Editorial: Toyota Production System practices as Fast-and-Frugal heuristics

Two Forum articles and an editorial in 2021 called for a rethink of how operations management (OM) scholars conceptualize the Toyota Production System (TPS) and Lean (the Western label given to certain elements of the TPS). In the lead article in that series, Hopp and Spearman (2021, pp. 10 and 11) observed that the evolution of Lean from a physics of flows to an organizational culture that supports “continual reduction of the cost of waste” requires us “to incorporate human behavior more scientifically.” They noted that “A more extensive, and largely untapped, resource is the wide array of cognitive research into heuristics and biases that has been developed by behavioral and decision scientists since the 1970s.” This brings to mind the description by Fujimoto (1999) of the TPS as a knowledge-management system, in contrast to the common understanding of the TPS (captured by the designation “Lean”) as buffer management. In this editorial, we continue the discussion started by Hopp and Spearman with a thought experiment in which we consider TPS practices as heuristics. An initial objective was to contribute to disentangling the TPS knowledge- and buffer-management roles, asking: Are buffer-management tools designed to support knowledge management, or do knowledge-management TPS tools exist to allow operations to run as lean as possible (i.e., manage buffers efficiently)? The heuristics lens revealed the mechanisms by which buffer removal can be used to create cues from the production environment that effectively inform decision making. More generally, we discovered that the exercise of interpreting TPS practices as heuristics provided insight into whether and how heuristics can contribute to an effective management of operations.

We analyzed a sample of common practices that have been observed to be used by Toyota as one approach to implementing the TPS: jidoka, andon, and kanban. These practices transform front-line employees into decision makers by clearly specifying the information to be considered and the decision rule to be followed in a precisely defined situation. The resulting heuristics can be described as “production” heuristics, as their objective is to contribute to the line running smoothly on a day-to-day basis. We then considered practices that Toyota has been observed to use to prepare the environment for the successful deployment of these production-heuristic practices, including, for example, respect for workers, gemba, kaizen, and “five whys”. These “exploration” heuristics are oriented toward problem solving through carving out regularities in what appears to be a chaotic landscape. Whereas the production heuristics use stopping rules to strictly limit the information to be considered and precisely define the decision rule, the exploration heuristics relax the search rules and strongly encourage the decision maker to maintain information in the decision process.ⁱ They also allow the goal of the decision process to be flexible. In the production context, humans may make the error of assuming that more information is always better. In the exploration context, humans may make the error of moving forward with a decision based on too little information. Heuristics can help to avoid both types of error: We see TPS practices as either limiting or augmenting the amount of information to be considered, either precisely specifying or explicitly refusing to specify the objective of the decision. In contrast to key performance indicators, kaizen encourages decision makers to think about what it means to make things better. The “five whys” instruct decision makers to keep asking questions even though they think they already know the answer. We will present examples in which TPS performance was decreased by failing to maintain these systems that cause exploration heuristics to avoid premature elimination of information and flexibility. Although conventional wisdom considers heuristics as always dramatically reducing information-in-use, our exploration of the TPS reveals that heuristics may direct decision makers to reduce or expand that information. TPS success can possibly be attributed in part to deploying heuristics that are designed to either produce efficiently or explore, with exploration heuristics creating an environment in which the production heuristics function well.

Gigerenzer et al. (1999) proposed a typology of heuristics that first divides “reasonableness” (rational decision making) according to whether rationality is bounded or unbounded (Simon, 1955). Bounded rationality—which underlies essentially all business decisions—requires the decision maker to reduce the information considered, along the lines that Savage (1954) described as a small world. Decisions made under bounded rationality set as their objective to satisfice (making a decision that is good enough, Simon, 1956) rather than optimize. Heuristics—the decision rules used in satisficing—can be more or less “ecologically rational,” that is, can vary in their ability to produce decisions that qualify as rational while requiring little in terms of data and computational capacity. Gigerenzer and Gaissmaier (2011, p. 454) defined a heuristic as “… a strategy that ignores part of the information, with the goal of making decisions more quickly, frugally, and/or accurately than more complex methods.” Rationality remains bounded for ecologically rational heuristics.

Ecologically rational heuristics—designated as “fast and frugal” by Gigerenzer, Todd and the ABC Research Group (1999)—have been observed to go beyond mere satisficing, sometimes performing as well as or better than optimization that uses considerably more data. Fast and frugal heuristics are exemplified by the gaze heuristic: a simple interception rule that can be used by athletes to catch balls when playing sports, by animals to hunt down prey, and suggested as a contributor to the Royal Air Force's victory over the German Luftwaffe in World War II (Gigerenzer, 2007; Hamlin, 2017). It may also have played a role in US Airways Flight 1549's spectacular life-saving water landing in the Hudson River in 2009 (e.g., Hafenbrädl et al., 2016). This heuristic considers only the angle of gaze (a single piece of information) and involves no mathematical analysis. “Fast-and-frugal trees” (e.g., Martignon et al., 2008) have been used in contexts such as medical, judicial, and military (e.g., Katsikopoulos et al., 2021). The “take-the-best” heuristic (Gigerenzer & Goldstein, 1996)—a lexicographic strategy for inference—has been observed to outperform extensive data analysis (Czerlinski et al., 1999; Gigerenzer & Brighton, 2009). In OM, Bendoly (2020) classified as fast and frugal the nearest-neighbor sequencing heuristic used in logistics, also heuristics used in project management that minimize either slack or processing time in assigning resources. He uses these examples to illustrate how restricting the information considered can yield a reasonably good decision that is easily determined.

Not all heuristics are fast and frugal. Heuristics are simple decision-making strategies that typically ignore much of the information that is potentially available. When that information turns out to be essential to making a good decision, not considering it may well produce irrational decisions, many of which can be attributed to a variety of biases. Hopp and Spearman cite hindsight, confirmation, and loss aversion as examples of bias in the context of Lean production. (see Eckerd & Bendoly, 2015, for an in-depth discussion of these biases in OM). Gray et al. (2017) identified a heuristic that they entitled “lowest per-unit landed-cost” in which the production-location decision was based on a single factor. Limiting the offshoring decision to this single factor—ignoring readily available information that would have brought to light offshoring risks—resulted in quality problems, loss of intellectual property, and other unexpected management problems that were sufficiently severe that the offshoring decisions were reversed and production reshored. In other words, basing the decision on a single factor may work well (as exemplified by the gaze heuristic), but—as in this case—may lead to bias. Furthermore, when the (biased) offshoring decision was reversed, it was done based on extensive data collection and analysis, closer to what Kahneman et al. (1982) referred to as System 2 thinking than to a heuristic. No effort was observed by Gray et al. to enlarge the production-location decision process that had led to the original decision to offshore based solely on minimizing the per-unit landed cost. The firms studied may thus be vulnerable to repeating the original myopic and biased decision. This example illustrates the potential for a heuristic to perform poorly in its current environment.

Heuristics are typically described as a statistical adaptation to a given context (as occurs in the development of some fast-and-frugal trees used in areas like medicine) or as a performance rule that indicates which activities to prioritize or delay (Browning & Yassine, 2016). They can be taught, learned by observation, and discovered via experimentation and trial and error.

Eckerd and Bendoly (2015, p. 5) referred to the tendency in the field of behavioral operations to consider cognitive limitations of individuals as yielding flawed mental models, contrasting that view with the more nuanced one by Katsikopoulos and Gigerenzer (2013) that heuristics may be either an asset (ecologically rational) or a liability (biased) when used in decision making. The fast-and-frugal heuristics research program (e.g., Gigerenzer et al., 2011) has contributed to the identification of heuristics present in decision making, and the characterization of what makes these heuristics fit and perform well in a particular decision environment. When a heuristic in use is observed to produce biased decisions, is debiasing better achieved by moving from the heuristic to a fuller analysis done by the decision maker (i.e., moving toward constrained optimization along the lines suggested by Little (1970)), by recalibrating the heuristic to improve its performance, or by moving between a production and an exploration heuristic?

SIMON (1955, 1956) developed the idea of bounded rationality, introducing the idea of satisficing as an alternative to optimizing. To economists, Simon (1955) emphasized the cognitive capacity of decision makers, and to psychologists (Simon, 1956) the environment (see Petracca, 2021, for a discussion of this division). Simon (1990, p. 7) brought these two sides together as he wrote, “Human rational behavior (and the rational behavior of all physical symbol systems) is shaped by a scissors whose two blades are the structure of task environments and the computational capabilities of the actor.” Heuristics are rules of thumb that can typically be decomposed into search, stopping, and decision rules. The ecological rationality of a given heuristic depends on how well these rules allow the “scissor blades” formed by the task environment and computational capacity of the actor to operate together. Production heuristics typically emphasize search limitation, whereas exploration heuristics may encourage maintaining more information in the decision-making process: A well-functioning heuristic may increase or decrease the information in use depending on the context.

The description by Daft and Weick (1984, p. 289) of organizations as “interpretation systems” provides useful input to understanding the interactions of the search (environment) and stopping (cognitive capacity) dimensions of heuristics. Daft and Weick defined interpretation modes in terms of (1) whether or not the goal of interpretation is to identify a right answer that is assumed to exist, which determines whether or not the environment is seen as “analyzable,” and (2) whether the organization relates to the environment intrusively or passively. The resulting 2 × 2 matrix is reproduced in Table 1.

The TPS emerged from Japan in the late 1970s (e.g., Sugimori et al., 1977) and quickly captured the attention of the entire world. Toyota overcame strong competitive disadvantages (such as the need to transport cars from Japan, and access to considerably fewer resources for research and development) to take market share from companies like General Motors. A key difference between the TPS and mass production concerned how to adjust the upstream production rate to match what the line needed or was able to handle downstream. Consider a downstream problem that caused in-process inventory to accumulate. TPS practices were designed to highlight problems arising so that attention would be directed to solving them, while also matching upstream production to downstream capacity. Rather than setting an objective to perform well with respect to traditional performance indicators like utilization or output, the TPS instead set as an objective to equip the employees encountering a production problem in their immediate environment to avoid inventory buildup and contribute to getting the problem fixed.

Mass production, in contrast, was based on a decision rule that called for output maximization irrespective of conditions in the environment. Under mass production, workers were expected to focus on their tasks rather than pay attention to machines that were not functioning correctly. A worker who was not able to complete their assigned tasks during a cycle was obliged to leave the work incomplete, resulting in a unit that did not conform to specifications, while a worker who was able to fill the intermediate buffer between their and an adjacent workstation was viewed as a good performer.

We here analyze three TPS tools that were designed to stop production when downstream needs decreased (either because demand was met, or because of a production problem): jidoka (also known as “autonomation,” or automation with a human touch), andon, and kanban.

Jidoka combines human intervention with automation. When an automated machine develops a problem (e.g., a piece gets stuck, it runs out of material, or it goes out of alignment), under jidoka, it is designed to stop automatically and signal to a nearby worker that action needs to be taken. When the signal is given, the worker either fixes the problem or notifies maintenance that repair is needed. The key difference with mass production is that the worker is personally involved in getting the problem fixed or getting its existence communicated rather than ignoring the problem to focus on maximizing output at their workstation. Monden (1983) described jidoka as often being used on a process with some degree of automation, but also being used as a concept in a manual process.

The andon cord provided at each workstation allowed the worker to flag a production problem and request help from a supervisor: A worker who saw at the 70% mark of the cycle that the work would not be completed would pull the cord and have a supervisor sprint over to provide help. If the supervisor was able to help get the work completed within the cycle, the cord was pulled a second time and the line continued. If, however, the problem was not yet resolved, the line would stop at the end of the cycle. Creation of these line-stoppage decision rules not only avoided assembly of a defective unit, but also provided clear data as to where workers on the assembly line were most likely to be stressed. Monden (1983) considered andon to fall under the general category of jidoka, but it was considered in the West to represent a quite startling departure from normal assembly line operation, both in giving workers the right to stop the line and in workers being willing to admit that they were not keeping up—knowing that this personal performance-related information would be collected and analyzed by management.

The kanban system was designed to limit the buildup of inventory between two adjacent workstations. An upstream worker is only allowed to begin production of an item if an unattached kanban is available. An inventory buffer between the two workstations is able to buffer to some degree, such that the effect of temporary slowdowns at one workstation on the adjacent workstation would be minimized. Once there are enough kanbans to buffer temporary slowdowns, adding more kanbans would only serve to increase system waiting time for the pieces in inventory. Toyota went a step further and implemented a system in which the number of kanbans was gradually reduced to draw attention to production-line imbalances. When one workstation would block or starve the other, the blocked or starved workstation would provide feedback that was expected to lead to learning. It is in the kanban system that we see most clearly Toyota's understanding of the relationship between buffer inventory and learning. While inventory buffers were used to smooth flow, there was a constant awareness of the ability of inventory to hide problems and line imbalances, and that careful management of inventory could lead to process improvement (see Suri & de Treville, 1986, for an in-depth discussion of the relationship between the exploratory stress created by reducing this buffer inventory and learning).

By extracting their search, stopping, and decision rules, these three tools can be conceptualized as heuristics, as shown in Table 3. They allowed Toyota to deploy the cognitive capacity of its entire workforce toward smoothing production of high-quality products. Not only were the heuristics themselves fast and frugal, but they also brought into use a massive cognitive capacity that tended to be neglected in mass production. Returning to the Daft and Weick (1984) distinction between active and passive interpretation systems, we suggest that the TPS represents active interpretation in contrast to the passive interpretation encouraged by mass production. The assembly line became an analyzable world in which front-line employees could confidently contribute to the company functioning well, because well-calibrated heuristics made clear to them what they were to do where, under which circumstances. There was no need for counterfactual inquiry, because the assumption that the local environment was analyzable yielded ecologically rational decisions: These three tools fit the description of production heuristics.

Two observations arise from this analysis. First, these production heuristics performed well when an active interpretation system was combined with an analyzable environment. Consider a front-line employee who observes a problem (e.g., defective raw material, not being able to complete their operation by the end of a cycle, or that their speed is blocking or starving an adjacent workstation). On a traditional line, the employee may observe the problem but is not in a position to take action to resolve it, either in the immediate or longer term. TPS practices enable the employee to take action by pulling the andon cord, reorienting a part correctly, and organizing with the adjacent workstation to rebalance capacity. Thus, one outcome of TPS is to make the employee interact more intrusively or actively with the local environment. Active interpretation then enables the cognitive capacity of the employees to be made available. Establishment of an analyzable, local environment then allows disconfirmation as the dominant method of inquiry in these decisions without risk of confirmation or other bias. Second, these well-calibrated heuristics produced rational and profitable decisions, contributing to a level of performance that continues to astound decades later.

Our thought experiment is built around the idea that key TPS practices can be conceptualized around the selection, design, and calibration of heuristics to increase the ecological rationality of the resulting decisions. Our above discussion suggests that this exercise should include consideration of whether—and where—the environment is analyzable. Where it is not analyzable because of changing goals and cues, then from Feduzi et al. (2022) we would expect to see counterfactual reasoning yielding activities like experimentation and trial and error (e.g., Daft & Weick, 1984; Sommer et al., 2009; Thomke, 2003). These counterfactual-reasoning activities would then be expected to lead to the discovery of analyzable sub-worlds that would lend themselves to formal search and disconfirmation as a method of inquiry. Conceptualizing TPS practices like autonomation/jidoka, andon, and kanban as heuristics reveals how meticulously the environment has been prepared to be analyzable: We can identify among the TPS practices not only well-calibrated heuristics but also practices that encourage exploration and experimentation that appears to be explicitly designed to identify analyzable sub-worlds that are stable enough for disconfirmation to operate without bias, and in which a heuristic can operate effectively.

In Table 4, we evaluate the effect on the environment of selected TPS practices that encourage search over setting the kind of well-defined stopping rules common to the three practices that we categorized as production heuristics. The TPS emphasis on respect for workers serves to increase overall cognitive capacity available to the organization. Muda (identifying and eliminating activities that do not add value) and muri (setting a policy to not require workers or equipment to run at an excessive pace or for an excessive duration) also serve to protect available cognitive capacity from being used ineffectively on non-value-adding tasks such as those created by unplanned downtime and product defects. Search and experimentation are encouraged by practices like gemba (decision makers go in person to observe what is happening where a problem is occurring), kaizen (a constant search for improvement by everyone everywhere in the organization), and “five whys” (encouraging problem solvers to ask “Why?” five times rather than immediately accepting the first answer as the root cause of the problem). TPS practices like heijunka (leveling demand so that the flow of work to the production line is stable), standardization of tasks, poka-yoke (organizing tasks, tools, and processes to make them “foolproof” and reduce the likelihood of errors), and mura (identifying and eliminating sources of variability that do not add value) improve visibility in the production process, increase the signal-to-noise ratio, and make the local environment analyzable. In particular, heijunka's artificial removal of external variability enables the local environment to become analyzable in terms of remaining sources of internal variability. The TPS devotes considerable attention to the effect of inventory buildup on how the organization interprets its environment. As discussed above, this takes two forms: first, ensuring that inventory does not build up as a result of a production imbalance or a large lot size, and second, use of exploratory stress to encourage local process improvement. The TPS practices of lot-size and setup-time reduction combine with use of kanban systems to avoid unnecessary inventory buildup and permit exploration of improvement possibilities. The kanban system serves as a heuristic device to make clear to workers when to commence or refrain from production of a piece.

These practices combine active interpretation with the assumption that the local environment is less analyzable, which then calls for counterfactual reasoning that takes the form of experimentation, trial and error, and invention. The combination of well-specified decision rules with mechanisms to increase available information and allow space for some environmental unanalyzability yields exploration heuristics. Monden (1983) and Sugimori et al. (1977) described the development of the TPS as relying heavily on trial and error. This is in sharp contrast to the fact that front-line employees are told exactly what to do when faced with an immediate problem on the line, where processes are documented and workers are expected to adhere exactly to those documents (Spear & Bowen, 1999). In fact, trial and error on the production line by front-line employees is strongly discouraged: A front-line employee that has a process-improvement idea is encouraged to submit the idea for testing, and the decision about whether to test or implement is made higher up in the organization (de Treville et al., 2005; de Treville & Antonakis, 2006). Each of the TPS practices we have considered can be clearly assigned to the production or exploration category, and is intended to operate either under disconfirmation or counterfactual reasoning. In the next subsection, we explore two examples from the TPS literature in which disconfirmation continued in a context in which the organization should have switched to counterfactual reasoning.

In this editorial, we have considered the TPS as an example of development and calibration of heuristics, suggesting that Toyota's ability to redefine competition in the global auto industry came in large part from Toyota's skill in managing and creating an appropriate environment for these heuristics. Toyota defined stop, search, and decision rules, creating heuristics that allowed them to successfully deploy the cognitive capabilities of front-line employees and contributing mightily to the TPS as a knowledge management system.

These heuristics did not arise spontaneously, but were described by Monden (1983) as resulting from many years of trial and error, with Toyota transforming the environment to facilitate exploration that was expected to result in analyzable sub-environments (local environments) in which heuristics could be used without the risk of bias. Rather than knowledge management in the service of buffer minimization, we see skillful use of inventory buildup to identify problems and maintain the spotlight on those problems until resolved. More generally, practices to prepare the environment for effective counterfactual reasoning served to increase cognitive capacity, expand search, avoid premature search truncation, and improve the signal-to-noise ratio. It is also worth mentioning Toyota's emphasis on separating problems from people, in contrast to the usual assignment of blame (This emphasis on blame at General Motors and Ford is captured well by MacDuffie, 1997).

These TPS practices took the form of production or exploration heuristics, depending on the need to enact the environment and remain open to adjustments to the “right answer” in use. Whether heuristics represented an appropriate decision process depended on which method of inquiry was appropriate, which in turn depended on whether the environment was reasonably assumed to be analyzable. If the environment was analyzable, then production heuristics were defined to permit quintessential fast-and-frugal decision making. If the environment was unanalyzable, exploration heuristics encouraged rich use of data and counterfactual reasoning. This is in contrast to the unfortunate blend of disconfirmation and counterfactual reasoning at General Motors and Ford that resulted in effort that did not result in solved problems. We also saw a case in which even Toyota was caught unaware of a change in the environment that made it unanalyzable: Inventory built up for several days, but eventually the inventory buildup led management to toggle from disconfirmation to counterfactual reasoning, and the problem was solved.

Let us return to the question: When a heuristic in practice is observed to produce biased decisions, how can it be recalibrated? A heuristic that is producing biased decisions suggests a need to shift from disconfirmation to counterfactual reasoning to encourage search. Observing Toyota, two things come to light. First, when a heuristic was observed to produce biased decisions and search was expanded, Toyota did not expand search only in terms of how much data was analyzed, but rather encouraged decision makers to be present in the problem and think widely and deeply about what was going on. They made efforts to maximize the quantity and diversity of the cognitive capacity available to process the problem. Second, the transition to counterfactual reasoning was not intended to be permanent, but rather to allow decision makers to organize and play with available data that would eventually get “routinized” (see the discussion by MacDuffie, 1997).

Toggling back and forth between disconfirmation and counterfactual reasoning not only gives insight into how to benefit from the ecological rationality of well-calibrated heuristics, but may also help in answering Little's (1970) call to facilitate managers making use of stylized models. Like production heuristics, stylized models rely on disconfirmation as a method of inquiry and can be ecologically rational when their assumptions are reasonable. Just as decision makers need to learn to trust and calibrate fast-and-frugal heuristics, so they may need to learn to trust and calibrate stylized models to ensure that they retain the appropriate information. And, when decision makers neglect to use apparently rational stylized models, it may be time to encourage counterfactual reasoning: Why is this stylized model not being used here? Thus, rather than replacing heuristics by models, successful heuristic use may provide insight into how to make models more useful to decision makers. As we gain this ability to toggle, we will be increasingly able to gain the ecological rationality of production heuristics in an analyzable environment, and the focused search capability provided by exploration heuristics in environments that are not analyzable.