Culture Shock and Recompiling AI Organization of Japanese Company Culture
Moving back and forth between large companies and startups — both Japanese and foreign-owned — I keep encountering a strange sensation every time I join a new firm. It is the feeling of two things at once: the friction of finding that the ways of my old job no longer apply, and the nostalgia of finding that the conventions of an even older workplace suddenly work again, as though a trait had skipped a generation. Each time, it makes for a good chance to learn the cultural differences between companies. But after enough of these experiences, a more analytical thought takes shape — a way of seeing, grounded in real experience, the larger differences in approach and politics: country by country, and startups versus large corporations.
One such observation is that administrative processes and onboarding tend to be far better organized at Japanese startups than at large foreign-owned corporations. Even compared with foreign-owned firms staffed mostly by Japanese employees, the Japanese startups still tend to feel more orderly. At Japanese companies, the first-day paperwork, expense reimbursement, the routing of applications, the points of contact for questions — all of it is carefully designed, and each employee’s administrative productivity feels high. Foreign-owned firms, by contrast, have rough internal processes even when the company operates at global scale, and I usually get lost. Every time I move between the two, this gap gives me a small culture shock.
And yet it is the foreign tech companies that make a name for themselves around the world. Of course, one large and rather blunt factor is that the scale of AI and startup investment in the U.S. and China dwarfs Japan’s. It is also surely true that, in the shadow of every fast-growing startup that becomes globally famous, countless companies in every country quietly fail without anyone ever hearing of them. Even so, it is clear that the leading examples of tech startups succeeding across borders in the age of AI are concentrated in the U.S. and China. In this essay, I want to dig into this tendency — and this apparent contradiction — from the angle of “organizational culture,” and in particular the organizational thinking around how companies engage with AI.
What is really at stake in organizational theory for the AI era is neither “which culture is superior” nor “whether you can eliminate friction.” What matters, I believe, is when and where you take friction on, how much of it you can convert into “learning,” and whether your people can survive within it. As a tendency, Japanese companies try to eliminate friction “before” adoption. U.S.-style tech companies more often run with the friction left in place “after” adoption, while Chinese-style companies seem to wash friction away through “sheer volume of execution on the ground.” This essay digs into these per-country tendencies — and a tendency broadly shared across companies is, after all, what we call culture.
On Generalizations and Gradients
From here on, I will repeatedly use broad subjects like “Japanese companies,” “U.S. companies,” and “Chinese companies.” But this is in no way a claim that “all Japanese companies are like this.” Every country has both extraordinarily meticulous firms and averagely messy ones. Even among Japanese startups, there are companies that put internal housekeeping on the back burner and simply keep piling on products. Plenty of Japanese startups and scale-ups have emerged that throw the whole organization into AI adoption and product development before tidying up internally, declaring themselves “AI-native.” Conversely, the U.S. has no shortage of large corporations whose procedures are organized to the point of being bureaucratic. In other words, the corporate culture of every country is really distributed along a smooth gradient.
Even so, the “center of gravity” of that distribution is, I would say, skewed differently from country to country. And it is the shift in that center of gravity that I experience as “culture shock” each time I change jobs. When I speak of “tendencies” in what follows, I would ask the reader to take it as referring not to a label pinned on any individual company, but to these differences in the center of gravity.
23× the Money, 2.7% the Gap
Let me begin with a pair of numbers that form the backbone of the argument.
According to Stanford HAI’s “AI Index,” released in 2026, U.S. private AI investment in 2025 reached $285.9 billion — more than 23 times China’s $12.4 billion. The number of newly funded AI startups in the U.S. alone came to 1,953, more than ten times the next-ranked country. And yet the performance gap between the top U.S. and Chinese AI models has narrowed from a 17.5–31.6 point spread as of May 2023 to just about 2.7%. China, moreover, installed roughly nine times as many industrial robots as the U.S. in annual terms (about five times as much in operating stock), and accounts for roughly 70% of the world’s AI patent filings.
There is a fact these paired numbers reveal: investment does not translate directly into “a capability advantage.” That 23 times the money buys only a 2.7% gap is a sign that model capability (benchmark performance) is converging, and converging cheaply. And the closer capability comes to parity, the more the focus of competition shifts elsewhere. It moves away from the sum invested and toward something else — how quickly and how widely you can deploy a capability from experiment to real-world use, and how much learning you can recover from doing so. In technical terms, the focus shifts toward “diffusion” (the spread and density of deployment). The fact that the U.S. and China are pulling away from Japan in robot-installation scale and in frontline usage rates — which we will see below — is nothing other than this difference in deployment.
Administrative Work and Labor Productivity
Whenever Japan’s low labor productivity comes up, the conversation quickly turns to “too many meetings,” “slow approvals,” and “too much paper.” Yet my own sense is rather the opposite: on average, Japanese companies have better-organized administrative processes, and they feel more efficient to me. The number of meetings is not all that different, either. I will take this contradiction as my starting point.
According to the Japan Productivity Center’s “International Comparison of Labor Productivity 2025,” Japan’s labor productivity per hour worked in 2024 was $60.1, ranking 28th among the 38 OECD member countries. On a per-worker basis, too, it remains the lowest among the major G7 economies. This cannot be explained by “the administrative competence of individual employees” alone. Labor productivity is a composite of much larger factors — capital, technology, industrial structure, pricing power, and organizational efficiency.
Here, the fog begins to lift once we separate two kinds of productivity. One is “administrative productivity” — reducing friction inside the company — and Japanese companies are, on the whole, strong at this. The other is “market productivity” — making things that sell for a high price in the market — and at this, Japanese companies are, on the whole, weak. And the decisive point is that labor productivity is dominated by the latter.
Labor productivity is “value added ÷ labor input.” Subtraction and division — that is, cutting costs and optimizing how resources are allocated — are operations that shrink the denominator, and their lower bound is zero. There is a limit to how much waste you can cut. Addition and multiplication, by contrast — new products, new markets, pricing power — are operations that enlarge the numerator, and their only ceiling is the size of the market. Because labor productivity is dominated by the numerator, no matter how much you polish the inside of the company, no matter how much you trim the denominator, the ceiling on value added does not rise. To say that “addition has value” is not a matter of mindset; it is a structural consequence of the fact that the growth of value added is fundamentally additive, not subtractive.
Messiness and Adaptation
Before we judge the meticulousness so common in Japan as either “careful and good” or “excessive and bad,” I think we should first read it as “an adaptation to an environment.”
Meticulous process design that aims for a net of plus-or-minus zero is the optimal solution in a world where the pie does not grow — a “conservation strategy,” so to speak. “Just add for now,” by contrast, is the optimal solution in a world where the pie is expanding — a “growth strategy.”
Where the culture of tidying up, so deeply rooted in Japanese companies, came from, I cannot say. Perhaps it was an adaptation to the low-growth equilibrium of post-bubble deflation — the so-called Lost Three Decades — and of a shrinking population. Or perhaps it is a tendency broadly shared among Japanese people from much further back. I will leave that analysis to the historians. Either way, it seems to me that many Japanese companies today carry an organizational OS adapted to a world of equilibrium.
The problem is that this adaptation reinforces the very equilibrium it adapted to. Once you carve the premise of “no growth” into an organization’s OS, the muscles for growing atrophy. The traits that were correct as adaptations are, precisely because they were correct, the hardest to discard. To frame Japan’s problem as “laziness” or “sluggishness” is to miss its essence. This is a story about structure — about how a clever optimization becomes obsolete before anyone notices.
The Four Operations and Addition
Let me try, as an experiment, applying the metaphor of an organization’s “four arithmetic operations” to the tendencies of each country. Think of addition as increasing new AI tools, products, and experiments; subtraction as reducing unnecessary meetings, approvals, and legacy work; multiplication as spreading AI laterally across product, sales, and development; and division as sharing out responsibility, authority, and risk.
What most sharply divides the tendencies is “the direction of the question.” Facing the same AI, companies differ in which question they raise first and most forcefully. Japanese companies tend to ask, “May we bring AI into this task?” This is a yes-or-no gate, and its ceiling is fixed. U.S. companies, it seems to me, more often ask, “What can we sell with AI?” This is an open-ended question, and its ceiling is the size of the market. Chinese companies tend to ask, “How much can we run with AI?” This is a question of scale, and volume and learning keep turning over. Of course, every company in the world raises all three of these questions. But the “order” — which one you grasp first — is where the difference in each country’s center of gravity shows up.
The summary below simply lays out the “centers of gravity” of the distributions, axis by axis. And of course, every country has plenty of companies that fit none of these descriptions.
- Where friction goes — Japan eliminates it before adoption; the U.S. leaves it in place after adoption; China washes it away through sheer execution volume.
- Favored operations — Japan leans on subtraction and division; the U.S. on addition and multiplication; China on addition scaling into multiplication.
- The question asked first — Japan: “May we adopt this?”; the U.S.: “What can we sell?”; China: “How much can we run?”
- Strategy type — Japan: conservation (protect the denominator); the U.S.: growth (extend the numerator); China: diffusion (data compounds).
- Relative strength — Japan: trust, quality, and codifying know-how; the U.S.: go-to-market, capital, and creating options; China: speed, deployment density, and iteration volume.
- Common weakness — Japan: tidying becomes an end in itself; the U.S.: sprawl, key-person dependence, and shadow AI; China: over-deployment, control, and questions of quality and ethics.
The point is not to choose one of the three. It is to design in what order, and where, to place the quality common in Japan, the go-to-market common in the U.S., and the implementation density common in China. An AI organization, I believe, should hold these three tendencies as a portfolio, switching between them depending on the situation.
Running Wild and Compounding
It was more than ten years ago that I first joined a foreign-owned company. Having moved from a large Japanese corporation, I found the internal messiness dizzying — all sorts of things had been introduced and then left to run wild. At first I saw this “introduce it and then leave it alone” approach in a negative light. But once you flip it around, an entirely different picture comes into view.
First, it is “the purchase of options.” You add a great deal, you don’t organize it, what gets used survives, and what goes unused quietly dies — this is “evolution (market selection)” with no central designer, computationally cheap and robust to uncertainty. The “tidy it up” approach common in Japan is, by contrast, “intelligent design.” It is efficient when the designer has good information and the environment is stable, but in the AI domain, where the premises change every six months, selection often beats design. “Just add AI for now and leave it to selection” is not sloppy under uncertainty — it is rational.
Second, it is a “flywheel.” Deploying AI generates real-world data, the data improves the model, and the improvement calls forth the next deployment. The reality we saw at the outset — “23 times the money, yet only a 2.7% gap” — is precisely this compounding at work. What must not be overlooked here is that an organization that starts late loses more than just “time.” It loses, in its entirety, the learning loop that only a first mover can spin — the compounding of the data that deployment produces.
The actual data backs this direction up. In BCG’s 2025 survey “AI at Work,” about 70% of frontline employees in the Asia-Pacific region used AI regularly — well above the global average of about 51%. Optimism about AI is also highest in China. So-called “shadow AI,” proceeding outside the company’s formal approval, is spreading too. In short, in many companies the front line is already making heavy use of AI without waiting for the rules.
Turning to Japan, meanwhile, the IPA (Information-technology Promotion Agency)’s “DX Trends 2025” paints a highly suggestive picture from a Japan–U.S.–Germany comparison. Japan’s DX (digital transformation) stays at the optimization of individual tasks — “inward-looking, local optimization” — while the U.S. and Germany show a clear tendency toward enterprise-wide optimization, “outward-looking, holistic optimization.”
The substance of the outcomes is contrasting, too. Japan’s gains are mostly inward — “cost reduction” and “shorter delivery times” — whereas the gains in the U.S. and Germany were mostly outward: “higher revenue,” “higher profit,” “greater market share.” And what is telling is that the share of companies that have set metrics to measure their DX outcomes was under 30% in Japan, but over 80% in both the U.S. and Germany. Japanese companies have a tendency not even to “measure” outward outcomes.
“Addition” is not a euphemism for recklessness. It is the act of buying compounding options cheaply and in bulk.
Tidiness and Handicaps
At the outset, I wrote that organizational theory for the AI era also has an axis of “whether your people can survive within it.” So far I have looked at this from the organization’s side, but if we shift the perspective to the members of the organization, something important comes into view that bears on the answer to that question: whether messiness becomes a handicap depends strongly on the individual’s ability.
A reasonably capable person, even with the company left messy by addition, will sort the necessary information from the rest using their own judgment and the power of AI, and slip through with ease. Even when old documents are mixed in, they can tell what matters from what doesn’t; even when they don’t know whom to ask, they can make a good guess. For such a person, neither the result tidied carefully by the four operations nor the result left in disarray by addition amounts to much of a handicap. Whether things are organized is not, for an individual with survival skills, all that big a problem.
Conversely, careful onboarding and well-ordered internal processes work best for the people who would otherwise get lost. Tidiness is an investment in raising the organization’s “floor.” The meticulousness common in Japan can be rephrased as a mechanism for making everyone — across a range of abilities — modestly productive. An organization that runs while staying messy, by contrast, tacitly assumes that a certain number of people who can carve their own path are present.
Get Shibui Yusuke’s stories in your inbox
Join Medium for free to get updates from this writer.
And here too, AI changes the premise. AI that searches, summarizes, and guides is, in effect, “navigation against getting lost, handed out to everyone.” The “power to survive messiness” that until now only a few strong individuals possessed, AI distributes to ordinary people as well. As a result, the greatest weakness of an organization made messy by addition — that ordinary people get lost — is substantially offset. The weakness shrinks while the strengths, speed and optionality, remain intact; so this is a change that makes “just add for now” even more advantageous.
The same holds from the working individual’s side. There is no need to weigh whether a company is well-organized more heavily than it deserves. A person with real ability, AI in hand, can draw value for themselves out of a tidy company and a messy one alike. If so, then what one should really be looking at is not “whether the company is organized internally,” but “whether it is growing fast in the market.”
Meticulousness and Assets
Why, in the first place, do Japanese companies try to tidy up? At bottom, it is the work of “compressing an organization’s tacit knowledge into an explicit, transferable form.” Designing onboarding is nothing other than compressing person-dependent know-how into procedures anyone can follow. Companies skilled at this “power to write an organization’s OS” are, it seems to me, relatively common in Japan.
And that is precisely what today’s large language models, RAG, and AI agents accomplish cheaply. Here lies a paradox.
One side is the “threat.” “The ability of a human to navigate well-ordered internal processes” is a moat when humans move through an organization, but it is no moat against an AI that reads the internal wiki in an instant. Ironically, before AI raises Japanese companies’ competitiveness, it neutralizes the very advantage Japan has been good at — a “bureaucracy built for humans.” What you are best at becomes what is most easily replaced.
The other side is the “opportunity.” The chaos common at foreign-owned firms — outdated information, key-person dependence, scattered chat logs — is hard for an AI to traverse as well. And because it is written in each country’s language and follows each country’s conventions, the difficulty rises further for an AI (though it is hard for multilingual humans, too…). Well-ordered Japanese process documents, by contrast, become a high-quality basis for RAG, training data, and execution rules. Japan’s meticulousness, handed to AI, may turn into an advantage.
The key to resolving this contradiction is “who holds the meticulousness.” As long as humans keep holding it, it is a moat that is filling in. But hand it to AI and it becomes rocket fuel. If business processes stay locked in PDFs, internal wikis, briefing sessions, and individuals’ heads, they remain a “tidy bureaucracy for humans.” But if you can convert them into APIs, permissions, business rules, workflows, knowledge graphs, ontologies, and agent execution environments, they become a “high-quality operational platform for the AI era.”
The task is not to discard the processes. It is to “recompile” the processes into a form AI can use. Meticulousness is not a past to be thrown away but can become the most valuable asset of the future — though only once it leaves human hands.
Caution and Survival
The tenor of the argument so far has leaned toward “add fast.” But for the sake of fairness, I want to lay out the merits of the caution so common in Japan as well. This essay is not a “Japan is hopeless” argument.
AI organizational theory as discussed in Japan, left unchecked, tends to lean toward “inward-facing completeness.” Guidelines for generative-AI use, security rules, prompt training, internal Q&A bots, operational efficiency, an AI CoE (Center of Excellence), usage applications, risk management, talent development, governance — all of these matter. Indeed, in March 2026, the Ministry of Internal Affairs and Communications and the Ministry of Economy, Trade and Industry revised the “AI Guidelines for Business” to version 1.2, going so far as to furnish checklists and worksheets. A definition of autonomous AI agents was newly incorporated as well. This is, unmistakably, a distinctly Japanese strength.
The problem is when AI organizational theory shrinks down to nothing but “a management doctrine for using AI safely inside the company.” Here, the contrast drawn by the cultural psychologist Michele Gelfand becomes useful — between “tight cultures” (common in Japan), with strong norms and little tolerance for deviation, and “loose cultures” (common in the U.S.). The early stage of AI — or really the early stage of software technologies in general — is essentially an exercise of “loose culture.” Without tolerance for failure, permission to deviate, and departure from established procedure, you cannot learn. “Just add for now” is a loose move; “only after we’ve run compliance and security through the four operations” is a tight one.
That said, caution is not simply an evil. In today’s phase, where capability is scarce, governance is a shackle; but in a phase where capability has commoditized into parity, “trust” becomes the decisive differentiator. In regulated industries and in enterprise B2B sales, the moment the performance gap disappears, governance and accountability turn into a bulwark. The caution common in Japan and the EU may have the character of being “too early, but eventually right.” For that reason, the one and only great problem is whether you can survive until that phase arrives. It is the risk of “premature optimization” — optimizing for a future that has not yet come.
Let me add one piece of irony here, by way of self-reproach. The “AI-native theory” and “AI organizational theory” now in vogue in Japan preach that “merely adding AI — ‘adoption’ — is no more than utilization; you must fundamentally redesign organizational structure, evaluation systems, and decision-making.” It sounds very progressive. But look closely at the structure, and it often looks like the same perfectionism: “redesign everything cleanly before you move.” Has the very same “four arithmetic operations” that produced the low-growth equilibrium crept even into the prescription for the AI era? — this is worth examining in ourselves. Of course, there are companies that back their redesign with fast implementation and speed. In companies that first wring everything they can out of general-purpose AI tools and then remake the work itself based on what they learn, that redesign works powerfully. The trap springs when the “discussion about redesign” degenerates into “meetings about redesign,” bringing the addition itself to a halt.
Inward-Facing AI and Outward-Facing AI
AI use splits broadly into two kinds.
One is “inward-facing AI” — making internal processes more efficient. Meeting minutes, approval workflows, expense reimbursement, internal help desks, knowledge search, code generation, recruiting admin, contract review, call-center support — all of it matters, but as a competitive advantage in the market it is weak, because many companies do the same thing. You lose ground if you fall behind, yet doing it creates no great differentiation. What is more, its ceiling on value is set by “your own waste — that is, your own payroll × headcount.” In a Japan with a shrinking population, that ceiling keeps falling.
The other is “outward-facing AI” — AI you sell to customers. Services with AI built in, new experiences powered by AI, industry-specific automation, AI-native data products, new pricing models that presuppose AI — the risk is high, but the return when it succeeds is large, and the ceiling is the global market. Moreover, according to the AI Index mentioned earlier, while the share of organizations using generative AI somewhere in their operations has already reached 70%, full-scale deployment of AI agents remains in the single-digit percent range across nearly every function. The frontier of outward-facing AI is still wide open.
AI can search, summarize, guide, and execute over scattered documents and vague processes. In other words, AI substantially offsets the weakness of “an organization made messy by addition alone.” AI makes the habit of adding things sloppily even more tolerable. But AI teaches nothing to an organization that says “we mustn’t use it yet,” “let’s make the rules first,” “once the company-wide policy is set.” Learning belongs to humans and organizations, and the chance to learn arises only once something is used. Paradoxically, an organization that uses AI while somewhat messy learns, and an organization that does not use it, however tidy, does not.
And the greatest trap is to add AI while keeping all the existing processes in place. That is not “productivity improvement through AI” but merely “an organization whose work has grown by exactly the amount of AI added.” Adding AI alone is not enough.
A Half-Right Argument Is More Troublesome Than a Wholly Wrong One
The “AI-native theory” we saw in the previous section is half right. Organizational redesign is indeed necessary. But precisely because of that rightness, it tilts toward perfectionism and ends up halting the addition itself. A wholly wrong argument you can simply discard; a half-right argument is troublesome precisely because you cannot bring yourself to throw it away. So what is needed is not abstraction but concrete prescriptions. The realistic solution, I believe, is not to treat culture as monolithic but to split it by purpose. Let me offer a few specifics.
First, split management into two tracks. Inward-facing AI (internal DX) can be left at plus-or-minus zero. Devote it to control, tidying, and building the foundation — and then repurpose that very foundation directly as a basis for RAG and an execution environment for agents. Outward-facing AI (products), by contrast, you run as a separate unit with a deliberately separate culture: rough addition, leaving things to run wild, and selection. Permit “lucky if it sells, withdraw if it doesn’t.” For an organization good at tidying, this is a design that turns its tidying strength into outward-facing fuel.
Second, design for withdrawal. The real bottleneck is, in many cases, probably not “slowness to adopt” but “clumsiness at withdrawing.” The structure is not that addition stops, but that, being bad at subtraction (withdrawal), you cannot move on to the next addition. Set up the machinery for lowering the cost of withdrawal in advance — a decision-maker empowered to decide, clear termination conditions, a fair third-party evaluator — and addition starts to turn naturally.
Third, move your KPIs from the denominator to the numerator. Evaluate not by administrative productivity (the denominator) but by new value added and outward-facing revenue (the numerator). As the IPA survey cited earlier shows, no small number of companies do not even have metrics to measure outward outcomes. As long as the yardstick of evaluation points at the denominator, the organization will keep optimizing the denominator.
Fourth, delegate the tidying from people to AI. Free your high-salary talent from internal process design. The more the internal administration is put in order, the more the organization leans toward “intelligence for not conceding points.” That is not in itself bad, but much of the work of avoiding mistakes can now be automated by AI. Shouldn’t high human intelligence be directed instead toward framing the questions, building the business, understanding customers, and turning things into products — that is, toward the numerator side?
And fifth, change the direction of the question. Shift from “May we bring AI into this task?” to “What can we sell with AI, and how much can we run?” Until you change the grammar of the question, the ceiling will not change.
Friction and Learning
Japanese companies, on the whole, tend to be good at building “organizations that don’t make mistakes.” The four arithmetic operations are a technology “for not making mistakes.” Put differently, they seem poor at designing organizations that “make mistakes fast.” Diffusion, rephrased at the organizational level, is nothing other than “the right to make mistakes fast, cheaply, and in great quantity.” It is not the organization that has driven friction to zero, but the organization that can turn friction into outward-facing learning, that is strong.
The meticulousness so common in Japan is not an ability to be discarded. As an asset to hand to AI, it is as valuable as anything could be. The problem lies not in meticulousness itself, but in humans continuing to hold it. And if AI takes over the tidying and the wayfinding, each working individual need not fret over “whether they can survive in a messy environment,” and can redirect that energy toward outward-facing creation (though of course it is not as simple as it sounds on paper). Hand the tidying to AI, and let humans move to the “make mistakes fast” side — isn’t that exactly the shift now being demanded? Friction is not something to be consumed entirely on the inside, but something to be burned outward.
AI organizational theory is discussed from many angles, but to let it end as an internal-management doctrine for using AI safely would be a waste. We should think of it as a theory of design: how to convert an organization’s cultural quirks into “outward-facing learning” and “business value,” and how to let people survive within it. The quality common in Japan, the go-to-market common in the U.S., the implementation density common in China — the question is not which is right, but in what order you combine them and feed them into learning. That, I believe, is where the key lies to turning the “culture shock” one feels working across borders into competitiveness.
Key Sources
- Japan Productivity Center, “International Comparison of Labor Productivity 2025” (published December 2025; Japan’s 2024 labor productivity per hour worked was $60.1, 28th of 38 OECD member countries; lowest in the G7 on a per-worker basis as well).
- Stanford HAI, 2026 AI Index Report (published April 2026; U.S. private AI investment of $285.9 billion in 2025 was about 23× China’s $12.4 billion, with 1,953 newly funded AI companies; the top U.S.–China model performance gap narrowed from 17.5–31.6 points in May 2023 to about 2.7%; China’s AI patent filings are roughly 70% of the global total; generative AI is used in about 70% of organizations, while full-scale AI-agent deployment is in the single-digit percent range).
- International Federation of Robotics, “World Robotics 2025” (published September 2025; in 2024, annual industrial-robot installations were 295,000 units in China versus 34,200 in the U.S., about 9×; operating stock was about 2.03 million in China versus about 390,000 in the U.S., about 5×; the primary source behind the AI Index’s robotics figures).
- BCG, “AI at Work (2025)” and “AI at Work: Is Asia Pacific Leading the Way? (2025)” (about 70% of frontline employees in Asia-Pacific use AI regularly, versus about 51% globally; optimism about AI is highest in China, with Japan at the bottom of the region; the spread of “shadow AI”).
- Ministry of Internal Affairs and Communications & Ministry of Economy, Trade and Industry, “AI Guidelines for Business (ver. 1.2)” (published March 31, 2026; the second revision since the first edition of April 2024; furnishes checklists and worksheets and adds a definition of autonomous AI agents).
- IPA (Information-technology Promotion Agency), “DX Trends 2025” (a Japan–U.S.–Germany comparison; Japan’s DX tends toward “inward-looking, local optimization” while the U.S. and Germany tend toward “outward-looking, holistic optimization”; Japan’s gains are centered on cost reduction, the U.S. and Germany’s on revenue, profit, and market share; the share of companies setting metrics for outcomes is under 30% in Japan versus over 80% in the U.S. and Germany).
- CSIS, “It’s Time for AI Policy to Get Serious About the AI Adoption Gap” (December 2025; argues that the U.S. gained an early edge through massive investment in the best general-purpose models, but that this strategy bears fruit only if the broader economy generates enough “pull” — demand — to convert that supply into lasting impact; supports the diffusion thesis).
- RAND, “Reality Checking a Major National R&D Investment in AI Trustworthiness, Safety, and Security” (March 2026; argues that because broad adoption of advanced AI depends on trust that systems will perform reliably, investing in trustworthiness can accelerate diffusion — complementing innovation rather than competing with it; supports the claim that “trust becomes a differentiator”).
- China’s “AI+ (人工智能+)” Action Plan (published August 2025; sets adoption targets of 70% by 2027, 90% by 2030, and ubiquitous levels by 2035).
- DeepSeek, “R1” technical report (arXiv, January 2025; an open model that approached frontier performance with far fewer compute resources; became a symbol of the diffusion thesis).
- Michele J. Gelfand et al., “Differences Between Tight and Loose Cultures: A 33-Nation Study,” Science 332(6033), 2011 (the original source of the tight/loose-cultures framework; in the 33-nation study, Japan is classified as a strongly norm-governed “tight” culture and the U.S. as relatively “loose”; see also the general-audience book Rule Makers, Rule Breakers, 2018).
- Friedrich A. Hayek, “The Use of Knowledge in Society,” American Economic Review 35(4), 1945 (a classic on dispersed knowledge and spontaneous order; the theoretical grounding for “evolution = market selection with no central designer, versus intelligent design”).
* Figures are as of each source’s date of publication. Citations in the text are summaries and differ from the original wording.
* The views expressed here are the author’s own and do not represent any organization. Apologies if any of the figures or data are wrong.
