Getting Control of Just-in-Time (2024)

Like all good revolutions, just-in-time manufacturing is producing revolutionaries who don’t know when to stop. It is also producing over reactions from people determined to make them stop. Consider the curiously vexed debate about how to get materials to, and work in process through, the shop floor.

Pick up virtually any manufacturing magazine these days and there will be some articles and pages of advertisem*nts by consultants extolling the virtues of JIT over such computer-driven control systems as materials requirements planning (MRP) or materials resource planning (MRP II)—as if JIT principles were opposed to MRP and to the use of computers. One recent ad put the choice this starkly: “JIT vs. MRP II—JIT is the key to your survival!”

MRP, proponents of JIT explain, is merely a “push” technique. An MRP II program promises manufacturing managers more precision than it can deliver, requires unnecessary information, and demands more formal discipline than the shop floor needs. In contrast, JIT people seem especially drawn to such computerless, “pull” techniques as kanban, the system used extensively in Japan’s auto and electronics industries. For JIT, presumably, human pull is good, computer push is bad.

What must be particularly confusing to manufacturing managers who get wind of this debate is that kanban systems are used most successfully by the same Japanese and American corporations that are famous for spearheading the use of advanced computer automation—Toyota or Hewlett-Packard, for example. In crucial respects, MRP II aims to be a JIT system, while kanban cannot. Worst of all, doesn’t kanban look suspiciously like the old order point, order quantity (OP, OQ) system that MRP once discredited and replaced?

This debate needs clarity, and then it needs to end. The idealized conception of the shop floor one gets from some extreme JIT advocates—a line, inherently flexible, inventoryless, even computerless, replenished by infinitely responsive suppliers—may actually prevent manufacturing managers from using the tools they need to run their operations. JIT principles should certainly not preempt the use of MRP II. Indeed, most advanced manufacturing companies find that they require a hybrid system of shop floor control systems—tailored systems, including innovative pull systems like kanban, as well as time-tested, computer-driven push systems like MRP II.

At the same time, shop floor managers should know just when MRP II is an unnecessary burden and when kanban can’t work—when push comes to shove and pull comes to tug. All managers can learn interesting strategic lessons from their choices. The question of how to manage inventory cuts quickly to the basics of manufacturing in an age of intense global competition: How much automation is enough? How should the factory respond to customers? How much can you load on workers? How do you deal with orders? What is waste? The shop floor is still a microcosm of the whole business.

Pulling and Pushing at JIT

The basic difference between pull and push is that a pull system initiates production as a reaction to present demand, while push initiates production in anticipation of future demand. Thus a fast food restaurant like McDonald’s runs on a pull system, while a catering service operates a push system.

At McDonald’s, the customer orders a hamburger, the server gets one from the rack, the hamburger maker keeps an eye on the rack and makes new burgers when the number gets too low. The manager orders more ground beef when the maker’s inventory gets too low. In effect, the customer’s purchase triggers the pull of materials through the system; the customer initiates a chain of demand.

In a push system, the caterer estimates how many steaks or lobsters are likely to be ordered in any given week. He reckons roughly how long it takes to broil a steak or serve a party of four; he figures out how many meals he can accommodate and commits to buying what he needs in advance. He can account for the special event that he knows is scheduled to take place in midweek. In other words, the caterer gets a picture of production in his mind and pushes materials to where he expects them to be needed.

What Is JIT Anyway? When you drop everything to take care of a walk-in client, you are reacting, implicitly, to a pull system. When you plan for a meeting, you are in the push mode. What does either method have to do with JIT?

Nothing directly. Think of JIT as a statement of objectives. It underscores the importance of lead-time management in all aspects of manufacturing. It asserts that incremental reductions in lead times are crucial indices of manufacturing improvement. JIT presumes that to achieve such reductions the system should deliver to every operator, in any conversion process, whatever he or she needs just when it is needed. It saves the money tied up in downstream inventories, protecting against long lead times. Shorter lead times mean improved responsiveness and flexibility.

JIT promises to preempt the delays and confusion associated with the stack-up of materials. Correspondingly, it saves the money that would otherwise go into indirect labor for storing and moving work-in-process (WIP) inventory and storing and handling buffer stock.

An analogy to commuter traffic may be useful. Say a thousand cars have to get through the Lincoln Tunnel every ten minutes. Wouldn’t it be ideal for the economy (and sanity) of New Yorkers if each driver left home at precisely the right time to fall into the line at the entrance, so each car entered, one after another, like cars in a train? Would New Yorkers need six-lane divided highways leading to the tunnel if they could enforce this JIT ideal?

To be sure, there are a number of established materials control techniques—pull techniques—associated with JIT, particularly as Japanese companies have realized it. There are synchronized deliveries from proximate suppliers, who can also deliver more or less as production needs fluctuate. There are floor layouts and kanban systems in which materials are constrained to flow consecutively along predictable paths and at a pace determined, in effect, by the last operator in the chain.

These pull techniques, excellent as they are, should not, however, be confused with JIT’s governing principles. JIT aims to manage lead times and eliminate waste. There is nothing inherent in push systems that makes them incompatible with JIT. On the contrary. The goal is to get drivers to leave home at just the right time to take their places at the tunnel entrance. JIT advocates shouldn’t care whether the signal to leave comes as a computerized call or as a wave from the next-door neighbor.

The Limits of Pull. There is no denying that pull systems are very effective in disciplining production to meet demand just-in-time and, in certain environments, are more effective than push systems. But are pull systems inherently JIT? No pull system, in fact, can constrain workers consciously to produce just-in-time for some future event—because pull systems do not recognize future events. The inventory level triggers production; the pull system aims to fill up depleted inventory—whether it’s Big Macs or machined parts.

Pull systems are fine if your McDonald’s franchise is downtown with a steady daily stream of customers. But if you’re next to a football stadium, how can a pull system alone prepare you for the day of a game? Similarly, it’s easy to see how implausible a pull system would be in solving the Lincoln Tunnel’s JIT problem. Wouldn’t it be more sensible to stagger the cars according to some kind of push system—say, “blue cars leave at 7:45”—than to expect all commuters acting individually to leave for work in perfect synchrony?

There is a paradox here. JIT advocates admire pull systems and look askance at computer-governed push systems like MRP. Yet the latter inherently aims to be a JIT system, while pull systems do not really recognize the future events you are supposed to be just-in-time for. This confusion has a history.

Forty years ago, the most common control system was OP, OQ, a pull system that seeks to exploit the presumed efficiencies of batch manufacturing. It was not known for its quick responsiveness to customers. Inventory managers determined the point below which materials and parts should not fall, and clerks ordered stock whenever it fell below that point. You put a line in a pail of bolts; when you exposed the line, you ordered more bolts. The order was based on average demand, and the fixed trigger point for releasing an order was based on typical behavior.

These techniques were more or less suitable for managing retail operations, though even in those settings there were problems like the McDonald’s franchise on the day of a game. If you ran a garage, for example, it would be tough to restock snow tires quickly after a big blizzard because every tire store would hit its order point and there would be a flood of reorders. The OP, OQ system would not react and place an early order when a blizzard was forecast.

If OP, OQ could be annoying to retailers, it could be disastrous to manufacturers. For want of a part, whole lines were shut down. And job lot orders, when they finally did come in, often forced manufacturers to tie up cash in more stock than they needed. Nor did OP, OQ integrate all information that was available when initiating production: How long does it take to build a product? What is likely external demand? In fact, parts and components are often built for in-house demand rather than for outside customers, and their requirements are well-known.

Managers often took obviously inappropriate action because of the old OP, OQ method. Demand surges invariably caused shortages; small shortages triggered production of parts even when people knew there would be no near-term demand for the part. Against this flawed system, MRP—the first generation of what is now MRP II—had obvious merit and won wide acceptance very quickly.¹

MRP to the Rescue. The concept behind MRP was straightforward, and MRP II is no real departure. In the same way that the caterer plans and conceives the whole week’s production, the MRP system explodes the entire manufacturing operation into discrete parts making up the whole. It then projects demand, the time it would take to meet it, and the materials needed.

The key to MRP is that you have to tell it the lead time to manufacture a part, component, or assembled product. If parts production is intended to support, say, final assembly of a telephone, MRP orders only the parts that are actually going into the phones you expect to sell—not some preset job lot determined by “efficient scale.”

So instead of building according to a fixed inventory position of various parts, MRP mandates building to the scheduled delivery of the final product. That, at least, is the theory. The best feature of MRP is its demonstrated capacity to work through the bills-of-materials relationships by which parts and subassemblies become the final product. MRP calculations start at the end items to be shipped and proceed stage by stage through bills of materials, releasing orders for the various parts or assemblies, according to a predetermined quantity and timing. The process then automatically repeats for the next level of parts going into each planned component or assembly.

Penetration of MRP methods into manufacturing has been substantial, especially in industries characterized by complex bills of materials, large numbers of open orders, and many needs for materials coordination among plants, vendors, and customers. Indeed, MRP has become so much the standard for materials management that it has led to the professionalization of the task, as exemplified by the American Production and Inventory Control Society. At the same time, owing to the heavy computer demands of MRP, systems managers and MIS departments have taken over a good deal of manufacturing management.

Push Comes to Shove

MRP II—more exact than MRP before it—initiates production of various components, releases orders, and offsets inventory reductions. MRP II grasps the final product by its parts, orders their delivery to operators, keeps track of inventory positions in all stages of production, and determines what is needed to add to existing inventories. What more could JIT ask?

A major barrier to MRP, though, is the cost of hardware and software for a complex computerized system—no minor barrier especially to smaller producers. And even more important are the costs of training and implementation. You have to teach your workers a lot that they don’t know about computers. And in order to enter the right data and the right relationships, you have to spend a great deal of time finding out things about your system that you don’t presently know: How should parts be timed to be put together just the way you want them? How long will it take for delivery of all the critical parts?

MRP does not conflict with JIT, but MRP must assume a fixed production environment with fixed lead times. Even with the best intentions, people who set up MRP systems base them on established and often flawed methods of conceiving conversion processes—methods that could be full of inefficiencies and that could be easily improved if workers were not constrained by MRP expectations.

In this context, the professionalization of materials control is a handicap, not a benefit. MRP standards have become a kind of orthodoxy, so people resist the introduction of new methods to the shop floor. New methods can threaten the positions of MIS managers, materials managers, MRP vendors, consultants, and educators who have become attached to the standards.

Of the many standard assumptions made by MRP, the fixed lead times are the most troublesome. Why is MRP so susceptible to getting lead times wrong? The best answer is that production lead times vary depending on the degree of congestion or loading within the shop. The fallacy in MRP is that its releases produce the very conditions that determine lead times, but these lead times have already been taken as known and fixed in making the releases. Consider again the cars traveling to the tunnel. The time it takes for any one car depends on traffic conditions and starting times. Change the pattern of departures, and you change the load on the system and the time it takes.

There is another way of looking at this. A single lead-time number must suffice in MRP for all situations faced on the floor. Consequently, the number has to be set high enough to cover all variations up to the worst case. If an order is ever late, people have an incentive to increase the planned lead time in the system so that the delay does not occur again. A commuter who might encounter an accident or traffic jam will leave early to protect against such contingencies. Similarly, orders will usually be released too early and will often complete early, thereby increasing inventories in the system.

Incentives for Improvement. Perhaps the most pernicious aspect of MRP is the removal of any responsibility for lead-time reduction from the shop floor. How can there be incentives to reduce lead times if there are no rewards for completing work faster than MRP’s fixed standards say?

Another big problem with MRP is its unnecessarily complex and centralized nature. MRP II systems plan and coordinate materials flow and produce order releases to the shop floor. But in many situations, the shop floor can be more flexible than MRP II.

For example, an assembly group might want to change its build schedule because parts aren’t available for some current schedule. Yet change is stymied because the appropriate paperwork is unavailable and won’t be available until the next run of the MRP system—say, next week. It often makes no practical sense to run an MRP plan every day. It takes time to collect and distribute all the data involved. Also, a good-size MRP system can tie up the central computer for hours. The same computer might be used for everything from word processing to payroll and general ledger accounting. Yet some shops would be better off working in just such short cycles.

Some MRP enhancements have addressed these problems. MRP vendors have created “shop floor control” modules—actually monitors, not controllers, which track progress on the shop floor. The resource management tools in MRP II analyze capacity and resource loading. Perhaps the best known of these systems is “rough-cut capacity planning.” This method analyzes the load that MRP order releases create on the shop floor. If this load exceeds the capacity of a work center, the implication is that the work in the shop will not get done within the time allowed. The human planner must now find some way to cure the problem diagnosed. Sophisticated techniques for evaluating the lead-time consequences of MRP releases are also available now, including order release methods (X-FLO, for example), scheduling techniques (OPT, CLASS, MIMI), and simulations (FACTOR).

While helpful, these methods increase MRP costs and can be subject to the same criticisms as the system they are meant to restore: they remove responsibility and incentives from the shop floor and they are only as good as the information put into them.

Does Pull Ever Come to Tug?

If MRP superseded OP, OQ, the kanban method is often prescribed as a JIT technique that overcomes the deficiencies of MRP. Presumably, if you set up a production system that works like a bucket brigade, you can forget about providing incentives for continual improvement or gathering what may prove to be incorrect information. The team will discipline itself according to the next customer’s needs.

To the extent that kanban works like a bucket brigade, it is indeed a JIT system. Everyone in the chain takes about the same amount of time to pass a bucket, and the system can work without any inventories of buckets between people. If the output end slows down, the whole chain will react and slow down; if it speeds up, the chain will react and speed up as much as possible, until limited by the slowest bucket passer.

Nor is kanban just warmed-over OP, OQ. With kanban systems, workers can see clearly the value of lead-time reduction. Unlike other pull systems, kanban combines production control with inventory control. The interaction between lead times and inventory levels becomes obvious to everybody on the line.

Moreover, the production supervisor owns the inventories that are produced; they are not pushed into other hands. He or she is thus forced to recognize that increasing the lead time of manufacturing increases WIP as well as finished inventory. This is completely unlike conventional pull systems like OP, OQ in which the inventory management function is separated from production or replenishment.

Indeed, the kanban method of posting circulating work orders makes the current work commitment of the manufacturing cell immediately obvious to everybody in the cell. Planning setups in advance, therefore, or opportunistically consolidating batches to save setups can become routine. The mix changes and demand surges that call for personnel reassignments become more transparent.

Kanban has another virtue that JIT people like. The fixed pool of cards in a kanban cell reduces the extent to which demand fluctuations are passed on by the cell to other upstream cells. The cards provide an upper bound that filters out extreme variations. At the same time, the system disciplines the downstream customer by punishing wide fluctuations or demand surges. A sudden surge will not be satisfied until the limited number of cards circulate many times. This encourages uniform demand and level schedules on the downstream side.

Kanban Is Reactive. Kanban is not without difficulties, though, which show up especially when it is forced to operate in complex operations where variations are too great or too intractable to be disciplined easily. Toyota’s kanbans discipline suppliers, but a supplier’s kanban cannot discipline Toyota.

The kanban method works best where there is a uniform flow—a level-loaded, synchronous, or balanced system. It does not plan well. JIT enthusiasts should realize that when a kanban system is implemented in an environment full of variations in supply and demand, it is even less likely than MRP to operate in a stockless manner—that is, without a burdensome amount of WIP. Variability causes the same extreme problems that it does in other pull systems Extra cards or containers—buffers, for example—have to be introduced to cover variability and avoid back orders. Nothing in a kanban system magically reduces inventory levels due to some internal rule or formula.

Since the system is reactive, changes in demand level percolate slowly from stage to stage. Even if it is perfectly obvious that demand is rising, there is no standard way to prepare for the situation. Some U.S. assemblers that work with Japanese suppliers using pull systems have commented that if there is a steep change in demand levels, the suppliers take from three to six months to adjust to it and encounter plenty of problems until the system reaches smooth operation again.

Tailored Controls, Hybrid Systems

Where does all this leave us? Which system should the manufacturing manager choose? The simple fact is that there is no need to choose between push or pull. These methods are not mutually exclusive, and each has its pros and cons. The best solution is often a hybrid that uses the strengths of both approaches.

Pull methods tend to be cheaper because they do not require computerization—hardware or software. They leave control and responsibility at a local level and offer attractive incentives for lead-time management. MRP systems are good at materials planning and coordination and provide a natural hub for inter-functional communication and data management. When it comes to work release, they are good at computing quantities even if they are weak on timing. A successful hybrid system can use each approach to its best advantage.

The key to tailoring production control lies in understanding how the nature of the production process drives the choice of control method. The accompanying exhibit summarizes various manufacturing control methods and process characteristics. For a continuous-flow process, ongoing materials planning is not essential and JIT supply techniques work well. Order releases do not change from week to week, so a rate-based approach can be used. At the shop floor level, JIT materials-flow discipline combined with pull release—kanban, for example—is effective.

In a repetitive manufacturing environment with fairly stable but varying schedules, materials planning can be a combination of MRP II and JIT methods. Order release may require MRP calculations if changes are frequent or if it is necessary to coordinate with long lead times or complex materials supply and acquisition. Pull methods work well on the shop floor.

As we move to more dynamic, variable contexts—like job shop manufacturing—MRP becomes invaluable for planning and release. Pull techniques cannot cope with increasing demand and lead-time variability. Shop floor control requires higher levels of tracking and scheduling sophistication. Materials flow is too complex for strict JIT.

Finally, in very complex environments, even job release requires sophisticated push methods. Where these are too expensive, the only option is to live with poor time performance, large inventories, and plenty of tracking and expediting.

The Best of Both. The dividing line between push and pull is obviously not sharp. In many situations, the two can coexist and are complementary. Most important, it is perfectly possible to take elements of one system and add them to the other. If pull systems have natural lead-time reduction incentives and push systems do not, for example, there is nothing to prevent managers from instituting a program of incentives in the context of a push system. Given the importance of lead-time reduction, in fact, it is crucial for managers to measure lead-time performance and provide feedback on response and turnaround times to each work center and shop. Though MRP systems do little to encourage good lead-time performance directly, managers can introduce measurement and incentive schemes based on MRP’s data collection capabilities.

There is nothing to stop managers from compensating for the deficiencies of pull systems either. Pull systems, for instance, have no means of lot tracking—pegging lots to specific customers. But customers may want to keep track of their orders, and there may be special regulatory or quality control reasons for maintaining a lot’s identity. So why not add lot tracking and data collection systems to a kanban line, leaving the release function as a pull system? (One simple and effective approach is to accumulate the information physically, with the lot itself as it moves through various process stages, and then record it electronically at inventory points in the process.)

Theoretically, there is no limit on the variety of control methods that can be developed. Most are hybrids. Attempts to implement pure push systems are usually accompanied by the growth of some informal, reactive pull procedures. The most common, alas, is the “hot list,” by which assembly tells manufacturing what parts it wants most on a given day.

In a way, such informal procedures are only piggybacking on the official MRP system, using short-term release information that MRP has not yet processed. The trouble with any informal procedure, however, is that it is very unsystematic; it may be based on assembly’s guess of what it can get from parts and does not take into account the actual position of open orders in parts. Moreover, it undermines the credibility of the official system. Since there can be no coordination between the two, disbelief in the official system becomes self-fulfilling. Instead of such informal overrides of MRP II, consider one of the following hybrids.

JIT-MRP. There are now several modifications of existing MRP II systems, which add pull elements and remove some of the problems connected with the system’s lack of responsiveness. Some such modifications are “synchro-MRP,” “rate-based MRP II,” and “JIT-MRP.” These systems are appropriate for continuous-flow or level-repetitive processes, where production is at a level rate and lead times are constant. In these situations, the order release and inventory management functions are of little value. The facility can be designed to operate in a JIT manner so that any material that enters the facility flows along predictable paths and leaves at predictable intervals. Work is released by a pull mechanism, so there is no WIP buildup on the floor.

Such a JIT-MRP line produces to meet a daily or weekly build rate rather than build to specific individual work orders. This means that inventory position isn’t necessary for release calculations. Inventory levels can be adequately calculated after the fact on a so-called “back-flush” or “post-deduct” basis by subtracting to allow for production that has already taken place. In short, MRP serves mainly for materials coordination, materials planning, and purchasing and not for releasing orders. The shop floor is operated as a JIT flow system.

Tandem Push-Pull. In a repetitive batch environment where lead times are fairly stable, either an MRP or a pull approach can achieve order release. MRP would be best for purchase planning of items with long lead times. Actual build routines closely correspond with the MRP II schedule, yet the timing of subassembly and assembly releases can be eliminated to allow the shop floor to change rapidly in response to short-term demand pull. Subassembly and assembly are flexible, short-cycle processes that can easily be run on a pull basis.

In this common situation, push and pull systems can simply be juxtaposed—MRP II to ensure parts availability based on end-item schedules and kanban for actual subassembly and assembly releases. MRP can be run only as frequently as necessary for parts purchasing and planning. Since the floor schedules can change quickly, the MRP database will always be playing catch-up with actual part withdrawals. This approach has been particularly successful in subassembly and assembly environments in which manufacturing cycle times are much shorter than parts purchasing and fabrication lead times.

Requirement-driven Kanban. Consider another situation where individual cells within the manufacturing chain can be run with kanban control, although MRP II runs much of the rest of the process. This can occur where final assembly schedules are unstable with respect to volume and mix, yet certain portions of the production process see fairly steady demand. A plastics injection molding cell that makes the same bottle for different shampoos is a good example. The MRP system can predict requirements for plastic parts quite well; kanban could run the injection molding cell.

One approach for such a case is to use MRP II to plan the number of cards in the cell on the basis of the gross requirements for all the parts produced by the cell. The MRP system doesn’t have to monitor the inventory level in the cell or match demand with available inventories since the system doesn’t make order releases. The gross requirements are an aggregate forecast of demand from the cell. Of course, as the gross requirements increase, additional cards are introduced into the cell in advance of the demand increase. They are withdrawn as the requirement level drops. MRP thus plays the role of planning adviser to the cell, setting the budget level in terms of the number of cards but not specifying the “expenditure” or release of the cards.

Many component manufacturing shops supplying subassembly and assembly operations, where the mix may change substantially but the total volume does not vary much, can use this approach. Other users are builders of common components or subassemblies like motors, similar components like PCBs, and metal-forming operations like blanking, shearing, and pressing.

Dynamic Kanban. Pull methods like OP, OQ typically do have some push component, such as seasonal expectations. Forecasts of demand patterns can be used to set new values for the order quantity and for the order point. In this manner, the otherwise passive pull system is able to anticipate predictable changes.

Similarly, the card quantity in a kanban system can be altered in response to regular changes in demand forecasts—not only seasonal variations but obvious trends or planned promotions. In these cases, the forecast can be used to calculate the number of cards necessary to support the changed level of demand. The cards become a planning parameter driven by forecasts of activity.

Looking Ahead to CIM

There are no panaceas for manufacturing management problems. A single approach will not suffice for all situations. Managers have to design and refine solutions. Kanban itself, like so many JIT techniques, evolved over many years.

I believe that future advances in pull systems will most likely accommodate even more computerized and automated factory environments. The challenge will be to create incentives for process improvements in the automated factory. Expert systems will have important roles in troubleshooting and diagnosing problems, sometimes even substituting for shop floor supervisors.

The fastest growing area for push methods is “factory management systems”—new methods oriented toward shop floor management rather than materials planning. These new systems monitor manufacturing and collect production data and merge with technologies like smart cards and bar coding. Even newer techniques of scheduling and cell management are leading to a bottom-up style of factory management. Indeed, as information technology evolves further, push techniques, like the pull approaches, will tend to decentralize control to the local, cell level.

In today’s manufacturing settings, we are witnessing a drift toward the ultimate JIT factory, in which the needs of a JIT cell are perfectly coordinated with the output of all others and matched to customers’ varying demands. Expressed in those terms, that’s the ultimate CIM factory too.

1. Early HBR coverage included Jeffrey G. Miller and Linda G. Sprague, “Behind the Growth in Materials Requirements Planning,” September–October 1975, p. 83.

A version of this article appeared in the September–October 1989 issue of Harvard Business Review.