MANUFACTURING & SERVICE OPERATIONS MANAGEMENT
Vol. 6, No. 2, Spring 2004, pp. 133–148
ISSN 1523-4614 EISSN 1526-5498 04 0602 0133
DOI 10.1287/MSOM.1030.0028
© 2004 INFORMS®
The terms pull and lean production have become cornerstones of modern manufacturing practice. However, although they are widely used, they are less widely understood. In this paper, we argue that while the academic literature has steadily revealed the richness of the pull/lean concepts, the practitioner literature has progressively simplified these terms to the point that serious misunderstandings now exist. In hopes of reducing confusion, we offer general, but precise definitions of pull and lean. Specifically, we argue that pull is essentially a mechanism for limiting WIP, and lean is fundamentally about minimizing the cost of buffering variability.
Key words: pull production; just-in-time; CONWIP; lean manufacturing
History: Received: August 15, 2002; accepted: July 23, 2003. This paper was with the authors 10 months for
2 revisions.
We have been teaching executive courses on manufacturing regularly since the mid-1980s. Starting at the height of the just-in-time craze and continuing through time-based manufacturing, business process reengineering, enterprise resource management, and lean manufacturing movements, we have been doing our best to explain pull production to people who really want to use it. We even wrote a book on the subject (Hopp and Spearman 2001). With all of our experience, and all of the attention pull has received in the practitioner literature, one would think our task would be getting easier.
However, it is not. In recent years, both of us have been finding increasing confusion among our students about how pull is defined, and decreasing willingness to accept our definition. The main controversy has usually stemmed from some participants (who have almost always received some sort of lean training prior to attending one of our classes) insisting pull means making products to order, as opposed to making them to stock or forecast. Interestingly, this is a fairly recent phenomenon. Ten years ago, our students equated pull with kanban, a make-to-stock system, but not with make-to-order. Nevertheless, it has been a disturbing experience for us because we do not feel that pull is properly defined as either kanban or make-to-order.
Have we been wrong all along? The sheer number of confrontations in our classes forced us to consider that possibility. To decide, we revisited fundamental issues we thought were long resolved, such as what constitutes a pull system? What makes pull work? What is lean, and how does it relate to pull? In addition, we carefully reviewed the practitioner and academic literatures related to pull. This paper is the result of our soul-searching process. In it, we:
To understand how and why pull came about, it is necessary to first appreciate the environment that preceeded it, namely the world of MRP. Prior to the dominance of the computer in manufacturing, inventory was controlled using reorder-point/reorder-quantity (ROP/ROQ) type methods. During the 1960s, Joseph Orlicky, Oliver Wight, and George Plossl along with others developed a new system, which they termed Material Requirements Planning (MRP). Orlicky obviously believed that they were on to something big; he subtitled his book on the subject The New Way of Life in Production and Inventory Management (1975). After a slow start, MRP began to gather steam during the 1970s fueled by the “MRP Crusade” of the American Production and Inventory Control Society (APICS). Orlicky (1975) reported 150 implementations in 1971. By 1981, the number had grown to around 8,000 (Wight 1981). As it grew in popularity, MRP also grew in scope, and evolved in the 1980s into Manufacturing Resources Planning (MRP II), which combined MRP with Master Scheduling, Rough-Cut Capacity Planning, Capacity Requirements Planning, Input/Output Control, and other modules. In 1984 alone, 16 companies sold $400 million in MRP II software (Zais 1986). By 1989, over $1.2 billion worth of MRP II software was sold to American industry, constituting just under one-third of the entire software industry (IE 1991).
While MRP was steadily dominating the American production control scene, history was taking a different course in Japan. There, perhaps because it lacked a strong indigenous computer industry, the computer was far less pervasive in production and inventory control. Instead, several Japanese companies, most notably Toyota, developed the older ROP/ROQ methods to a high level. Starting in the 1940s, Taiichi Ohno began evolving a system that would enable Toyota to compete with American automakers, but would not depend on efficiencies resulting from long production runs that Toyota did not have the volumes to support. This approach, now known as the “Toyota Production System,” was designed to “make goods; as much as possible, in a continuous flow” (Ohno 1988).
According to Ohno, the Toyota Production System rests on two “pillars”: (1) “autonomation” and (2) just in-time production (JIT). Autonomation, or “automation with a human touch,” is the practice of determining the optimal way to perform a given task and then making this the “best practice” standard method. Autonomation also involves “fool proofing”or “poke yoke,” which uses devices to quickly check dimensions and other quality attributes to allow workers to be responsible for their own quality. If problems were found, the production line stopped until the problems were corrected. This eliminated the need for rework lines and, eventually, eliminated most scrap. The Toyota Production System also promoted “5S,” Seiri, Seiton, Seiso, Seiketsu, and Shitsuke, which are organizational and housekeeping techniques aimed at achieving Autonomation and Visual Control.
Just-in-time production according to Ohno involved two components: kanban and level production. Kanban or “pull production” became the hallmark of the Toyota Production System (which was also frequently referred to as just-in-time) to the point where many thought they were synonymous. However, kanban was just a means to an end. Ohno famously described his inspiration for kanban while returning from a visit in the United States during the 1950s in which he was more impressed with American supermarkets than with American manufacturing. The idea of having all goods available at all times was, to Ohno, novel and revolutionary. He said:
From the supermarket we got the idea of viewing the earlier process in a production line as a kind of store. The later process (customer) goes to the earlier process (supermarket) to aquire the needed parts (commodities) at the time and in the quantity needed. The earlier process immediately produces the quantity just taken (re-stocking the shelves) (1988, p. 26).
To do this, Ohno had to make some major system changes. Because the supermarket was to replenish only what was just taken in a timely manner, lot sizes had to be drastically reduced. To achieve the efficiencies needed, Ohno and his Toyota colleagues found many creative ways to reduce change-over times. Although change did not happen overnight (the JIT revolution was really an evolution), the results were substantial. In 1945 setups on large presses took two to three hours. By 1962 they had been reduced to 15 minutes, and by 1971 some were down to 3 minutes (Ohno 1988). With such short change-overs, Ohno could achieve “one-piece flow” and JIT production.
By the early 1980s, American manufacturers had become acutely aware that they had fallen behind in manufacturing innovation and even in manufacturing efficiency (especially in the automotive sector). Although MRP sales continued to climb, many were thinking that MRP had been a mistake. A 1980 survey showed that less than 10% of the firms interviewed had recouped their investment within two years (Fox 1980). JIT began being hailed as the next great thing.
American managers quickly became enamored with everything Japanese. American professors went to Japan to learn first-hand what was going on, and of course they wrote books. The first JIT book, published in 1981, was Hall’s Driving the Productivity Machine: Production and Control in Japan. This was followed by Schonberger’s Japanese Manufacturing Techniques: Nine Lessons in Simplicity in 1982. By 1983 Yasuhiro Monden, a Japanese professor, got on the bandwagon with The Toyota Production System. Shigeo Shingo, who worked with Ohno, published a book on setup reduction in 1985, The SMED System. Ohno’s book finally appeared in English in 1988.
Despite the frenzy of interest in JIT throughout American industry during the 1980s, results were mixed. Firms implementing JIT were faced with a deceptively simple philosophy and a complicated array of techniques (see Zipkin 1991 for an elegant discussion of the disconnect between “romantic” JIT (philosophy) and “pragmatic” JIT (techniques)). Managers had to contribute considerable site-specific innovation to produce workable systems. Depending on how creative and insightful they were, JIT sometimes worked and sometimes did not.1
1The early confusion about JIT had many philosophical, cultural, and technical explanations. However, it may also have been partly intentional. In a 1990 interview, Ohno claimed that Toyota considered the system so powerful that they deliberately coined misleading terms and words to describe it. “If in the beginning, the U.S. had understood what Toyota was doing, it would have been no good for us.” Eventually, however, Toyota became very open and invited the whole world to see their factories in the 1980s and 1990s (Meyers 1990).
By the end of the 1980s, JIT began being eclipsed by the next great thing—Enterprise Resources Planning (ERP). With the development of the client/server information technology architecture, it became feasible to integrate virtually all of a corporation’s business applications with a common data base. ERP offered both near-total integration and “best-of-breed” software in the specific applications. Of course, ERP was much more complex than MRP II, containing modules for every business function imaginable, from accounting and financial functions to human resources. And it was correspondingly more expensive, with implementation costs at some companies soaring as high as $250 million (Boudette 1999). In spite of the price tag and the growing number of implementation horror stories, ERP continued to grow in popularity. As the 1990s drew to a close and fear of the Millennium Bug intensified, ERP was being installed at a feverish rate.
As ERP began its rise, it appeared that the JIT movement had run its course. Even though Toyota continued to rank at the top of the automotive industry in quality and efficiency metrics, interest in the Toyota Production System was on the wane. However in 1990, a landmark case study conducted by MIT was published in The Machine That Changed the World by Womack et al. (1990). This study compared American, European, and Japanese automobile manufacturing techniques and concluded in no uncertain terms that the Japanese methods, particularly those of Toyota, were vastly superior. In addition, the authors freshened JIT by recasting it as “Lean Manufacturing.” With a new name and a new set of stories to rekindle interest, the system created by Taiichi Ohno again became a hot topic in the world of manufacturing.
The JIT movement also spawned a separate movement that ultimately became larger than JIT itself—Total Quality Management (TQM). Originally cast as a means for facilitating smooth production flow, TQM grew into a popular management doctrine institutionalized in the ISO 9000 Certification process. The focus on TQM in the 1980s also spurred Motorola to establish an ambitious quality goal and to develop a set of statistical techniques for measuring and achieving it. This approach became known as “Six Sigma” and was eventually adopted by companies such as Allied Signal, and then General Electric. Six Sigma entered the mainstream management lexicon when Jack Welch, the charismatic CEO of GE, declared that it played a major role in his company’s financial success. Today, Six Sigma carries on the legacy of TQM just like lean carries on the legacy of JIT.
Despite traveling separate paths since emerging from JIT, it now appears that the Lean Manufacturing and Six Sigma movements are about to merge, as suggested by the recent best-selling management book Lean Six Sigma (George 2002). At the same time, the computer approach characterized by MRP/ERP is also undergoing consolidation. After Y2K proved to be a nonevent, ERP became passe as a term, but was quickly replaced by SCM (Supply Chain Management). Remarkably, SAP, the world’s largest provider of ERP software, changed its entire product line in a matter of months (or so it would appear when a search of their website revealed that all references to ERP were gone and had been replaced with the new SCM acronym). Because both ERP and SCM trace their roots back to quantitative production and inventory control, it is perhaps not surprising that they would wind up embodied in the same software products. However, all this leads one to wonder when the Lean Six Sigma SCM movement will begin.
While corporate America tried to navigate this sea of buzzwords and to comprehend the Toyota Production System along with the mysterious concept of pull, those in academia were busy as well. As with any new management trend, academic researchers were quick to get on the JIT bandwagon. Hall and Schonberger were both academics who went to Japan to study JIT first-hand. Monden, of the University of Tsukuba, did not have to travel as far. It appears the first academic paper describing kanban was published by Japanese researchers in 1977. The title of the paper by Sugimori et al. (1977) is telling: “Toyota Production System and Kanban System: Materialization of Just-In-Time and Respect-For-Human System.” Like Orlicky’s “New Way of Life,” the Toyota Production System is something big that goes beyond production control. The title of this first paper may also have contributed to the mistaken equivalence between the Toyota Production System and kanban. It was followed four years later by another paper by two other Japanese researchers, Kimura and Terada (1981). Both of these papers described the mechanics of kanban and the requirements for its implementation. As such, they set the stage for the books of Schonberger, Hall, and Monden. Hence, by the early 1980s the mechanics of kanban had been widely described. Because it represented the first system to be termed a “pull” system, and is hence central to the subject of this paper, we briefly describe these mechanics below.
The classic version of kanban as pioneered by Toyota is called a “two-card kanban” system, as shown schematically in Figure 1. Production begins when a material handler with a move card removes a standard container of parts from the outbound stock point. The move card authorizes that handler to take these parts and tells him where the parts are needed. Before the container is removed, the production card is removed from the standard container and placed on the production card board.2 Production cannot begin without a production card, a container of the appropriate incoming parts, and an idle work station. When all three are available, the worker removes a standard container of parts from the inbound stock point, removes the move card from the container, places it in the hopper for move cards at the work station, and begins to process the parts. Periodically, a material handler collects the move card(s), locates the needed parts, transports them to the work station, and the process repeats at the next work station upstream.
It is easy to see that the two-card kanban system is the result of an artificial distinction between parts processing and material movement. If we include material movement as a separate process, we see that the two-card kanban system becomes a one-card system with the move card becoming a “production” card for the move process. This is illustrated in Figure 2, which also shows that a kanban system is essentially a serial production system with blocking.
2Some authors indicate that the term “kanban” refers to the card, while others indicate that it is the board holding the cards.
Figure 1 A Two-Card Kanban System
If kanban is nothing more than a serial production system with blocking, then why all the fuss? Systems with blocking occur naturally (e.g., in automotive systems where limitations on space and the number of pallets can cause downstream congestion to shut down upstream production). Why would Toyota deliberately constrain the flow in their system? Researchers and practitioners have had to address this fundamental question since the beginnings of JIT.
The first, and easiest, step in understanding kanban in specific, and pull in general, is to characterize its benefits. These have been widely cited as (see, e.g., Cheng and Podolsky 1993, Hirano 1988):
Figure 2 Equivalence of the One- and Two-Card Kanban Systems
Another, and more subtle, problem is to identify what it is about kanban that yields these benefits. This is important if we are to understand how much kanban can be simplified or generalized and still be effective. In essence, this is the question that determines what really constitutes pull.
The water and rocks analogy suggests that the benefits of kanban stem from environmental improvements (removing “rocks” or problems). Some research has supported this view. For instance, Huang et al. (1983) performed a simulation of kanban and concluded that without significant changes in the manufacturing environment, kanban would not be successful in the United States. Krajewski et al. (1987) performed an even more comprehensive simulation study and concluded that the benefit of kanban is due more to an improved environment than to any fundamental change in logistics.
However, ascribing kanban’s benefits to environmental improvements does not explain exactly what it is about kanban that leads to improvements. Furthermore, most researchers were not willing to credit environmental improvements as the sole explanation of all of JIT’s benefits. Ohno clearly regarded kanban as essential to JIT, and many companies had used versions of it to great success. Moreover, if the flow control method did not matter, then successful JIT companies could have reverted to MRP after improving their environments. However, this did not happen. Hence, to more fully understand what makes kanban work, researchers turned to mathematical models.
Karmarkar (1986, 1991) appears to have been first to note the similarities of kanban to earlier base-stock systems in two simple and insightful papers. The system he discusses is very similar to a traditional base-stock system as described by Simpson (1958). Simpson attributes this system to George Kimball in an unpublished Arthur D. Little report that Karmarkar had published in 1988. Simpson describes the base-stock system as follows:
When an order is placed, it is filled from inventory if the inventory is not zero. If the inventory is zero, the order is placed in a backorder file, to be filled when an item arrives. In any event, a manufacturing order is immediately placed with the preceding manufacturing operation to produce an item to replace the item that has been consumed. The manufacturing operator, in turn, immediately places an order for the required raw materials against the preceding inventory, and as soon as this order is filled [i.e., he has the needed inventory], proceeds to “operate” on it to produce the required item. In this way an order against the last inventory for a finished item is immediately transmitted all the way back along the line to all the manufacturing operations, each of which is galvanized into production.
However, while kanban and the base-stock systems are similar, they are not identical. The difference was described by Spearman (1992, p. 950) as follows:
Kanban would not place an order for more parts if a demand (in the form of a move card) had arrived when there was no stock in the outbound stock point. Instead, there would be one or more production cards already in process and whenever one of these was completed, the waiting move card would be immediately attached to the recently completed container of parts and the production card would be sent back into production. In this way a kanban system bounds the amount of WIP there can be in the system, while the base-stock system does not.
Spearman et al. (1990) and Spearman and Zazanis (1992) found that while specific environmental improvements are certainly influential (e.g., setup reduction, production smoothing), there are three primary logistical reasons for the improved performance of pull systems:
Based on these findings, Spearman and Zazanis proposed a hybrid push/pull system known as CONWIP that possesses the benefits of kanban But can be applied to more general manufacturing settings. A number of properties of CONWIP systems were demonstrated using simulation in Spearman et al. (1990), which also described several practical implementation considerations. The book Factory Physics also discussed several ways in which CONWIP can be implemented, including using CONWIP in maketo-order environments (Hopp and Spearman 2001). Independent work by Veatch and Wein (1994) also concluded that the WIP constraint is a major source of the benefit of kanban.
During the late 1980s and early 1990s a number of researchers compared different versions of kanban, CONWIP, and other systems. A good overview of the work before 1996 can be found in Huang and Kuisiak (1996).
As more variants of kanban and related systems were proposed, researchers began to seek ways to unify them. For example, Buzacott (1989) described a “generalized kanban policy” and introduced the concept of “production authorization cards.” This concept was developed further in Buzacott (1993) where it was shown to represent MRP, kanban, CONWIP, as well as many hybrid systems. Using this framework and sample path arguments, Buzacott and Shanthikumar showed that CONWIP is the optimal structure for maximizing throughput for any WIP limiting system. Tayur (1993) also considered the question of the optimal pull structure, viewing systems as a continuum from kanban to CONWIP, depending on how they are partitioned into WIP controlled segments and considering how to optimize the allocation of cards in such systems so as to maximize throughput.
In another unification paper, Axsäter and Rosling (1993) showed that kanban is a special type of “installation stock” policy (i.e., one in which the reorder point is an integer multiple of the reorder quantity and where backlogs are not subtracted from the inventory position). They also defined an “echelon stock policy” in which the inventory position is the sum of the installation inventory positions for the installation and all downstream installations. They then demonstrated that echelon stock policies will always dominate installation stock policies and; therefore, kanban will be dominated by a echelon stock policy. In later work, Axsäter and Rosling (1994, 1999) showed that a kanban system can be reproduced within MRP with a suitable selection of parameters (i.e., a zero lead time and the safety stock less one is a multiple of the order quantity). The result is completely general except that MRP is a periodic review system, while the kanban system is a continuous review system (although there is no reason why the review in MRP could not be done continuously). This matters because if kanban is a subset of MRP, then it is clearly dominated by MRP. However, this dominance will be limited to the experimental conditions because the Axsäter and Rosling model does not consider robustness issues, such as those arising from accidentally releasing too much work into the line.
Other generalizations of kanban include those of Glasserman and Yao (1994) and Yao and Cheng (1993). These papers introduce a “block-and-hold-k” card that determines how much work in process should be kept before moving. However, because justification of this approach relies on the existence of separate move cards, the benefits disappear when we consider the equivalent one-card kanban system.
Suri (1998) introduced a generalization of kanban called “Paired-Cell Overlapping Loops of Cards with Authorization” or POLCA. POLCA is different from kanban in that cards are assigned to pairs of cells, rather than particular parts within a cell. The result is a more general construct than kanban that can be applied to make-to-order situations. Because the initial authorization is from a MRP system but cards are used to limit WIP within pairs of cells, POLCA possesses both push and pull characteristics. One drawback, however, is that one must set a card count for every set of consecutive cells. Nonetheless, Suri describes several examples where POLCA has been used to greatly reduce overall cycle times and WIP.
More recently, Liberopoulos and Dallery (2002) (building on Dallery 2000 and Chaouiya et al. 2000) developed another generalized pull system in the form of a make-to-stock system with separate limits on finished goods (S) and WIP (K) that we will refer to as a (K, S) system. This model incorporates classical base stock (K =∞, S < ∞), CONWIP (K = S), single-stage, reserved-stock kanban as in Buzacott (1989) (K<S), single-stage backordered kanban policy as in Buzacott (1989) and Dallery and Liberopoulos (2002) (K>S), and several other formulations. They define a “critical WIP,” Kc, as a minimum WIP level that supplies enough throughput from the production system so that for any K ≥ Kc, the optimal base-stock level is equal to the base-stock level that would occur with K =∞, noted as S∞. They then conjecture that the policy parameters that minimize total WIP and finished goods (with identical carrying costs) are Kc and S. If this conjecture is true, then one can easily improve on a CONWIP make-to-stock policy by providing separate limits on WIP and finished goods. A trivial case is that of a single-server, production-inventory system, where the optimal values are clearly Kc = 1 and S set to whatever value is required to maintain the fill rate.
The research literature gave a clear definition of kanban and a detailed summary of its benefits. It also showed that kanban could be generalized in a wide variety of practical ways. However, all these variants of kanban suggested that pull was certainly a more general idea than kanban. So, what exactly is it?
Industry did not have a clear answer. One reason for this was that Ohno and the other early practitioners of JIT discussed pull only in very general, highlevel terms.
Manufacturers and workplaces can no longer base production on desktop planning alone and then distribute, or push, them onto the market. It has become a matter of course for customers, or users, each with a different value system, to stand in the front line of the marketplace and, so to speak, pull the goods they need, in the amount and at the time they need them (Ohno 1988, p. xiv).
In other words, one should not simply make a large amount of stock and then try to go and sell it. One needs to be aware of the market and pay attention to the customer.
While this might sound like common sense today, it was revolutionary for the mid-20th century. However, it only described the strategy of pull, not the tactics of pull. Although Ohno did outline the elements needed to make pull work: (1) standard work methods (autonomation) and (2) level production, his writings fell short of providing a working description of pull.
Because of the lack of a more precise definition of pull at the shop floor level, pull quickly became equated with its first manifestation, kanban. Symmetrically, push became nearly synonymous with MRP (see, e.g., Hall 1983). Most magazine articles and trade press books provided vague and self-referential definitions, such as “pull is the opposite of push.”
A rare exception was the 1993 book Just-in-Time Manufacturing: An Introduction that states, “ …the pull mode of manufacturing only allows parts to be moved from the previous operation to the next when the subsequent operation is ready to process” (Cheng and Podolsky 1993, p. 42). They identified one of the advantages of pull as follows:
Pull systems by far outreach the responsiveness of a push system. The responsiveness of the system to changes and problems that arise in upstream processes allows the downstream processes to be shut down. This prevents the accumulation of inventory on the plant floor (Cheng and Podolsky 1993, p. 43).
In short, a WIP cap prevents a WIP explosion. Sadly, this book was not typical.
In the mid-1990s “pull” shifted in popular usage from being synonymous with kanban to shorthand for make-to-order. A key catalyst for this change was the 1996 book by Womack and Jones Lean Thinking, which was a follow-up to their highly successful The Machine That Changed the World. Unfortunately, while this book was widely read and provided many details and case histories related to lean techniques, including pull, it did not provide clean definitions of basic concepts. (In a sense, this book was the opposite of Ohno’s, which was short on details, but clear on basic philosophy.) For instance, in the chapter titled “Pull” Womack and Jones begin,
Pull in simplest terms means that no one upstream should produce a good or service until the customer downstream asks for it, but actually following this rule in practice is a bit more complicated.
While this may look like a reformulation of Ohno’s description of pull, it is not. While Ohno was speaking at the strategic level about the basic connection between production and demand, Womack and Jones were talking about the tactics of implementing lean. At the tactical level, waiting for customers to ask for goods or services does indeed cause things to get “a bit more complicated.” For instance, consider Ohno’s supermarket example. The grocer would have to wait for the customer to ask for his groceries before stocking them!
Lean Thinking was just one of many books and articles that confused the concept of pull with the simpler idea of make-to-order. However because of its popularity, it has served to muddle the understanding of pull significantly. It has become common for managers to state that pull is make-to-order, while push is make-to-stock. For example, when Alcoa announced its heavily publicized Alcoa Production System, one of the basic tenants was: “Produce for use, not for inventory” (2001), even though its principle method of production control was kanban.
Unfortunately, such “lean thinking” has become pervasive in industry. Boeing, now a champion of lean and a provider of lean training to its suppliers, gives the following definition for pull on their website:
Pull production is the opposite of push. It means products are made only when the customer has requested or “pulled” it, and not before (The Boeing Company 2002).
Using this definition, a make-to-order MRP system would be an ideal pull system, which clearly contradicts the historical intent of pull.
Despite the confusion over tactical pull, companies implementing lean did appreciate Ohno’s insight that leveling production was key to “strategic” pull and that the way to accomplish this was to use a “takt time” or “takt-paced production.” For example, Boeing (2002) defined “takt-paced production” as:
Takt-paced production describes the rate of assembly in a factory. Lean does not mean doing things faster; it means doing things at the right pace. Essentially, the rate of demand establishes the pace, or takt time. So, rather than simply maximizing the rate of work, lean sets the pace in the factory, ensuring that the customer’s needs are met on time.
Taken at face value, it would appear that demand must be extremely regular because otherwise following customer demands would be hugely inefficient. In reality, however, setting a pace (instead of chasing demand) is exactly what Boeing (and Toyota) do to smooth the demand that is seen by the plant. This means that they set the takt time based on a current backlog of orders and then adjust it from time to time. Because most releases can be connected to a customer order, these systems are, in an overall sense, make-to-order.
In an analytical sense, production smoothing is really a simple matter of buffering the production line from demand variability. This buffering is done with either a time backlog or inventory. If demand temporarily increases, orders are backlogged. If orders are needed later, the line will build up some inventory.
Furthermore, while the takt time drives final assembly, component parts must be available for them to be “pulled” from internal fabrication centers or outside suppliers. Clearly, any component with a lead time longer than the time between the start of the final assembly and when the unit is needed must be made-to-stock.
If actual demand varies enough (e.g., due to seasonality or random surges), the order backlog may occasionally run dry. When this occurs, the firm must adjust the takt time and/or release some jobs without explicit customer orders. Hence, it is possible for a takt-time-based system to prerelease jobs in maketo-forecast mode.
The bottom line is that although the practitioner literature has commonly defined pull to be make-toorder, this only applies at the strategic level. At the tactical level, the systems actually used to implement pull can be make-to-stock, or even make-to-forecast. Given this, it is no wonder that many practitioners have found the literature on lean confusing and difficult to implement.
To remove this confusion, we need a definition of pull that captures what is essential for obtaining its benefits, but does not overly restrict implementation (e.g., by equating it with a specific system such as kanban).
As we noted above, the term “pull” can be applied at both the strategic and the tactical levels. Establishing a takt time to set the output of the plant to be equal to demand is a way to establish strategic (or market) pull. This was an important part of Ohno’s original vision (around 1950) and was adopted (or, more likely, independently discovered) by those who developed computer-based push systems. For instance, Wight (1970) described very similar logic (i.e., “Input/Output Control”) to be used with MRP. Hence, strategic pull can be implemented even in what have been traditionally called push systems.
Therefore, the key question we must answer to fully define pull is what constitutes tactical pull as characterized by kanban? To be meaningful, we must define pull so that it includes those systems that people tend to think of as pull (e.g., kanban) but not those that have been considered as push (e.g., MRP). Moreover, if pull is somehow intrinsically better, our definition should provide a means to determine if a system belongs to a class of systems that dominate in performance. Finally, although admittedly less important, such a definition should settle long standing debates among academics as to whether various systems (e.g., base-stock models and other make-to-stock variations) are fundamentally pull or push systems.
Given common usage, we begin with the assumption that kanban is a pull system, while MRP is a push system. To define (tactical) pull, we need to characterize what is fundamentally different about these systems.
Axsäter and Rosling (1994) show that MRP is more general than an installation stock (Q, r) policy so that any (Q, r) policy can be replaced by a MRP system. This implies that MRP should also dominate a kanban policy because it is simply an installation stock policy with an additional limit on the number of outstanding orders (Axsäter and Rosling 1993). However, the additional constraint appears to be the key to the effectiveness of the pull system. The fact that kanban does not continue adding orders to the system beyond a certain point puts a natural limit on WIP. Thus, no matter how wrong the forecast or how great the demand, the system cannot be overwhelmed beyond its capacity. Because the output of a production line is an increasing but bounded function of WIP (bounded by the bottleneck rate), while flow time begins to grow almost linearly beyond a certain point, it is pointless to add more work to a system that is already “saturated” (Spearman 1991). Findings by Veatch and Wein (1994) also point to the beneficial effects of having a WIP limit and we discuss these at length in Factory Physics (Chapter 10).
Based on this observation we propose the following definition:
Definition (Pull and Push). A pull production system is one that explicitly limits the amount of work in process that can be in the system. By default, this implies that a push production system is one that has no explicit limit on the amount of work in process that can be in the system.
If we are interested only in the essence of push or pull, it is quite easy to construct a mathematical model that is either purely push or purely pull. For instance, a closed queueing network, with a rigidly set limit on the number of entities, is a pure pull system, while an open queueing system (e.g., a GI/G/1 queue) is a pure push system. The former has a clear limit on WIP, while the latter does not.
However, in the real world there are no pure push or pure pull systems. For example, while a kanban system establishes a clear limit on WIP via the production cards, there are almost always circumstances under which this limit will be overridden (e.g., a downstream nonbottleneck machine goes down). Conversely, while MRP does not establish a limit on WIP, there is almost certainly some level of WIP that will cause management to ignore the planned order releases, or revise the master production schedule to prevent further WIP growth. Indeed, there is presumably some limit on WIP for every system (e.g., an amount equivalent to 1,000 years of demand). The distinction, however, is that the WIP limit in practical pull systems is explicitly stated and is generally small. Any WIP limit in a practical push system is implicit, large, and usually comes into play too late (i.e., after WIP is out of control).
Hence, our definition gives a black and white distinction of push and pull among mathematical models. However the real world, as is generally the case, is a matter of shades of gray. The extent to which a system will obtain the benefits of pull depends on how sharply the up-front WIP limit is imposed.
Note that our definition of pull does not involve the concepts of make-to-order (MTO), make-to-stock (MTS), or make-to-forecast (MTF). Indeed, as we illustrate in Table 1, the push/pull distinction is orthogonal to the MTO/MTS/MTF decision. These examples show that all combinations are possible. Hence, while MTO may indeed be superior to MTS or MTF in many environments, this distinction does not capture the essence of why tactical pull systems improve production efficiency. In the end, it is the explicit limit on WIP that defines a tactical pull system and makes it work.
| Make-to-forecast | Make-to-order | Make-to-stock | |
| Push | MRP with forecas | MRP with firm orders | Q r with pull from FGI |
| Pull | Kanban with takt time & forecast | Kanban with takt time & orders | Q r with pull from FGI Kanban with pull from FGI |
Table 1 Examples of Push and Pull
The good news inherent in our definition of pull is that it implies that pull can be implemented in a variety of ways. Kanban is certainly one way to limit WIP. But there are others. To illustrate this we consider some of the most common systems found in the industry and the literature, and classify them as either push or pull.
The confusion surrounding pull has led to confusion around other issues. These include the “push/pull interface” and the essence of what defines lean production.
If we accept the above definition of push and pull systems, the “push/pull interface” described by Lee and Billington (1995) and by ourselves (2001) is actually a misnomer. What we and Lee and Billington were pointing out was the fact that virtually all production systems include make-to-stock and make-toorder segments. For instance, consider the stylized McDonald’s system shown in Figure 3. During rush hour, when there are specified targets for the warming table, production up to the warming table is make-tostock; production beyond this point (i.e., bagging and checkout) is make-to-order.
The warming table generally establishes a WIP cap on the front end of the line, but there is no firm limit on WIP in the back end of the line. Therefore, we defined these segments as pull and push and called the interface (warming table) the push/pull interface. However, because theoretically one could envision systems like this with no WIP limit on the front end and/or a finite limit on WIP in the back end, this is not a proper definition. A more accurate term would be the inventory/order (I/O) interface because it represents the point in the production process where the stimulus for work movement shifts from make-tostock to make-to-order.
This correction in terminology does not change the fundamental point that we and Lee and Billington were making, which is that these interfaces exist in all production systems, and that positioning them is a management decision. Locating the inventory/order interface at raw materials produces a conventional make-to-order system, while locating it at finished goods produces a classic make-to-stock system. Locating it at an intermediate point (e.g., by stocking components or subassemblies) leads to an assembleto-order system, which can be used if the firm wants to be able to quote customer lead times that are shorter than the overall manufacturing cycle time.
Figure 3 The Inventory/Order Interface
Just as kanban was part of the larger construct of JIT, pull is part of the larger construct of lean production. And just as there is considerable confusion about pull, there is confusion about lean. Most sources describe the essence of lean production as waste reduction. However because this does not sound very deep, lean books and articles are fond of speaking in terms of elimination of muda. But whether one speaks in English or Japanese, the idea of reducing waste in search of efficiency is hardly new. Andrew Carnegie, Fredrick Taylor, and Henry Ford were aggressive champions of waste elimination 100 years ago. Therefore, either lean production is simply a new name for an old concept, or it means something more than waste reduction.
Our opinion is that while lean is certainly concerned with driving out waste, it represents a more fundamental framework for enhancing efficiency. Using the language of factory physics, we can define lean as:
DEFINITION (LEAN). Production of goods or services is lean if it is accomplished with minimal buffering costs.
The first source of excess buffering is obvious waste. Such waste includes operations that are not needed, excessive setup times, unreliable machines that can be made more reliable, rework that can be eliminated, etc. Indeed, to the dismay of some lean practitioners, Shingo once said, “ ‘Eliminate waste’ is a nonsensical slogan.” We take it as a given that the elimination of “obvious waste” is obvious. Unfortunately, this tends to be the focus and limit of most lean implementations that never get to the real “meat” of the problem.
The less obvious source of buffering costs is variability. This can take on many forms, including variability in process times, delivery times, yield rates, staffing levels, demand rates, etc. Anything in the system that is not absolutely regular and predictable exhibits variability. The causes of variability can be classified into internal factors (such as setups, downtime (scheduled and unscheduled), operator-induced fluctuations in production rates, yield loss, rework, engineering change orders, and many others) and external factors (including irregular demand, product variety to meet market needs, customer change orders, etc). External variability is often the consequence of a firm’s business strategy, such as offering high levels of product variety to achieve a competitive advantage. In such instances, a certain amount of variability is simply a cost of doing business. However, external variability is not always justified by strategic considerations. For example, offering discounts, pushing to make end-of-quarter sales targets, interrupting operations to satisfy a specific customer, and so on, may not generate sufficient revenue to justify their cost.
Regardless of its source, all variability in a production system will be buffered. A fundamental principle of factory physics is that there are three types of variability buffer: inventory, capacity, and time. For example, safety stocks represent inventory buffers against variability in demand and/or production. Excess capacity can also provide protection (i.e., a capacity buffer) against fluctuations in demand and/or production. Finally, safety lead times provide a time buffer against production variability. While the exact mix of buffers is a management decision, the decision of whether or not to buffer variability is not. If variability exists, it will be buffered somehow.
This framework suggests a few simple steps for lean implementation:
There are other paths to becoming lean besides the one Toyota took. For instance, one of us has been working with Moog, Inc., a producer of precision servo valves. Their approach was quite different from Toyota’s, but consistent with the basic principles outlined above. Moog’s basic problem was that lead times were too long for the changing market, costs were up, and customer service was down. There was a real danger of losing customers. To address these issues, we developed and implemented the following strategy:
The result has been much greater responsiveness to the customer with improved service. The improved flow also resulted in an unexpected (for management) benefit—a greater than 5% improvement in productivity.
In this section we have defined lean in terms of the cost of buffering. This is critical because the fundamental objective is not to reduce inventory, increase utilization, shorten lead time, or even the “perfect value stream” (Womack 2003). The objective is to make money. Because the cost of the various options for reducing or buffering variability will vary between environments, no one solution is right for all systems. The real challenge of lean is to find the mix of policies that is best for each particular environment.
Pull systems have been a part of the manufacturing lexicon for a quarter of a century. During this time, academics have incrementally described the mechanics of pull systems, characterized their bene- fits, extended kanban into an array of variants, and offered unifying frameworks for integrating and comparing the range of possible pull systems. This work has steadily enhanced our understanding of how and why pull works. Industry practitioners, in spite of their many successes at implementing pull, seem to have focused on describing pull in ever simpler terms. Starting with a tendency to confuse the general concept of pull with the specific practice of kanban, the popular literature has evolved into a simplistic view of pull as nothing more than making products to customer orders. We think this is a serious mistake that compromises the ability of firms to construct effective pull systems for their environments.
The magic of pull is the maintenance of a WIP cap. While pull systems can take on many forms to suit different sets of circumstances, all of them have in common the fact that releases are regulated according to internal system status in a manner that prevents inventory from growing beyond a specified limit. Because this is fundamental to achieving the benefits of pull, it is essential that pull be defined in terms of the WIP cap. Decisions of whether to maketo-order or make-to-stock and how to rely on forecasting are certainly important but are orthogonal to the push versus pull decision. The WIP cap definition of pull given in this paper will prevent confusion of these important, but separate, issues.
The sin of oversimplification has been carried over from pull to lean. Practitioner literature that describes lean production only in terms of reducing waste does the profession a disservice by overlooking important but indirect sources of waste. Moreover, there are many causes of variability that result in a buffer that are not “waste.” For instance, adding variety to a product mix to accommodate customer’s demands is hardly considered waste, but it will reduce efficiency and increase buffers. Lean is better defined as “best buffer” production than “low waste” or even “low buffer” production. Thinking in terms of the more fundamental concepts of variability and buffering encourages more comprehensive consideration of effi- ciency improvement alternatives.
While there have been plenty of cases where academics have confused or missed the point of a practical problem, (see, e.g., Dudek et al. 1992 for a criticism of scheduling research) we feel the pull arena is one where the academic community has been a voice of reason. The question for us now is: Is anyone listening?
The authors would like to thank Lee Schwarz for encouraging them to write this paper. They would also like to thank the truly wonderful referees Dr. Robert Inman, Dr. William Jordan, Professor Uday Karmarkar, and Professor John Buzacott, who provided much guidance in the preparation of the manuscript. All of their input was extremely valuable in sharpening the authors’ thinking and clearing up misunderstandings. Finally, the authors would like to thank Professor Scott Moses who pointed out an important logical inconsistency in the exposition on pull in their book Factory Physics.
