Prologue: a lighthouse does not steer the ship
A lighthouse sends its beam through the fog. But that is all a lighthouse does. Reading the rocks, turning the helm and adjusting speed in the swell are, in the end, the work of the people on board. When the World Economic Forum (WEF) decided at Davos in 2018 to seek out the factories leading the adoption of Fourth Industrial Revolution technologies, and chose to call them “lighthouses,” the metaphor could hardly have been more apt.1 They show the way; they do not travel it for you.
Yet the discourse on manufacturing innovation often remembers only half of the metaphor. Each time a new list of lighthouse factories is announced, many companies copy down the technologies those factories use: digital twins, machine vision, generative AI. The scene is familiar. In the 1980s, Western companies copied down Toyota’s kanban cards, and many of them failed. The tools travelled; the way of thinking that produced the tools did not.2
This essay is an attempt to test a narrative heard often of late: that “the era of the Toyota Production System (TPS) is over, and the era of the Global Lighthouse Network (GLN) has arrived.” To state my conclusion first, I believe the narrative is only half right. The GLN has undoubtedly added something new. But the evidence it has gathered over eight years points most strongly, and paradoxically, to the place TPS has been pointing to for more than seventy years: the person who learns on the shop floor.
1. Reframing the question: what are we comparing?
We should begin by clarifying categories. TPS is a production philosophy and operating system that a single company evolved over decades. As Toyota itself explains, its roots lie in Sakichi Toyoda’s automatic loom and Kiichiro Toyoda’s idea of just-in-time production; Taiichi Ohno laid its foundations through repeated trial and error at the Honsha Machinery Plant from the late 1940s to the early 1950s, and it was rolled out to all plants from 1960.3 The GLN, by contrast, is a recognition and learning network that the WEF co-founded with McKinsey & Company, a forum in which designated sites share their experience with one another.4
One is a system for how things are made; the other is a mechanism for selecting and sharing who is making them well. Placing the two on the same scale to declare a winner is a kind of category error. The better question is this: what in TPS does the GLN’s eight years of evidence refute, what does it reconfirm, and what does it add?
The term “paradigm shift” also calls for care. When Thomas Kuhn introduced the concept, a paradigm shift was a rupture accompanied by incommensurability, in which old and new frameworks could not easily be measured on a common scale.5 If the GLN’s success factors rest on the principles of TPS, what we are witnessing is not a rupture but cumulative evolution. That is the hypothesis of this essay.
2. What TPS left behind: two pillars, and people
According to Toyota’s official description, TPS stands on two pillars.7 The first is jidoka (自働化). As its Japanese characters suggest, it is not mere automation (自動化) but “automation with a human touch.” When an abnormality occurs in equipment or quality, the machine stops itself or a worker pulls a cord to halt the line; the root cause of the problem thus exposed is then removed to prevent recurrence. The second is just-in-time (JIT): the principle of flow, making only what is needed, when it is needed, in the amount needed. Toyota describes the premise of the whole system as making work easier for workers.7 Jidoka itself originated in Sakichi’s invention, born of a wish to ease his mother’s labour at the loom.8
Ohno classified the waste (muda) that obstructs flow into seven types: overproduction, waiting, transportation, over-processing, inventory, motion and defects.9 The “unused human creativity” often added as an eighth type is not part of Ohno’s original classification; it was added later by Western lean literature to make explicit TPS’s spirit of respect for people.10 This seemingly minor distinction shows that TPS is not a system frozen in its original form but one that has been interpreted and extended. The naming of “lean production” by MIT researchers in 1990, which made it a global standard, was one phase of that extension.11
Spear and Bowen asked why companies that imitated TPS failed. Their answer was that the essence of TPS lies not in tools such as kanban or andon, but in rules that specify all work as testable hypotheses and have shop-floor workers continually test them.2 Tools are visible; rules are not. That is why transplants that move only the tools fail.
The humanity of TPS has its shadows, too. Satoshi Kamata, who worked as a seasonal labourer at a Toyota plant in the early 1970s, documented the pressure of conveyor speed and work intensity.12 Adler, by contrast, studying NUMMI, the GM–Toyota joint venture, argued that standardized work could be either a tool that constrains workers or a tool for learning that workers themselves design and improve. He called the latter “democratic Taylorism.”13 Read together, the two accounts leave a single lesson: respect for people in TPS is not guaranteed by its principles. It is an outcome of how the system is run. As we will see, the same tension recurs in the digital factory.
3. Where do the limits of TPS come from?
Much has been written about the limits of TPS. Those limits come into focus, however, only when separated into three distinct kinds.
Limits of transplantation, not of principle
As noted, many failed TPS implementations stemmed from imitating tools. What is striking is that digital transformation fails according to exactly the same grammar. In their first lighthouse white paper in 2019, the WEF and McKinsey reported that more than 70% of industrial companies remained in so-called “pilot purgatory,” with only 29% deploying new technologies at scale.1 In 2025, researchers at MIT NANDA reported from preliminary research that 95% of organizations investing in generative AI saw no measurable P&L impact, attributing this not to model quality but to a “learning gap”: systems that do not retain feedback or adapt to context.14 Sample limitations aside, the direction is clear. The tools changed from kanban to AI, but the grammar of failure did not. Tools transplanted without a learning organization do not take root.
Limits of design conditions
JIT presupposes stable supply. The less inventory there is, the leaner the flow, but the more vulnerable it becomes to shocks. That vulnerability is real. Yet the story does not end there. After the 2011 Great East Japan Earthquake severed its supply chains, Toyota introduced a business continuity plan (BCP) requiring suppliers to hold two to six months’ worth of stock for long-lead-time components such as semiconductors. This measure has been cited as a reason Toyota suffered comparatively less than its rivals during the global chip shortage of 2021.15 The case says two things at once: JIT’s vulnerability is real, and TPS learned from it and corrected itself. Much of the criticism of TPS targets not the living Toyota but the TPS of 1980s textbooks.
The received wisdom that heijunka (levelling, 平準化) is ill-suited to high-mix production also deserves a second look. Toyota describes TPS as producing vehicles of different specifications one at a time on the same line.16 Levelling is not a principle that rejects variety but a technique for absorbing variety into flow. The limit lies not in variety itself but at the point where the speed and unpredictability of change exceed human planning capacity.
Limits of data scale: the genuinely new condition
Genchi genbutsu (現地現物), the principle of going to the actual place to see the actual thing, presupposes human eyes and feet. But in today’s factories, vast numbers of sensors pour out data by the second, and value chains span continents. This scale exceeds human cognition. Here AI becomes not a luxury but a necessity. Direction still matters, however. As Polanyi put it, we know more than we can tell.17 The tacit knowledge in a skilled worker’s hands cannot be fully converted into data. Data should not replace the shop floor; it should be a telescope that lets the shop floor see further and sooner.
4. Eight years of the Lighthouse Network: a verified chronology
Now to the GLN. Launched following a decision taken at the 2018 Davos Annual Meeting, the project scanned more than 1,000 manufacturers, contacted over 150, conducted site visits and, through an expert panel, selected its first 16 lighthouses.1 Ten more were added in July 2019,18 and by January 2020 the count had reached 44.19 In Korea, following POSCO (2019) and LS Electric (2021), LG Electronics’ LG Smart Park in Changwon became, in March 2022, the first Korean home-appliance site to join.20
The network grew to 114 sites in October 2022, 153 in December 2023 and 189 in January 2025, reaching 201 in September 2025.4 At that point the designation framework itself changed: instead of the factory, value-chain and sustainability categories, sites are now recognized for distinction in five areas: customer centricity, productivity, supply chain resilience, sustainability and talent.4 In January 2026, 23 new sites took the network past 220,21 and in June 2026, 16 new awards announced in Dalian, China, brought it to 238.22
I read this chronology in five phases. I should make clear that the phase divisions in the table below are not an official WEF or McKinsey classification but my own interpretation, based on shifts in emphasis visible in public sources.
| Phase and period | Shift in emphasis | Milestones confirmed in public sources |
|---|---|---|
| Proof of concept 2018–2019 | Demonstrating that 4IR technologies work at scale, beyond pilots | First 16 sites designated. White paper sets out two routes: production-system innovation and end-to-end value-chain innovation1 |
| Extension to the value chain 2020–2022 | Connecting the whole supply chain, beyond the factory walls | 44 sites in January 202019; 114 in October 2022. Sustainability lighthouse designations gain momentum4 |
| Designing for scale 2023–2024 | Designing for rollout from the outset | 21 new sites in December 2023, 153 in total. 82% of lighthouses cite scale as the most important design consideration23 |
| Rediscovering people and process 2024–2025 | Organizational absorptive capacity over technology | Investment in people and process about 4× that in new technology. Concepts of process debt and the scaling slump emerge24,25 |
| Rewiring operations 2025–2026 | Embedding AI in core operations and redefining performance categories | Five-category framework (Sept 2025)4; Lumina industrial AI platform launched (Jan 2026)21; 238 sites (June 2026)22 |
5. Is what the lighthouses discovered really new?
The lesson the GLN has lately stressed most is not about technology. According to McKinsey’s 2025 analysis, for every $2 lighthouse sites spend on technology innovation, they earmark $3 for reducing process debt and $5 for scale and adoption. That amounts to roughly four times their technology investment going to the foundations of people and process, whereas other manufacturers spend closer to one or two times.24 “Process debt” here refers to the inefficiency that accumulates when new technology is layered on without changing the way work is done: in effect, a tax that erodes the value of new technology.26 McKinsey further diagnoses that the wall today’s leaders face is no longer pilot purgatory but a “scaling slump,” in which early successes fail to translate into enterprise-wide impact, and observes that lighthouses are overcoming it through localized adoption in partnership with their frontline workforces.25
What would Ohno have said to these lessons? Probably that there is nothing new in them. That automating waste merely speeds up waste is long-standing common sense in the lean community. The essence of jidoka was never a device for stopping machines but a structure for exposing problems so that people could learn. Process debt is a new name, for the digital age, for muda, and frontline-led adoption is another way of saying genchi genbutsu and kaizen. The WEF’s very first white paper in 2019 already reveals this connection: BMW’s Regensburg plant, one of the first lighthouses, was presented as a highly advanced factory with lean processes that had used digital manufacturing to reach the next level of performance.1 The first lighthouses stood on lean foundations.
Academic research points the same way. In their systematic review of the relationship between Industry 4.0 and lean, Buer and colleagues noted that both make use of decentralized control and aim to increase productivity and flexibility.27 Follow-up research provided empirical evidence that lean and digitalization have complementary effects on operational performance,28 and Tortorella and Fettermann, analysing Brazilian manufacturers, reported larger performance improvements when lean practices and Industry 4.0 technologies were implemented together.29 German researchers had already proposed “lean automation” enabled by Industry 4.0 technologies in 2015.30 In short, the mainstream academic view is not substitution but combination.
The most compelling evidence comes from Toyota itself. In 2022, Toyota’s Production Digital Transformation Office developed an in-house “AI platform” that allows shop-floor staff without AI expertise to build the machine learning models they need, and released it to the plants.31 At the Takaoka plant in Aichi Prefecture, a model built on the platform now inspects the adhesive that robots apply, freeing two people who had spent their days on visual inspection for higher-value work.31 In presenting the project, Toyota refers directly to jidoka and JIT.32 Moving AI from the experts’ toolbox to the shop floor’s translates into digital form TPS’s long-standing principle that the people closest to a problem should lead its improvement. This is not the end of TPS; it is TPS taking up AI.
| TPS principle | Lesson shown by the GLN | Relationship |
|---|---|---|
| Respect for people; frontline-led improvement | About 4× technology investment in people and process; localized adoption with frontline workforces24,25 | Inherited |
| Elimination of muda | Reduce process debt first, then introduce technology24 | Inherited (new name) |
| Standardized work as the basis of improvement | Assetizing proven solutions and scaling them through standard platforms25 | Inherited and extended |
| Jidoka: stop on abnormality and remove the root cause | Predicting abnormalities in advance with AI and digital twins20 | Extended (reactive to predictive) |
| JIT and kanban: pull from the downstream process | End-to-end value-chain synchronization from factory to consumer21 | Extended (factory to value chain) |
| Heijunka: levelled, mixed-model flow | 58 refrigerator models produced simultaneously on one line20 | Inherited and extended |
| Use only reliable, thoroughly tested technology10 | Rapid proofs of concept and fast scale-up1 | Tension |
| No explicit metric | Sustainability and talent measured as independent performance categories4 | New |
6. So what is genuinely new?
Strong traces of inheritance do not mean there is nothing new. I see four contributions unique to the GLN.
First, a shift in the time horizon. If jidoka is the principle of stopping at the moment an abnormality occurs, digital twins and AI read the signals before it occurs. LG Smart Park’s intelligent process system collects and analyses factory data every 30 seconds to forecast line conditions ten minutes ahead, and uses deep learning to detect likely defects and equipment failures in advance.20 The move from reaction to prevention does not change the purpose of TPS, but it shortens the time needed to reach it.
Second, an expansion of scope. The January 2026 cohort included cases that stripped out multi-tier distribution networks and built value chains running directly from factory to consumer, sharply reducing lead times and inventory.21 Where the flow of TPS was designed within the boundary of the factory and its suppliers, the flow of the GLN runs from demand forecasting all the way to installation service.
Third, an expansion of the value being measured. Since the 2025 revision, talent and sustainability have become independent categories on a par with productivity.4 The definition of the talent category explicitly includes work design and safety. Schneider Electric’s Wuhan site, for example, cut onboarding time for new staff from 75 days to 15 through digital apprenticeships co-designed with schools and AI-driven personalized upskilling.21 I would add one point. The most honest indicator of being “people-centred” is safety. Safety shows first whether an organization treats people as an end rather than a cost. How much weight the lighthouses’ talent category gives this indicator remains to be seen.
Fourth, the infrastructuralization of learning. In January 2026 the WEF launched Lumina, an AI-powered industrial intelligence platform that consolidates eight years of lighthouse data and more than 1,000 industrial transformations. According to this data, 94% of successful transformations combine multiple technology domains, and lighthouses that pair technology with workforce and sustainability initiatives outperform peers by an average of 16% or more.21 Where TPS accumulated learning within a single organization, Toyota, the GLN seeks to make that learning a shared asset at the level of industry. In the same announcement, the WEF emphasized that competitiveness today is no longer defined by efficiency alone but by the speed with which organizations sense, adapt and respond, which captures the heart of this change.21
7. Re-examining a case: LG Smart Park
The more often lighthouse cases are cited, the more easily their numbers inflate. This essay therefore uses only the figures LG Electronics has officially disclosed about LG Smart Park. According to the company, the site was designated a WEF lighthouse in March 2022. It developed in-house an intelligent process system combining AI, big data and digital twins, and is equipped with overhead conveyors carrying materials of up to 30 kg, 5G-based logistics robots moving loads of up to 600 kg, and an intelligent automated warehouse that monitors its own inventory and requests replenishment. Robots handle welding that produces open flame and the assembly of refrigerator doors weighing about 20 kg. As a result, productivity rose by 20%, the time needed to set up a production line for a new model fell by 30%, and energy storage systems and a building energy management solution improved energy efficiency by about 30%.20
The point I find most noteworthy in this case lies elsewhere. Through modular equipment and AI, the site produces 58 refrigerator models differing in colour, size and destination market simultaneously on a single line.20 This is less a move beyond TPS than a digital elevation of the mixed-model flow that heijunka has always aimed for. The design choice to hand dangerous and heavy tasks to robots while people concentrate on observing and controlling the process is likewise not far from the spirit of jidoka, which began with Sakichi’s wish to ease his mother’s work.8
8. The shadows of the lighthouse: a critical reading
To use the lighthouses well, we also need to know how their light is made. Reading GLN materials with a researcher’s eye calls for caution on at least five points.
First, selection bias. The GLN recognizes sites that have succeeded. Sites that invested in the same technologies and failed are not recorded. Lighthouse results are therefore evidence of what is possible, not an average expected value, and one cannot conclude that the technologies observed at lighthouses caused their results.
Second, self-reported metrics. As the performance tables in the 2019 white paper already credit the lighthouse sites themselves as their source,1 the figures are largely submitted by the sites, with base years and metric definitions differing from case to case. Comparisons across cases require caution.
Third, the structure of interests. McKinsey is a co-founder of the GLN and a member of its advisory board. The WEF and McKinsey state that all designations are made by an independent peer-review panel of experts from industry and academia, and that whether a site received external advice has no influence on the outcome.4 Readers should hold these two facts together.
Fourth, geographic concentration. Of the 23 sites announced in January 2026, 16 are in China (author’s count).21 This reflects the dynamism of Chinese manufacturing, but it also suggests that lighthouse designation has become a kind of certification race intertwined with industrial policy. The number of lighthouses should not be read directly as a measure of national manufacturing competitiveness.
Fifth, the under-measurement of the human dimension. The 2019 white paper reported that lighthouses raised efficiency with minimal displacement of workers,1 but independent studies tracking work intensity, the restructuring of skills and long-term employment effects remain scarce. Just as the humanity of TPS was an outcome of how it was run, the humanity of the digital factory is determined not by technology but by operations. The European Commission’s articulation of “Industry 5.0,” which names human-centricity, resilience and sustainability as the direction for industry, can be read as a response to this gap.33
9. Now what? Five propositions for Korean manufacturing
Translated into the language of the Korean shop floor, the discussion above yields five propositions.
- A lighthouse is a coordinate, not a destination. The moment designation itself becomes the goal, an organization starts accumulating a list of technologies for the assessors. Lighthouse cases are most useful as coordinates for locating where one’s own organization stands.
- Budget for people and process alongside technology. The lighthouses’ ratio of roughly four to one is a reference point, not a prescription.24 But if the ratio stays at one or two, the organization is likely heading for pilot purgatory or a scaling slump.
- Fix the process before you automate it. Before digital investment, redrawing the value stream and eliminating the steps that should not exist is the cheapest way to avoid process debt.
- Let the shop floor build the AI. As Toyota’s AI platform shows, democratizing AI is a twenty-first-century form of jidoka.31 Improvement never stops only when those closest to a problem can build its solution.
- Measure people-centredness. Unless safety, skill formation, turnover and learning time are managed as metrics with the same weight as productivity, “people-centred” remains a slogan.
Epilogue: before the lighthouse, once again
The transition from TPS to the GLN is not replacement but extended inheritance. The GLN has added a predictive time horizon, a scope widened to the whole value chain, a view that measures sustainability and talent as performance, and an industry-wide learning infrastructure. Yet what sustains those results, as the lighthouses’ own eight-year record confesses, lies not in technology but in people and processes.
In the end, the lighthouse beam sends us back to the question Ohno held onto on the shop floor: what are people learning in this process? AI does not answer that question for us. It only lets us ask it faster and further. A lighthouse only lights the way. It is always the people on the shop floor who steer the ship.