administrative Behavior: “Although any practical activity involves both ‘deciding’ and ‘doing,’ it has not commonly been recognized that a theory of administration should be concerned with the processes of decision as well as with the processes of action. This neglect perhaps stems from the notion that decision-making is confined to the formulation of over-all policy. On the contrary, the process of decision does not come to an end when the general purpose of an organization has been determined. The task of ‘deciding’ pervades the entire administrative organization quite as much as does the task of ‘doing’- indeed, it is integrally tied up with the latter. A general theory of administration must include principles of organization that will insure correct decision-making, just as it must include principles that will insure effective action.”↑
Cyert, R. M., Simon, H. A., & Trow, D. B. (1956). Observation of a business decision. The Journal of Business, 29(4), 237-248.
Cyert, R. M. (1994). Positioning the organization. Interfaces, 24(2), 101-104.
Dong, J., March, J. G., & Workiewicz, M. (2017). On organizing: an interview with James G. March. Journal of Organization Design, 6(1), 14.
Thanks again for reading and for all the positive feedback. Please keep it coming. If you haven’t read any of these yet, the gist is that I’m writing a book about mental models and writing these notes up as I go. You can find links at the bottom to the other frameworks I’ve written. If you haven’t already, please subscribe to the email and share these posts with anyone you think might enjoy them. I really appreciate it.
Credit: Organizational Charts by Manu CornetI first ran into Conway’s Law while helping a brand redesign their website. The client, a large consumer electronics company, was insistent that the navigation must offer three options: Shop, Learn, and Support. I valiantly tried to convince them that nobody shopping on the web, or anywhere else, thought about the distinction between shopping and learning, but they remained steadfast in their insistence. What I eventually came to understand is that their stance wasn’t born out of customer need or insight, but rather their own organizational chart, which shockingly included a sales department, a marketing department, and a support department.
“Organizations which design systems (in the broad sense used here) are constrained to produce designs which are copies of the communication structures of these organizations.” That’s the way computer scientist and software engineer Melvin Conway put it in a 1968 paper titled “How Do Committees Invent?” His point was that the choices we make before start designing any system most often fundamentally shapes the final output.1 Or, as he put it, “the very act of organizing a design team means that certain design decisions have already been made.”
Parkinson aside (who did so mostly in jest), very few have the chutzpah to actually name a law after themselves and Conway wasn’t responsible for the law’s coining. That came a few months after the “Committees” article was published from a fan and fellow computer scientist George Mealy. In his paper for the July 1968 National Symposium on Modular Programming (which I seem to be one of the very few people to have actually tracked down), Mealy examined four bits of “conventional wisdom” that surrounded the development of software systems at the time. Number four came directly from Conway: “Systems resemble the organizations that produced them.” The naming comes 3 pages in:
Our third aphorism-“if one programmer can do it in one year, two programmers can do it in two years”-is merely a reflection of the great difficulty of communication in a large organization. The crux of the problem of giganticism [sic] and system fiasco really lies in the fourth dogma. This — “systems resemble the organizations that produced them” — has been noticed by some of us previously, but it appears not to have received public expression prior to the appearance of Dr. Melvin E. Conway’s penetrating article in the April 1968 issue of Datamation. The article was entitled “How Do Committees Invent?”. I propose to call my preceding paraphrase of the gist of Conway’s paper “Conway’s Law”.
While most, including Conway on his own website, credit Fred Brooks’ 1975 Mythical Man Month with naming the law, it seems that Mealy deserves the credit (though Brooks’ book is surely the reason so many know about Conway’s important concept).3Back to the questions at hand: Why does this happen, where does it happen, and what can we do about it?
Let’s start with the why. This seems like it should be easy to answer, but it’s actually not. The answer starts with some basics of hierarchy and modularity that Herbert Simon offered up in his Parable of Two Watchmakers: Mainly, breaking a system down into sets of modular subsystems seems to be the most efficient design approach in both nature and organizations. For that reason we tend to see companies made up of teams which are then made up of more teams and so-on. But that still doesn’t answer the question of why they tend to design systems in their image. To answer that we turn to some of the more recent research around the “mirroring hypothesis,” which (in simplified terms) is an attempt to prove out Conway’s Law. Carliss Baldwin, a professor at Harvard Business School, seems to be spearheading much of this work and has been an author on two of the key papers on the subject. Most recently, “The mirroring hypothesis: theory, evidence, and exceptions” is a treasure trove of information and citations. Her theory as to why mirroring occurs is essentially that it makes life easier for everyone who works at the company:
The mirroring of technical dependencies and organizational ties can be explained as an approach to organizational problem-solving that conserves scarce cognitive resources. People charged with implementing complex projects or processes are inevitably faced with interdependencies that create technical problems and conflicts in real time. They must arrive at solutions that take account of the technical constraints; hence, they must communicate with one another and cooperate to solve their problems. Communication channels, collocation, and employment relations are organizational ties that support communication and cooperation between individuals, and thus, we should expect to see a very close relationship—technically a homomorphism—between a network graph of technical dependencies within a complex system and network graphs of organizational ties showing communication channels, collocation, and employment relations.
It’s all still a bit circular, but the argument that in most cases a mirrored product is both reasonably optimal from a design perspective (since organizations are structured with hierarchy and modularity) and also cuts down the cognitive load by making it easy for everyone to understand (because it works like an org they already understand) seems like a reasonable one.4 The paper then goes on to survey the research to understand what kind of industries mirroring is most likely to occur and the answer seems to be everywhere. They found evidence from across expected places like software and semiconductors, but also automotive, defense, sports, and even banking and construction. For what it’s worth, I’ve also seen it across industries in marketing projects throughout my own career.
That’s the why and the where, which only leaves us with the question of what an organization can do about it. Here there seem to be a few different approaches. The first one is to do nothing. After all, it may well be the best way to design a system for that organization/problem. The second is to find an appropriate balance. If you buy the idea that some part of mirroring/Conway’s Law is simply about making it easier to understand and maintain systems, than its probably good to keep some mirroring. But it doesn’t need to be all or nothing. In the aforementioned paper, Baldwin and her co-authors have a nice little framework for thinking about different approaches to mirroring depending on the kind of business:
As you see at the bottom of the framework you have option three: “Strategic mirror-breaking.” This is also sometimes called an “inverse Conway maneuver” in software engineering circles: An approach where you actually adjust your organizational model in order to change the way your systems are architected.5 Basically you attempt to outline the type of system design you want (most of the time it’s about more modularity) and you back into an org structure that looks like that.
Dominant organisations are prone to stumble when the new technology requires a new organisational structure. An innovation might be radical but, if it fits the structure that already existed, an incumbent firm has a good chance of carrying its lead from the old world to the new.
A case study co-authored by Henderson describes the PC division as “smothered by support from the parent company”. Eventually, the IBM PC business was sold off to a Chinese company, Lenovo. What had flummoxed IBM was not the pace of technological change — it had long coped with that — but the fact that its old organisational structures had ceased to be an advantage. Rather than talk of radical or disruptive innovations, Henderson and Clark used the term “architectural innovation”.
Like I said before, it’s all quite circular. It’s a bit like the famous quote “We shape our tools and thereafter our tools shape us.” Companies organize themselves and in turn design systems that mirror those organizations which in turn further solidify the organizational structure that was first put in place. Conway’s Law is more guiding principle than physical property, but it’s a good model to keep in your head as you’re designing organizations or systems (or trying to disentangle them).
He was writing mostly about software systems, but as you’ll see it’s much more broadly applicable.↑
As an aside, it’s hard not to think that Mealy’s third point about what one programmer can do versus two sounds a lot like Fred Brooks’ “mythical man month” concept. Mealy worked with Brooks on OS/360 and in the book Computer Pioneers by J.A.N. Lee it’s mentioned that Mealy’s Law was also named at the 1968 symposium: “There is an incremental programmer who, when added to a project, consumes more resources than are made available.” Sounds pretty similar to me.↑
There’s a very interesting point about the role of “information hiding” in pushing companies into Conway’s Law. Essentially the idea is that companies naturally hide information within teams or departments for the sake of simplicity across the rest of the company. It would only make things more complicated, for instance, if the finance team exposed the detailed rules of GAAP accounting instead of just distributing a monthly GAAP accounting report. “Information hiding as a means of controlling complexity is a fundamental principle underlying the mirroring hypothesis. With information hiding, each module in a technical system is informationally isolated from other modules within a framework of system design rules. This means that independent individuals, teams, or firms can work separately on different modules, yet the modules will work together as a whole (Baldwin and Clark, 2000).”↑
Henderson, R. M., & Clark, K. B. (1990). Architectural innovation: The reconfiguration of existing product technologies and the failure of established firms. Administrative science quarterly, 9-30.
Hvatum, L. B., & Kelly, A. (2005). What do I think about Conway’s Law now?. In EuroPLoP (pp. 735-750).
Lee, J. A. (1995). International biographical dictionary of computer pioneers. Taylor & Francis.
MacCormack, A., Baldwin, C., & Rusnak, J. (2012). Exploring the duality between product and organizational architectures: A test of the “mirroring” hypothesis. Research Policy, 41(8), 1309-1324.
MacDuffie, J. P. (2013). Modularity‐as‐property, modularization‐as‐process, and ‘modularity’‐as‐frame: Lessons from product architecture initiatives in the global automotive industry. Global Strategy Journal, 3(1), 8-40.
Mealy, George, “How to Design Modular (Software) Systems,” Proc. Nat’l. Symp. Modular Programming, Information & Systems Institute, July 1968.