A theory of technology and organizations

Introduction

This article develops a general theory of technology tailored for scholars interested in strategy and organizations. It builds on my earlier work (Baldwin, 2008, 2024; Baldwin and Clark, 2000; Colfer and Baldwin, 2016) and draws inspiration from thinkers like Thomas Kuhn, Giovanni Dosi, and Richard Nelson who explored scientific and technological paradigms. ¹

Put simply, a technology is a mix of practical knowledge and skill aimed at achieving a human goal (Arthur, 2009). Technologies are everywhere—from agriculture, cooking, and the arts, to engineering, business, politics, and sports. Each is based on a different mix of technical knowledge and hands-on expertise. This diversity makes it hard to generalize about “technology” as a single concept. ²

One useful way to simplify this complexity is through the idea of a paradigm. Originally meaning a model or example, Kuhn expanded the term to describe the shared assumptions and methods that guide scientific communities. Dosi, Nelson, and others later showed that technological paradigms exist across disciplines and industries, each with its own models and methods. So while the idea of paradigms helps, it doesn’t fully solve the problem of technological heterogeneity.

For strategy and organization scholars, this heterogeneity is inconvenient. Technologies are constantly evolving. What works today may be outdated tomorrow. As technologies shift, organizations and strategies must adapt—or risk becoming obsolete.

Some technological changes can be dramatic. As Joseph Schumpeter noted, no matter how many mail coaches you add, you’ll never create a railway. In cases like this, the new technology may require an entirely different organizational structure and capabilities (Teece, 2008, 2014). When railways replaced mail coaches, many coach companies couldn’t make the leap. Later, planes and automobiles displaced railways. Shifts like this make technological change a central concern for both strategists and organizational designers.

To better understand these transitions, we need a theory of technology that applies across different eras and industries. This article proposes such a theory and uses it to examine a range of technologies and their organizational consequences.

Building blocks

To build a theory of technology, we first need to identify its basic components. In organizational economics, scholars often focus on transactions, contracts, or property rights. For instance:

In all these views, technology plays a background role. It’s usually treated as a simple function—something that converts inputs and effort into outputs, defined by vague terms like “scalability” or “asset specificity.”

But if we want to understand how real technologies influence the structure of organizations, we need to dig deeper. We must look at how technologies are actually built and used—in the same way a skilled practitioner would.

As indicated, when viewed from inside an organization, the basic units of a technology are:

A working technology is simply a network of tasks and transfers that, when performed correctly, produces something useful. To visualize this, we use a Task Structure Matrix (TSM). In a TSM, tasks appear on the diagonal, and transfers show up as marks off the diagonal. Arrows show how outputs from one task feed into others. Transfers going forward (below the diagonal) reflect movement toward the endpoint; backward transfers (above the diagonal) reflect cycles, feedback loops, or rework. ¹⁴

Transfers involve a secondary set of tasks: the movement of material, energy or information from the point of origin to the point of use. Most technical recipes specify inputs but leave methods of transfer to the implementors of the recipe.

A transfer from Task A to Task B perforce creates a dependency of B on A. Dependencies are transitive: if C depends on B and B depends on A, then C depends on A. A backward transfer from B to A, whether direct or indirect, creates a cyclic interdependency.

Figure 1(a) shows a simplified TSM for designing a laptop computer. Figure 1(b) adds a view of the product’s main components—screen, motherboard, drive, and casing. ¹⁵ Notice that transfers are denser within components than between them, but cross-component dependencies still exist. That means the components aren’t fully modular—we may call them proto-modules

OPEN IN VIEWER

Figure 1.

Task structure matrix for design of a laptop computer.

TSMs add a crucial new layer to traditional organizational economics. Figure 2 contrasts two views: the mainstream view where a firm is simply a bundle of contracts, and a modified view where we can see the underlying task and transfer network. Technologies determine this inner structure. They specify what needs to happen inside a firm for it to turn raw material, labor and capital into finished products.

OPEN IN VIEWER

Figure 2.

Two views of a firm.

Herbert Simon was one of the few economists to take this internal view seriously. In a memorable metaphor, he imagined an alien observer looking down at Earth through a special telescope. Firms would appear as green blobs filled with internal structure, connected by red lines representing market transactions. The observer would see a landscape of large green zones, not just a web of red lines. In short, organizations—not markets—would dominate the picture. ¹⁶

In this article, we will follow Simon’s lead and look inside the green zones. We will treat technologies not as black boxes, but as structured systems of tasks and transfers. This lets us see how changes in technology reshape the structure of work—and how organizations must adapt in response.

How technologies change

I have said that “technologies change.” It would be more correct to say that “technologies can be changed” by modifying tasks and transfers. For example, suppose the designers of the laptop computer depicted in Figure 1 decided that for the sake of efficiency and flexibility, they should change the interdependent system into a modular system. How would they proceed? ¹⁷

If the nature of a set of transfers between two components is known, designers can eliminate the transfers by agreeing to ex ante rules that both components must obey. The transfers would then disappear, replaced by a rule, which would constrain both components. (For example, the rule “all screws the same size” can eliminate the costly “fitting” of screws into holes of different diameters. This in turn is an example of the principle of interchangeable parts: “Every part must be produced to fit at once into the design for which it is made. In mass production, there are no fitters.” ¹⁸ )

The end result of a systematic effort to eliminate cross-component transfers is a modular task structure as shown in Figure 3. Here the leftmost column is new. It represents all the “design rules” necessary to make the components compatible without ongoing transfers between them. Design rules must be formulated ex ante before work on the components begins.

OPEN IN VIEWER

Figure 3.

Task structure matrix for a modular laptop computer.

The last six rows, also new, indicate a final stage called “system integration and testing.” At this point, the modularized components are brought together, assembled into whole computers, and then tested to ensure the functionality of the entire system. Necessarily, integration and testing take place ex post, when the modules are ready to be assembled.

In Figure 3, horizontal arrows indicate connections between individual design rules and various tasks within the separate components. (The total number of transfers from the design rules to the individual modules is much larger than the number of arrows actually shown. Most arrows have been omitted to simplify the picture.)

Notice that, in the modular system, there are no transfers between the individual modules. The design rules ensure that the modules will be compatible once assembled, even without ongoing coordination between their respective designers.

Symmetrically, a set of vertical arrows shows transfers of completed modules from the last task in each module’s TSM to the first task in the final integration and testing phase. There are only as many of these arrows as there are modules: in this fashion the steps required to design and produce a module are “hidden” from the other modules and from the integration and testing process. However, the tasks involved in integration and testing of the assembled system must be consistent with how the system actually works. Thus, the design rules constrain the integration and testing stage: again, most of those arrows are not shown to reduce clutter in the figure.

A subtle change in timing is embedded in the transition from the integral system in Figure 1 to the modular system in Figure 3. Given an integral system, the firm cannot scale up to full production until most of the design work is complete. Furthermore, once production begins, it is very difficult to change any part of the design. Design and production must occur in a strict sequence.

In contrast, the absence of transfers between modules in Figure 3 means that design and production tasks do not have to be strictly segregated in time: modules can be redesigned and new designs substituted into an existing system. Thus, the modular system depicted in Figure 3 contains many more options than the integral system depicted in Figure 1. We will come back to this difference in the discussion of “platforms” below.

Transactions, thin crossing points, and the boundaries of firms ¹⁹

A transaction is a voluntary transfer of goods and/or services from one party to another in exchange for something of corresponding value. By definition, it is a reciprocal exchange: people are not forced into transactions, and each side is made better off by the trade.

Transactions are different from the transfers depicted by off-diagonal x’s in a TSM. For example, suppose Bob needs a knife to slice a loaf of bread, and his partner, Ann, is near the knife drawer. In their own kitchen, Ann may simply hand the knife to Bob. In this case, (1) Ann and Bob do not have to count the number of knives transferred; (2) they do not have to define what the knife is; and (3) Bob does not have to pay Ann for the knife. The transfer of the knife can occur even if none of these conditions is satisfied. However, the transfer cannot be a transaction unless all three conditions are satisfied. ²⁰

If there is no appropriate knife in the kitchen, Bob and Ann must go to a store and buy one. At the store, the couple together with the store’s employees must first (1) define the type of knife that is needed—in this case, a bread knife. Both parties must also (2) agree on how many knives will be purchased—one. Finally, the couple must (3) compensate the store for the knife, using an agreed-upon means of payment. Only then, can the couple take the knife back to their kitchen and slice the loaf of bread.

The three steps—defining, counting, compensating—add further tasks to the overall task network as well as transfers of material, information and money between the couple and the store. Thus, a transaction is more than a plain transfer: it is a transfer with several added, costly features. I have called these added costs mundane transaction costs to differentiate them from the “opportunistic” transaction costs studied by Williamson and the property rights theorists.

As Adam Smith noted, modern economies are characterized by a fine division of work across different specialized establishments connected by transactions. In general, many more transfers take place within establishments than transactions between establishments. This is precisely because of mundane transaction costs: if every task or transfer in an establishment had to be recorded and paid for on the spot, then very little work would get done. Thus, it is generally efficient to segregate interdependent tasks and transfers within specialized organizations (firms or households) and place transactions only at the organizations’ boundaries. In this way, complex activities related to production and/or consumption can take place without the frictional costs of making every transfer a transaction.

Firms and households generally have few actual transactions in their interiors: ²¹ in this sense, they are “transaction-free zones” (TFZs). Inside these zones, rather than enter into formal exchanges, participants can cooperate freely to get a job done. (The contributors may be bound by longer term contracts that make cooperation in their best interest in the short term.)

Figure 4 shows the relationship of technologies and transactions in a market-based economy. The figure depicts TSMs for two organizations, a store and a household. The store sources and sells kitchen supplies: the household carries out the tasks of daily life, including cooking. The two organizations are non-overlapping: both operate internally as transaction-free zones.

OPEN IN VIEWER

Figure 4.

Two organizations linked by a transaction.

The store and the household are connected by the transfer of a knife from the former to the latter by means of a transaction. As indicated earlier, the transaction is more complicated than a simple transfer from Ann to Bob. To purchase the knife, Ann and Bob must pass through a checkout line: the knife’s barcode is scanned, the register records its price, and adds it to the bill. The couple then chooses a means of payment, taps a card on the scanner and presses a few buttons. The household now owns the knife: its cost will be deducted from their bank account or added to their credit card balance.

Centuries of technological development, starting with marks on clay tablets, continuing through double-entry book-keeping, to today’s online banking and payment systems, lie behind modern transactions, which may take less than a minute to complete. In fact, transactions are a complex form of joint production involving goods, information, legal ownership, and money. The two arrows placed at the thin crossing point between the two transaction-free zones indicate the presence of a highly scripted and automated exchange.

In summary, mundane transaction costs—defining, counting, compensating—are lowest at the “thin crossing points” of the task network, where two transaction-free zones come together. Thin crossing points in turn can be created via modularization—the process of resolving and removing dependencies via design rules described above.

TSMs and thin crossing points are fundamental building blocks for a theory of technology and organizations. Their purpose is to describe technologies in an abstract and general way. Any technology can be represented by tracing tasks and transfers and constructing a corresponding TSM. More complex technical systems in turn can be modeled by combining the TSMs of component technologies via thin or thick crossing points. The thin crossing points in turn are favorable locations for transactions, defined as formal, recorded transfers of ownership from one party to another within a designated economic domain.

The spectrum of complementarity ²⁸

Technical systems are distinguished by the fact that their elements and components are complementary, meaning that the value of the parts together is greater than the sum of the parts individually. In the early 1990s, Paul Milgrom and John Roberts constructed a formal theory of complementarity based on the concept of “supermodularity” proposed by the mathematician, Donald Topkis. ²⁹ In its simplest form, supermodularity is a property of some functions with multiple arguments V ( x , y ) such that an increase in one argument increases the positive effect of an increase in the other argument.

Formally, two or more inputs to a system are complementary if more of one makes more of the other more valuable. Thus, let V ( x , y ) be a function that is (by definition) increasing in both arguments. Without loss of generality, let V ( 0 , 0 ) = 0 . V ( x , y ) is supermodular and x and y are complements if and only if:

In other words, increasing both arguments together increases the value of the function by more than increasing one or the other separately. Colloquially, y makes x more valuable. This is known as the property of “increasing differences.” ³⁰ It is the essence of complementarity.

The function V can have any number of arguments, and can be continuous or discrete. This opens up the possibility that there are degrees of complementarity—a spectrum—ranging from super strong to strong to weak to non-existent. The concept of a “complementary surplus” provides a way to locate a particular system along the spectrum. Setting V ( 0 , 0 ) = 0 , we can define the complementary surplus, CS as:

A high complementary surplus indicates a high degree of complementarity: the presence of both (or all) inputs creates a system with much more value than the sum of the values of the inputs separately. Correspondingly a low complementary surplus indicates a lesser degree of complementarity: assembling the inputs into a single system does not increase the value of the collection very much.

Centripetal and centrifugal forces ³³

The spectrum of complementarity itself is the product of economic “forces” that “push” or “pull” firms in opposite directions. The forces reflect different degrees of complementarity in the underlying technologies.

Centripetal forces are the result of strong complementarities in a technical system. They pull organizations together and reward high levels of integration in centrally managed enterprises. Centripetal forces are strongest when the technical recipe calls for task interdependence and tight scheduling of transfers such as arise in high-volume mass production.

Centripetal forces are also caused by non-contractible, unobservable effort. In these cases, the technology generally requires many uncodified and contingent transfers of material, energy and information. The people performing the tasks—for example, cooks in a restaurant kitchen—will develop organizational ties that are flexible and can be responsive to new information. In such organizations, there are no thin crossing points where transaction costs are low. Thus, ideally, the tasks, assets and people should all be placed within a single transaction-free zone—the kitchen—and report to a single boss—the chef.

Centrifugal forces are the result of weak complementarities in the system. These forces push organizations apart by rewarding decentralized, wide-ranging exploration of multiple alternatives. They arise in the presence of diverse preferences and dispersed knowledge, when the best alternative is not known a priori. They are reinforced when the underlying technical system is modular, making it easy to swap components in and out. So-called network effects, which arise when adding more agents to a network increases its value to everyone, are also a source of centrifugal forces.

Figure 5 summarizes the impact of complementarities on the choice of an organizational design. At one extreme when complementarities among subsets of tasks are “super strong,” the separate tasks have value only when the other tasks are also performed. If ownership of task outputs is distributed across different actors, then early actors (whose costs are sunk) are at risk of holdup by later actors. It is then desirable to create organizational ties that encourage cooperation (common employment, common dispute resolution processes). High levels of synchronization also require ongoing communication, which can be achieved by via dedicated communication channels and/or co-location. Thus “super-strong” complementarities call for highly integrated organizations capable of performing the required tasks reliably and without interruption.

OPEN IN VIEWER

Figure 5.

The impact of complementarity on organization design.

At the other extreme, with very weak complementarity, there is little reason to spend resources on coordinating actions: centrifugal forces will then push the organizations toward higher levels of autonomy and independence.

In the middle ground, however, there is a balance of centripetal and centrifugal forces. In these cases, very tight coupling of actions and investments is unnecessary and often self-defeating because it reduces flexibility. Autonomous firms are better able to explore a wide range of technical configurations and recipes. In this region of the spectrum, loosely coordinated ecosystems of firms and individuals will create more value than either a single integrated firm or many independent, unconnected firms.

Technological paradigms

TSMs, thin crossing points, the mirroring hypothesis, and the spectrum of complementarity provide us with basic conceptual tools that can be used to address the problem of heterogeneity in technologies identified at the beginning of this article. First, any technological process can be represented in the form of a task structure matrix (TSM). TSMs depict the connections between technological tasks. They reveal the direct and indirect dependencies between tasks including forward sequences and backward cycles.

Second, to succeed in implementing a specific technical recipe, there must be agents (people or machines) positioned to carry out each task and transfer in the corresponding TSM. Thus, organizations carrying out a given technological process must “mirror” the TSM corresponding to the process. Furthermore, transactions costs are lowest at “thin crossing points” in the TSM where transfers are few and simple. However, task structures can be changed to make crossing points thinner or thicker.

Finally, the “complementary surplus” measures the value of a set of inputs used in combination (as directed by a technical recipe) minus the value of the components deployed separately. The size of the surplus relative to the next best alternative is proportional to the economic incentives to implement the technology in question vs some other technology. “Centripetal forces” caused by task interdependency, synchronized steps, and non-contractible effort cause strong complementarities, which increase the surplus. These forces make it advantageous to manage the resources within a single, centrally controlled organization. In contrast, “centrifugal forces” caused by diverse preferences, dispersed knowledge, modularity, and network effects reward more loosely coupled, decentralized organizations including business ecosystems and markets.

The next three sections describe three representative technological “paradigms” which call for different forms of organization: (1) job shops; (2) synchronized flow processes; and (3) platform systems. Each technological paradigm has a distinct, generic TSM. Using the paradigms, real technologies can be organized into groups according to their structural similarities. ³⁴ I begin with job shops, then discuss flow production processes, then platform systems.

Job shops ³⁵

A job shop is an organization that sells goods produced in small quantities (sometimes as low as one). Each job completed by the shop may be different from the one before, reflecting what customers want to purchase at a given time. The order of steps may be similar for each job or it may vary. Work often cycles backwards to earlier stages in the process.

In Adam Smith’s time (1750-1800 approximately), job shops were the most common form of organization, and they remain important today. The butcher, the brewer, and the baker, who provided Smith with his dinner, managed job shops. ³⁶ The shops did not each sell a single standard product: the butcher sold different cuts of meat, the brewer made different types of ale, and the baker offered different kinds of cakes, pies, and bread.

Workers in a job shop assume different roles at different times. Everyone contributes as needed, and transfers material, energy, and information to everyone else. Examples of job shops today include most restaurants and small retail shops, doctors’ and dentists’ offices, classrooms (but not lecture halls), emergency rooms, local police and fire departments, as well as R&D labs, and entrepreneurial startups.

The generic task structure matrix for a job shop is shown in Figure 6. There is likely to be a central coordinator (indicated by C in the TSM), but transfers do not follow a linear flow. Instead, as the arrows show, transfers can go in many directions, and cycles are common:

OPEN IN VIEWER

Figure 6.

Task structure matrix for a representative job shop.

The advantage of job shop organizations is that they are versatile and responsive to new information. They can also offer high levels of variety, within some fixed scope. As such, they permit experimentation and make product innovation relatively easy.

In Chapter 3 of The Wealth of Nations, Smith explains why job shops were common during his time. He observed that “the division of labor is limited by the extent of the market.” In essence, the best organization to supply products in a given location depended on the cost of placing the product in the hands of potential buyers. When the “accessible market” was small, organizations had no choice but to produce a variety of goods.

A country carpenter . . . is not only a carpenter, but a joiner, a cabinet-maker, . . . as well as a wheel-wright, a plough-wright, a cart and waggon maker. ³⁷

In these settings, making one product was not sufficient to make a living. Given small demand and high transportation costs, a skilled worker supported by a family-sized organization had to be versatile—a jack of all trades.

However, managing high variety and fluctuating demand has significant drawbacks for organizational efficiency. First, the time spent on each step will vary across jobs. There will always be extra capacity somewhere in the system. Idle people or machines give the shop owner incentives to take on multiple jobs at the same time: this in turn creates the need to change paths and/or employee assignments to reflect temporary imbalances.

Coordinators of job shops must thus be prepared to redeploy resources in response to transient surpluses and shortfalls in different departments. Unfortunately, addressing a surplus or shortfall in one instance will generally not lead to permanent improvement: the imbalances are likely to be different the next time round.

Job shop production was common in the United States until the middle of the 1800s:

At one time a metalworking factory would be willing to make pumps, steam engines, farm implements, tools, locomotives, in brief, just about anything in metal that their craftsmen could handle. ³⁸

However, when a market is growing rapidly, shops capable of making a broad variety of related goods may find it profitable to concentrate on a few products or even just one. This happened in the U.S. in the aftermath of the Civil War:

By the end of the Civil War a number of specialized manufacturers emerged who made just pumps or locomotives or machine tools. During the next two decades this specialization of product increased to where firms were making just one type of a general product, such as one type of pump, or just lathes instead of general machine tools . . .. ³⁹

Specialization in turn makes it worthwhile to create a standardized workflow applicable to all jobs in the shop. When individual tasks are repetitive, dividing the work into smaller jobs can reduce setup/transition times and thus increase throughput. Taken to its limit, the result is a synchronized flow production process. This technological paradigm is discussed next.

Synchronized flow production ⁴⁰

Although job shops were the most common form of organization during Adam Smith’s lifetime, he began The Wealth of Nation, not with the shops on the streets of Edinburgh, but with the obscure process of making pins. He admitted that this was “a very trifling manufacture,” but looking at the pin factory allowed Smith to study the impact of arranging tasks as a series of successive steps performed by different people (Peaucelle and Guthrie, 2011).

He called this technological arrangement “the division of labor.” We may call it flow production. According to Smith, “The division of labor, insofar as it can be introduced, occasions, in every art, a proportionable increase of the productive powers of labor.” ⁴¹ In other words, synchronized flow production is always and everywhere the key to increased productivity and wealth.

Smith’s evidence for this claim was weak at best. First, he made an error in his calculations: the productivity gain due to the division of labor in the pin factory was not 240 times as Smith claimed but closer to 2.4 times. ⁴² Furthermore, he provided no reasons other than “the extent if the market” why or when flow production should be preferred to versatile, job shop production.

Why were pins amenable to flow production, when so many other products were not? Pins are small objects, with a standardized design. They weigh very little and are very durable. In the 1700s, seamstresses, tailors, and lawyers ⁴³ used pins by the hundreds. Thus, relative to other products of this period, the extent of the market for pins was very large. Creating a group of dedicated pin factories, located close together, using an efficient flow production process, could be a profitable community effort.

In fact, the pins Smith described were probably made in the vicinity of the village of Laigle in Normandy. This district is close to Versailles and Paris (major centers of dress-making and coat-making), and to seaports like La Havre and Honfleur with easy access to London and the interior. Thus, the pins produced at Laigle could easily be transported to large numbers of users in France, England and beyond. Merchants passing through France from Italy and Spain to England would know where they could buy pins in volume. At the same time, families in Laigle could count on a steady stream of merchants passing through to purchase pins. ⁴⁴

Despite his miscalculation, Smith’s instinct about the future importance of synchronized flow production was prescient. In the late 1700s, population densities in cities and towns were growing as the enclosure movement forced workers to move from the country to cities and towns. ⁴⁵ At the same time, transportation costs were falling because of the building of roads and canals. ⁴⁶ These trends continued for the next two centuries.

At the same time, machines (the spinning jenny, the cotton gin, automatic looms) were invented that automated many simple repetitive tasks formerly done by hand. The trend toward automation also continued for the next two centuries and beyond.

Thus, during the 19th and 20th centuries, flow production in its many forms turned out to be the best way to produce physical goods at low cost, and transport these goods everywhere in the world.

Synchronized flow production technologies specify a sequence of steps, each of which is essential to the final product. At the end of the sequence, a complete product or service is delivered to the customer. Because each step is essential, the steps are strong complements. In addition, for maximum efficiency, each step should take the same amount of time, that is, they should be synchronized. The time needed to complete one step is called the “takt time” (takt = beat or pulse in German). The step that takes the longest is called the “production bottleneck.”

Generic task structure matrices for two representative flow production processes are shown in Figure 7. They differ in the way organizational ties are used to coordinate the flow.

OPEN IN VIEWER

Figure 7.

Task structure matrices for two representative flow production processes.

In the first panel (A), a central coordinator (which may be a computer) communicates with all operators. The last operator lets the coordinator know when the sequence of steps is complete. This communication may also take place through a computer. For example, in the automated checkout procedure described earlier, when a transaction is complete a green light may come on, and a recorded voice will indicate that the register is open “Welcome, . . . please scan your first item . . ..” The next person in line can then begin a new checkout process.

In the second panel (B), after using up the materials needed to complete a step, the downstream operator signals to the upstream operator that they are ready to begin the next task. The signal may be as simple as an empty box. In a task structure matrix, these organizational ties can be indicated by placing an s for “signal” in the cells above the main diagonal. In theory, the backward-signaling arrangement can be self-coordinating as long as the initial input stock does not run out. It was pioneered by Toyota Motors and is now known as the “kanban” system. It has the advantage of limiting work-in-process inventory and production rates to match demand. It can be used both within a factory and across factories in a supply chain. ⁴⁷

As a technological paradigm, flow production has several important advantages relative to job shops. First, flow processes are easy to scale and to combine. To increase the scale of a process, one simply increases the capacity of each step. This can be done by subdividing steps, setting up multiple parallel flow lines, and/or increasing the speed and size of machines at each step.

Flow processes are also easy to combine end-to-end in a chain-like fashion. These recombinations are feasible because the beginning and end of a flow process are generally self-evident and throughput is easy to measure. This is not true of the more complicated and circular processes with varying sequences of steps such as are common in job shops.

Finally, flow processes are easy to improve. In the late 19th and early 20th centuries, the principles of flow improvement were codified under the title “systematic management.” ⁴⁸ The pioneers in this endeavor were a group of managers and consultants, mostly based in the U.S. who gained experience in railroads, steelworks (U.S. Steel), chemical factories (DuPont), appliance companies (Singer), meat packers (Swift), catalog companies (Sears), and automobile companies (Ford and GM). The same principles applied to a wide range of technologies, thus there was cross-fertilization among the leading manufacturers through the late nineteenth and early twentieth centuries. ⁴⁹

For example, William S. Knudsen, President of General Motors’ Chevrolet division, recalled practices in unsystematically managed factories:

Any one of us who worked in shops in the old days remembers how piles of work to be done, . . . rendered confusion around the machine inevitable. ⁵⁰

In contrast, at GM by the late 1920s:

[Every] time we build a car, we handle approximately one ton of material. Now, to handle successfully 100,000 cars per month, this immense burden must flow smoothly and without confusion to its destination. ⁵¹

By the late 1920s, the principles of systematic management were understood, but only by a small number of people. The leaders in improving practice were the railroads and the large companies listed above, especially the automakers Ford and GM. ⁵² Most U.S. factories were still managed inefficiently, however. During the Great Depression of the 1930s, many of these inefficient producers went out of business.

Then, at the very beginning of World War II, President Franklin Roosevelt appointed William Knudsen Director of Production for the US War Department. Knudsen was responsible for the production of war materiel, including aircraft, naval and cargo ships, submarines, tanks, guns and ammunition—essentially everything mechanical needed for the war effort. Very quickly, most of GM’s factories were converted to military production. At the same time, scores of GM’s divisional managers were sent to other companies to assist them with similar conversions. GM’s managers taught workers and managers throughout the eastern United States the principles of systematic management and efficient flow production. (Henry Kaiser played a similar role on the West Coast.) ⁵³

I have said that flow production entails performing a sequence of essential steps in a well-defined order. “More efficient” means achieving higher output with less input, where inputs include material, energy, labor, machinery and time. The linear structure of the process means that the principles for making flow processes more efficient are the same everywhere: ⁵⁴

In an optimized flow production system, the flow will be synchronized (each defined task will take the same amount of time), all workers will be capable of performing their assigned tasks during one work cycle, and work-in-process inventory will be proportional to the number of steps in the end-to-end process.

Efficiency in a flow production process is most easily achieved by placing the sequence of steps within a vertically integrated organization subject to unified governance, direct authority, and a managerial hierarchy. Vertical integration means that all steps in the process lie within the control of a single organization. Control describes both ownership of the material assets used to complete the technical process and authority over the actions of its employees. Governance of both assets and employees is unified.

Vertical integration and unified governance stand in contrast to the “inside contracting” method of organizing work, which was used by many manufacturing organizations during the middle of the 19^th Century. “Inside contracting” is the practice of outsourcing the work of different departments in a factory to contractors who are not employees. This form of governance often leads to disputes between the company managers and the contractors in charge of different parts of the flow process.

David Hounshell provides a vivid account of the weaknesses of inside contracting at Singer Sewing Company in the 1880s. As Singer attempted to scale up to supply a global market, its managers tried to change to a system of interchangeable parts and automated assembly. However, they could not gain the cooperation of their inside contractors in making the switch. Poor quality, high levels of rework, and hand assembly continued to plague their plants until they took direct control of all departments. ⁵⁵

This case demonstrates that a major source of inefficiency in flow processes is the need to resolve disputes between contending parties. Vertical integration and unified governance together remove that cause of inefficiency from the highest ranks of the organization.

Unified governance avoids another source of inefficiency: disagreements between owners and managers. Even if the shareholders who own the enterprise are numerous and have conflicting views, their conflicts must be settled off line (e.g. by a vote of the Board of Directors) and will not create costly real-time conflicts within the flow process itself. ⁵⁶

However, as a practical matter, a flow process with many steps cannot be completely managed by a single person. To solve the problem of delegating decision rights, the best form of organization is often a nested hierarchy. In these organizations: (1) authority relationships are ordered; (2) each superior has links with more than one subordinate; and (3) each subordinate is linked to one and only one superior. These conditions give rise to the inverted tree structure depicted in a typical organization chart or a military chain of command. ⁵⁷

In a flow process managed hierarchically, work generally flows laterally as directed by the sequence of steps, while data flows up the hierarchy and decisions (orders) flow down the hierarchy. This structure is parsimonious with respect to linkages: at each level, a decision maker is “in charge” of a contiguous section of the process and has a limited number of “direct reports.” Superiors generally exercise direct authority over subordinates and disputes between equals can be settled by referring the decision up the chain. There is a risk that decisions in different parts of the hierarchy may be inconsistent, but the inconsistencies can themselves be resolved by formulating general policies and communicating them to the whole organization. This practice gives rise to a classic line-staff organization structure of the type described by Alfred Chandler. ⁵⁸

However, the optimality of nested hierarchical organizations in the management of flow production comes with an important caveat: workers in the organization must accept its design and follow the orders given to them.

The work itself is planned down to the last detail. Individual tasks are defined to be unambiguous, short, easy to perform and part of a larger process. In classic flow production, there is no room in the design for individual skill or problem-solving initiative. ⁵⁹

For example, Henry Ford described a “typical operation”—the processing of a “spring leaf” through the furnace to a bending machine through grinding and painting to the loading dock. The process involved 11 men each performing a different task taking less than a minute. ⁶⁰ The factory had to produce 25,000 leaves per day using “a great battery of lines” operating in parallel. Furthermore, each step “must have its own supply of material delivered in sufficient quantities at indicated places . . . steel at 1; heat at 2; power and oil at 3; molten nitrate at 5; bolts at 7; nuts at 8; clips at 9; paint at 12.” All this for what in the end was a very minor part of an automobile. “It goes to meet other parts of the motor car which have come from other parts of the plant by similar processes.” ⁶¹

The design of this minor part of the system is impressive in its completeness and attention to detail. But the experience of workers on the line was repetitive, physically arduous, demanding of attention, hot, boring, and wholly detached from the final product and its users. In the words of Peter Drucker:

For the great majority of automobile workers, . . . work appears as something unnatural, a disagreeable meaningless and stultifying condition of getting to a pay check, devoid of dignity as well as of importance. No wonder that this puts a premium on slovenly work, on slow downs and on other tricks . . . No wonder that this results in an unhappy and discontented worker—because a paycheck is not enough to base one’s self respect on. ⁶²

This is not the place in which to discuss alternative ways of designing flow production processes. ⁶³ What is true is that flow production under the principles of systematic and scientific management taught by Frederick Taylor, Henry Ford, William Knudsen and others gives rise to work that is back-breaking and mind-numbing. This in turn in turn can lead to denial by workers of the authority relations needed to coordinate both large and small flow production processes.

Despite these drawbacks, the savings achievable by using high-volume flow production methods were such that the years 1750–1980 witnessed the conversion of many existing technologies from job shops to flow production. In addition, hundreds, even thousands of new products were invented during this period based on new sources of power (electricity, oil and gas, internal combustion), new chemical processes (plastics, nylon), new methods of transportation (railroads, automobiles, trucks), and new distribution channels (department, catalog, and chain stores). In these years, “advanced production technology” was practically synonymous with “flow production.” Once a product came out of the lab, the evident path to profit was to use synchronized flow production to supply a mass market.

Table 1 contains a list of technologies that opted to use synchronized flow production between 1750 and 1980, in rough order of their appearance. The business organizations that emerged during this period operated at a scale and scope never seen before. ⁶⁴ They invented and consistently improved new methods of structuring work, increasing efficiency, and speeding up production. Although flow production continued to be dominant until the 1980s, the high point of this era was during and right after World War II.

OPEN IN VIEWER

Table 1.

Technologies using synchronized flow production 1750–1980.

Cannon	Gunpowder	Telegraph
Thread (spinning)	Oil and Gas	Electricity
Ginning Cotton	Chemicals	Wireless Radio
Power Loom (Weaving)	Construction	Automobiles
Wooden Ships	Printing	Airplanes
Rifles and Revolvers	Insurance	Television
Iron Ships	Food Processing	Home building
Railways	Glass	Container shipping
Concrete	Heavy Equipment	Banking and credit
Steel	Farm Equipment	Financial Trading
Cameras and Film	Logistics	Digital computers
Clocks	Adding machines	Semiconductor chips

Platform organizations, and ecosystems

Platform systems represent a fundamentally different technological and organizational paradigm from synchronized flow production and flexible but limited job shops. Platforms are modular systems designed to provide options.

In contrast to job shops and flow production processes, which evolved over centuries, modular platform systems were invented. In 1961, a group of IBM engineers retreated to a motel in Greenwich, Connecticut with the mandate to design a comprehensive plan for IBM’s next generation of computer hardware. IBM’s main problem at this time was the lack of compatibility between its different computer lines. For customers, buying a new computer entailed rewriting all the software running on its existing machines: this deterred IBM customers from buying new equipment as it became available.

The so-called SPREAD group recommended a fully compatible product line where “each processor [would] be capable of operating correctly all valid machine language programs of all other processors with the same or smaller memory configurations.” ⁶⁵ This almost-impossible goal was achieved through modularization via design rules and system integration (see Figures 1 and 3 above). The result was IBM System/360, a fully compatible line of computers. ⁶⁶ Almost immediately after its introduction, the new product line catapulted IBM to a dominant position in the nascent computer industry. ⁶⁷

Then, in 1981, the IBM PC became the first open digital platform system. The PC was based on the same modular design principles as System/360 but relied to a much greater extent on outside suppliers and complementors to provide its users with a wide range of hardware and software options. The result was a so-called open platform with a very large potential market. ⁶⁸ Platforms, especially open platforms with ecosystems, are the third representative technological paradigm.

A platform system is a modular technical architecture in which a core set of components—the platform—supports a variety of complementary options. Each option depends on the presence of the platform, but no particular option is essential. The platform and its options are economic complements: each increases the value of the other. However, because no option is essential (if it were, it would not be an option), the platform and options are weak complements.

A generic task structure matrix for a modular platform system is shown in Figure 8. Figure 3 shows a more realistic task structure matrix for the design of a modular laptop computer system.

OPEN IN VIEWER

Figure 8.

Task structure matrix for a representative modular platform system.

As shown in the figure, a basic platform system has three parts: (1) “design rules” that are communicated to all platform participants; (2) modules containing “hidden information”; and (3) system integration tasks which bring the modules together to make a useful whole system. ⁶⁹ Notice that there are no x’s in the white spaces below and to the right of the two modules in the diagram. This reflects the fact that module providers do not need to transfer or receive material, energy, or information from other module providers. If such transfers are required, the blocks indicated are not modules.

Rather than interacting with each other, module providers receive information in the form of design rules from the architect of the system. Their products must conform to the design rules to work with (be compatible with) other modules and the system as whole.

There can be only one set of design rules—that is, one architect—for the core of a platform system. However, modern digital platforms generally allow additional design rules to be added to the core platform, thus creating a layered modular architecture. ⁷⁰

In the lower part of the matrix, each module transfers material, energy, and/or information to the system integrator. However, these transfers are much simpler (contain less information) than the transfers that take place within each module. In this sense, modules hide information. Information hiding reduces the cost of constructing low-cost yet effective transfers between the module provider and the downstream integrator. Each party can attend to its own tasks without observing exactly what others have done or are doing.

As long as they understand and obey the relevant design rules, there can be many system integrators. Indeed, end users often serve as integrators for their own systems of use. ⁷¹

Open platforms and ecosystems

While the technology of flow production emphasizes tightly-coupled, sequential tasks, the technology of platforms is built around a stable core of design rules plus a set of loosely coupled modules. As we saw in Section 2, design rules replace lateral dependencies between modules, reducing the need for real-time coordination and enabling independent innovation by multiple actors as long as they obey the rules. This structure in turn leads to a different organizational mandate for platforms systems than for flow production: instead of controlling bottlenecks and synchronizing flows, platform leaders must “offer a lot of valuable options” and stimulate innovation. External innovation, outside the boundaries of the platform sponsor, can be encouraged by opening the platform to outside contributors.

Platforms, by definition, create numerous “thin crossing points” in the task network, which have low transaction costs. Unaffiliated contributors can attach themselves to the platform (making use of the platform’s functionality) at those points. Thus, platform technologies support the formation of platform ecosystems—groups of independent firms engaged in complementary activities linked by the platform. ⁷²

There are four basic types of platform systems, each serving a different function in the economy at large. They are shown in Figure 9. Standards-based platforms enable the creation of families of related products, by providing interoperability across components. Logistical platforms coordinate the movement of goods, services, and information. Transaction platforms facilitate transactions between different parties, taking a fraction of the complementary surplus created by the exchange. Communication platforms enable communications, both broadcast and point to point. They are often funded by third-party advertisers or the sale of data “scraped” from the platform.

OPEN IN VIEWER

Figure 9.

Four types of open platforms.

Each type of open platform has existed since ancient times. However, starting in the 1980s, the spread of digital technology brought digital platforms, including personal computers, mobile phones, the Internet, and the Worldwide Web to the forefront of the economy. Automated computations and algorithms reduced the cost of creating and disseminating all kinds of information, including standards, locations, transactions, and messages. This in turn made it easier to create and coordinate the platforms and their ecosystems.

A single firm may simultaneously operate more than one type of open platform. For example, Amazon originally combined a retail transaction platform with a logistical fulfillment platform. ⁷³ Today, the company operates two other transaction platforms: Amazon Marketplace and Amazon Web Services. For its part, Apple controls two standards-based platforms (Apple iOS and mac OS), with corresponding logistics platforms, plus digital and physical transaction platforms (Apple Stores online and physical) and several communication platforms (Apple Mail, Apple News). Google controls a massive communication platform (Search). Through its APIs, Google also controls numerous standards-based platforms. Its logistical platform for distributing content through its data centers is in a class by itself. And it is participating in the development of new chip designs (standards) optimized for data centers.

Digital platforms can also be combined with other technological paradigms at all levels of the economic system. For example, digital technologies allow production lines and other flow processes to attain levels of speed, flexibility, and variety not possible before. ⁷⁴ Digital platform technologies also influence job shops: many complementors and customers of open platforms are quintessential job shops. The platforms give these businesses the ability to advertise and provide customer service (delivery, repair, advice) to consumers everywhere in the world at a very modest cost.

In summary, platform systems have a different structure, hence constitute a different technological paradigm from synchronized flow production and job shop production. Technologically, platforms are modular systems organized around a core set of design rules and optional complements. The TSMs of platform systems reveal a shift from all-to-all transfers in job shops and tightly synchronized linear transfers in flow production processes to loosely coupled modules coordinated by design rules, standards, and APIs. Organizationally, because of their modularity, platforms can be open, admitting contributions by multiple independent organizations coordinated by design rules rather than managerial authority.

Within the modules of a platform system, we may see flow production processes, job shops, and platforms built on platforms, as well as hybrid forms such as divisionalized enterprises, temporary clusters, and various supply and distribution networks. By strategically hiding information via design rules, platforms enable heterogeneity in the organizations that supply the modules of the platform system. The end result is often a complex overlapping network of organizations—an extended multi-platform ecosystem—whose value as a whole greatly exceeds the sum of its separate parts

Is generative AI a new technological paradigm?

Is generative AI, the newest revolutionary technology, a separate paradigm in its own right, or is it a combination of the paradigms discussed above?

In one sense, this is an easy question to answer. A functioning AI computer is the product of interactions within an array of chips with a pre-defined architecture, a curated training data set, validation and testing protocols, and a process of “iterative refinement” that continues until the machine passes all “essential” tests. ⁷⁵ This is a well-defined four-stage process. The actual balance of modularity vs flow production in the creation of the AI can be observed by mapping the TSMs of the AI designers as they implement this four-stage process. It probably has a proto-modular structure similar to the laptop computer design process shown in Figure 1. Within the proto-modules, the teams in each stage function as job shops, but consultations across teams are both common and encouraged.

What we cannot observe about AIs are the steps a trained machine will use to arrive at its answer to a given prompt. There have been some pioneering studies, most recently Anthropic’s “Tracing the Thoughts of a Large Language Model,” ⁷⁶ and related work. However, the “tasks and transfers” inside an AI model at this point do not allow construction of a corresponding TSM. Indeed, “tasks and transfers” may not be the right metaphor for the physical processes instantiated within the circuits of the machine.

In summary, Gen AI is a revolutionary new technology, which greatly expands our concepts of what machines can do. It is a new technological paradigm in the Kuhn-Dosi-Nelson sense—where each new scientific field or industry can beget a new scientific or technological paradigm. ⁷⁷ However, we do not (yet) need to define a new organizational architecture to describe the organizations engaged that design and deploy GenAI systems. They form a four-stage proto-modular platform ecosystem dominated by a few lead firms. In this sense the GenAI ecosystem is similar to the classic supply networks studied by Ron Adner and Rahul Kapoor (2010) and Adner (2017) as well as the clusters of startups in the portfolios of many high-profile venture capitalists.

Comparing the three paradigms

The mirroring hypothesis predicts that the structure of technical systems as revealed by their task structure matrices will be reflected in the organization of the firms that implement the systems. From the perspective of mirroring, platform systems, synchronized flow processes, and job shops each give rise to distinct patterns of organizational design, governance, and coordination.

Job shops are capable of small-batch, custom work with high variability in both tasks and task sequences. Each job may follow a different path through the system, while transfers often loop back to earlier steps within the shop. Organizations that implement job shop technologies tend to be informal, craft-based, and person-dependent. Coordination relies on local communication and mutual adjustment rather than formal hierarchies or design rules. Job shops are versatile and responsive but difficult to scale and hard to improve systematically. Their mandate is to increase the quality and range of their products by investing in workers’ skills. They evolve slowly.

In synchronized flow production, strong complementarity caused by the interdependence of steps and the need for tight timing generally calls for vertical integration and some form of unified governance. (The integration may be temporary as in the filming of a movie or the production of a concert.) Every step in the flow must proceed in a fixed sequence. Organizational control is hierarchical, with a mandate to identify and eliminate bottlenecks. These systems are optimized for speed and efficiency. They can evolve rapidly by adding steps at the beginning or end of the process. But they are vulnerable to disruptions and generally not flexible in the face of changes in technology or demand.

Last, open platform systems combine modularity with distributed governance, decentralized innovation, and thin crossing points enabled by design rules. The platform sponsor’s mandate is to offer users a large and diverse set of options. Often the best way to do this is to give outside complementors access to the platform, thus fostering a platform ecosystem. Coordination of the ecosystem occurs through modular interfaces and standards (rules), rather than direct authority. The systems can evolve rapidly by adding and subtracting new modules and/or layers. They are generally robust to localized failures and adaptable to changing technologies and markets. However different parts of a platform system may not survive a wave of external innovation.

Table 2 contrasts the three technological paradigms in terms of their structure, methods of coordination, governance, sources of innovation, efficiency, and evolvability.

OPEN IN VIEWER

Table 2.

Comparison of three structural paradigms.

Paradigm	Job shop	Synchronized flow	Platform system
Structure	Loosely ordered, cyclic tasks	Sequential, interdependent tasks	Modular, optional components
Coordination	Local adjustment	Centralized control	Design rules & System integration
Governance	Informal	Hierarchy	Distributed
Innovation Source	Craft-based, local	Internal (R&D)	External complementors
Efficiency	Low	High	Variable
Evolvability	Low	Rapid through extension of lines and scaling	Rapid through addition/subtraction of new modules and layers

Conclusion

This article presents a theory that explains how technologies and organizations interact at the level of basic structure to shape the economic landscape. The basic units of analysis are tasks and transfers specified by technological recipes and the agents and organizational ties needed to carry out the recipes.

Quite possibly, all the competing recipes aimed at a specific set of users will use the same structural paradigm. This has been true of automobiles, steel, oil and many other goods since the early 20th century. It generally leads to some form of oligopolistic competition based on scale.

In other cases, different structural paradigms aim to serve different users and thus occupy distinct market positions. This is true in the case of apparel. Today, most of the apparel market consists of ready-to-wear clothing produced using flow production methods. A very small fraction of the market consists of bespoke apparel made in small shops (ateliers) using craft production methods.

However, the latter group includes many prominent fashion houses, which aim to be the source of new ideas and trends. In addition, the most celebrated fashion leaders often have complementary mass-produced product lines, although they are careful to segregate their mass-produced products from the hand-crafted ones. As a result, the corporate entity becomes a platform, offering multiple options. The apparel-making companies in turn are part of an extended ecosystem of non-exclusive suppliers as well as wholesale and retail distributors.

In summary, my goal in this article has been to suggest a language and a set of tools (tasks, transfers, transactions, TSMs) that cut across all technologies. This theory can be used to refine debates about technological strategies and related organizational designs. The increase in precision made possible by this theory can then lead to a more comprehensive picture of how we as humans interact positively or negatively with the larger technological systems we inhabit.