1. What deploying AI in an enterprise means (the short version)

• "Using AI" covers a wide range. It runs from a general-purpose chat tool for drafting emails, to a system that answers from internal documents, to a system that hands part of a fixed procedure to AI, to AI built into a product customers see. Each needs different data, a different unit of evaluation, and a different weight of responsibility Sourced [Ref. 9, 14].
• A trial and daily operation are different problems. Daikin Industries announced a detection accuracy of 91% in an experiment at a training facility and 76% in a verification on actual field work, for an AI agent that checks work for omissions (company-announced values) Sourced [Ref. 11]. The same system gives different numbers once the data comes closer to the field.
• Benefit depends on how it is measured. In a field deployment in customer support, productivity rose by 15% on average, with a small drop in quality among the most experienced agents; in a randomized trial with experienced developers, completion time actually increased by 19% Sourced [Ref. 15, 17]. "It feels more convenient" and a measured result can disagree.

2. Where adoption stands: what public statistics show

Public statistics are the starting point for how widely AI is used. Surveys differ in who is covered (size and sector), how the question is worded and when it was run, so this article does not line numbers up to rank countries. We read each statistic for what it actually measures.

EU: Eurostat (2025)

According to Eurostat, 19.95% of EU enterprises with 10 or more employees used at least one AI technology in 2025. The list covers text mining, speech recognition, generation of written or spoken language or code, generation of images or audio, image recognition, machine learning for data analysis, automation of workflows or decision support, and machines that move autonomously. The figure is 6.47 percentage points higher than in 2024. By size, 17.00% of small, 30.36% of medium and 55.03% of large enterprises used AI; by country, Denmark was highest at 42.03% and Romania lowest at 5.21% Sourced [Ref. 1]. By sector, information and communication (62.52%) and professional, scientific and technical services (40.43%) stood out; every other sector was below 25% [Ref. 1].

Among enterprises that used AI in 2025, the most common purposes were marketing or sales (34.70%) and organising business administration processes or management (31.05%); logistics (6.08%) was the least common. In manufacturing, too, the leading purposes were marketing or sales (30.41%) and management (27.05%) [Ref. 1]. Among enterprises that had considered AI but did not use it, the most common reasons were lack of relevant expertise (70.89%), lack of clarity about legal consequences (52.52%) and concerns about data protection and privacy (48.83%) [Ref. 1].

United States: Census Bureau BTOS (December 2025 to May 2026)

The US Business Trends and Outlook Survey (BTOS) is a biweekly national survey. In data for 14 December 2025 to 3 May 2026, the share of businesses reporting AI use stayed in a 17-20% range, with a national value of 19.8% on 3 May 2026; 20-23% expected to use AI within six months. Among firms with 250 or more employees, 37% reported using AI; for firms with 100-249 employees it was 32% (period ending 3 May); and for firms with four or fewer employees it was under 20% Sourced [Ref. 2]. Caution: the Census Bureau explains that in November 2025 it changed the question from AI use "in producing goods or services" to AI use "in any business function" [Ref. 2]. Numbers before and after that change should not be read as one continuous time series.

Japan: MIC's 2026 White Paper (FY2025 corporate survey)

In the FY2025 corporate questionnaire run by the Ministry of Internal Affairs and Communications (MIC) in Japan, the United States, Germany and China, 68.9% of Japanese companies (49.7% in the previous year's survey) said their policy is to use generative AI actively or to use it in limited areas (74.0% of large companies, 58.1% of small and medium-sized ones). The white paper notes that Japan remains lower than the other three countries [Ref. 3]. Separately, excluding "don't know" answers to the policy question, 86.4% of Japanese companies said they use generative AI in at least one business function, up sharply from the previous survey and narrowing the gap with the other three countries [Ref. 4]. The two percentages have different denominators. The business types where Japanese companies use generative AI most are drafting minutes and emails (about 70%), sales (about 60%) and internal help desks (about 50%); about 70% also reported a sense of benefit for minutes and email drafting [Ref. 4].

For the expected effects of generative AI, the white paper reports that in Japan "greater efficiency through shorter work time" was highest at 61.6%, while in the other three countries "better products and services, for example through more accurate quality control or personalization" was highest [Ref. 5]. The white paper also offers the observation that in the other three countries companies are using generative AI in customer-facing products and services, and that pre-deployment review and post-deployment evaluation structures may be developing accordingly [Ref. 6]. That is the white paper's reading, not proof of cause.

3. Patterns of deployment: from personal tools to AI in products

The MIC white paper describes the staged approach of advanced companies as moving from "improving individual work with general-purpose AI" to "redesigning whole business processes on the premise of AI" [Ref. 9]. On the developer side, Anthropic distinguishes workflows, where large language models (LLMs) and tools are orchestrated through predefined code paths, from agents, where the LLM dynamically directs its own process and tool use, and recommends finding the simplest solution possible and adding complexity only when needed (the developer's own view) [Ref. 14]. The AI Guidelines for Business (version 1.2) define an AI agent as "an AI system that senses its environment and acts autonomously to achieve a specific goal" [Ref. 34]. The figure below is this article's own arrangement based on those sources. The division into patterns is ours and may not match how individual companies label their systems.

View original SVG, values and source annotations

All entries are statements in published sources (Sourced). The pattern arrangement and the caution column are our reading.

4. How a PoC differs from operation at scale

A PoC (proof of concept) asks "can it technically be done?"; operation asks "will it keep running in daily work, at a set quality, under someone's responsibility?" (our reading). Three examples from official sources show the difference.

All entries are company announcements or white paper statements (Sourced); performance figures are company claims. Where the definition is not published, we do not compare them with other examples.

5. Data preparation — what "readable by AI" means

The RAND report names five root causes of failed AI projects, and two of them concern data: the organization lacks the data needed to train an effective model, and it lacks the infrastructure to manage its data and deploy completed models Sourced [Ref. 13]. The MIC white paper likewise quotes Professor Sato of the National Institute of Informatics (Japan) as saying that the first step to successful AI deployment is whether the data is ready for AI ("AI-Ready"), and that for AI agents in particular the accuracy and freshness of data matter more Sourced [Ref. 36]. In the same white paper's survey, more than half of companies in the US, Germany and China answered that they train generative AI on internal data or have built databases that generative AI can consult; the share in Japan was much lower [Ref. 37].

The international standard series for data quality is ISO/IEC 5259 (data quality for analytics and machine learning); Part 1, ISO/IEC 5259-1:2024, covers an overview, terminology and examples. ISO's pages also list a package of the five parts (ISO/IEC 5259-1 to -5) [Ref. 24, 30]. On the management of scientific data, the FAIR principles (Findable, Accessible, Interoperable, Reusable) were proposed in Scientific Data in 2016 [Ref. 23]. These are recommendations and guidance; compliance is not legally required (our reading). The standards' full texts are paid and were not read, so their detailed requirements are not described here.

The items and examples are our reading, not requirements listed by the standards or the white paper.

6. Evaluation — measuring on three levels

It helps to split the evaluation of AI into three levels according to the question asked (our arrangement): 1. the general capability of the model, 2. performance on your own tasks, and 3. business outcomes. A high result at level 1 can coexist with a low one at level 2, and a high result at level 2 may not translate into the outcome at level 3.

All entries are Sourced (as reported by each source). The level arrangement is our reading. We do not rank numbers from different settings, versions or evaluators.

Keep your own evaluation set

Public benchmarks are a guide for comparing models; they are not quality assurance for your operations. In the Daikin Industries and Hitachi example, the evaluation cases were divided by equipment type (5 types by 5 cases) and rated by experienced engineers [Ref. 10]. Points for building your own evaluation set are as follows (our reading).

• Collect from real questions and cases: build from past inquiries and defect reports, not only from imagined questions.
• Fix the criterion for correctness first: who scores, against what standard. If several people score, check how well they agree.
• Include hard cases on purpose: exceptions, old versions, contradictory documents and questions with no answer. Also evaluate whether the system can say "there is no answer".
• Rerun at every change: when the model version, retrieval settings or documents change, confirm no regression on the same set.

7. Measuring cost and benefit — our calculation

The cost of AI is not only usage fees (token charges or licences). The main components are usage fees, data preparation, labour for evaluation and verification, platform operation and maintenance, training, governance and review, and handling changes to models and documents (our reading). On the benefit side, some companies, such as Chugai Pharmaceutical, have a structure in which the digital division works alongside the front line and handles effect measurement [Ref. 36]. Below, hypothetical numbers show how costs break down and how to think about break-even.

Basics of measuring benefit

• Measure outcomes, not usage: logins and question counts show only that the tool is used. Choose indicators you can measure from before deployment, such as rework counts, lead time, missed inspection findings or time to resolve inquiries (our reading).
• Provide a comparison: the QJE study used the staggered rollout of a generative AI assistant to compare before and after, and adopters with non-adopters [Ref. 15]. If everyone gets the tool at once, its effect cannot be separated from seasons or other initiatives (our reading).
• Measure quality as well: looking only at speed misses quality declines. Even in the QJE study, quality dropped slightly among the most experienced agents [Ref. 15].
• Assume models will be updated: Chugai Pharmaceutical explains that AI technology evolves quickly and that a new technology can give similar output at lower cost, so it provides an in-house environment (an AI platform) where the latest tools from several vendors are always available [Ref. 7]. The cost assumptions are not fixed.

8. Workflow automation — what happens as steps are added

When tasks are chained for automation, the more steps there are, the lower the overall success rate. If each step succeeds independently with probability p, the probability that all n steps succeed is p to the power n.

Anthropic summarises that workflows offer predictability and consistency for well-defined tasks, while agents are the better option when flexibility and model-driven decision-making are needed at scale (the developer's view) [Ref. 14]. Chugai Pharmaceutical says that in redesigning business processes it is important to decide which steps to leave to AI and where humans must take responsibility, and to keep humans in the loop for judgment and confirmation [Ref. 7]. The design points are as follows (our reading).

• Put human checks before the steps where an error is costly: checking every step defeats automation; checking none lets errors accumulate.
• Record the output of each step: keeping intermediate outputs, not just the final result, shows which step broke.
• Make actions reversible: allow executed operations to be undone, or put an approval step before execution.
• Decide what AI is not given: the AI Guidelines for Business note a proposal that, when a person approves an AI evaluation or judgment, the person should first think independently about the reasons and grounds for approving [Ref. 34].

9. Why projects fail

A source close to a public analysis of why AI projects fail is the RAND report RR-A2680-1, published in August 2024. Based on interviews with 65 experienced data scientists and engineers (at least five years of experience building AI/ML models, in industry and academia), it summarises five root causes Sourced [Ref. 13]. The report cites the estimate that "more than 80 percent of AI projects fail" as an estimate "by some estimates"; it is not a measurement made by the report itself [Ref. 13]. We therefore do not use that number as a statement of the failure rate.

One of the report's five principles, to choose problems that are worth sustained effort (the report asks that each team commit to a specific problem for at least a year), does not map one-to-one onto the five causes above, so it is not in the table [Ref. 13]. The right-hand column is our reading.

Public statistics and the white paper show barriers from other angles. In the EU, among enterprises that had considered AI but did not use it, the most common reasons were lack of expertise (70.89%), lack of clarity about legal consequences (52.52%) and data protection and privacy concerns (48.83%) [Ref. 1]. The risk Japanese companies are most concerned about is "security risks such as leakage of internal information", followed by "problems with the accuracy of outputs" and "the possibility of infringing copyright and other rights" [Ref. 6]. In the white paper's survey, 27.0% of Japanese companies answered that they had "no organized effort" on business transformation with generative AI, notably higher than in the US, Germany and China, with a stronger tendency among small and medium-sized companies [Ref. 37].

10. Governance structures

As AI use spreads from personal tools to business processes and products, an organization needs a structure that says who decides what, who checks what, and who explains when something goes wrong. Japan's AI Guidelines for Business (version 1.2) were published by the Ministry of Internal Affairs and Communications and the Ministry of Economy, Trade and Industry (METI) on 31 March 2026, and are a non-binding soft law (guidance with no legal force) Sourced [Ref. 34, 35]. The versions were published as 1.0 (19 April 2024), 1.01 (22 November 2024), 1.1 (28 March 2025) and 1.2 (31 March 2026) [Ref. 35]. The guidelines distinguish three actors (AI developers, AI providers and AI users), and set out common guiding principles and the building of AI governance. For governance they say that, rather than fixing rules and procedures in advance, it is important to practise agile governance: multiple stakeholders continuously and rapidly run a cycle of environment and risk analysis, goal setting, system design, operation and evaluation [Ref. 34].

Among the leading companies the white paper interviewed, many started from the needs of business units on the front line, with the AI promotion (digital transformation) unit supporting them with idea generation, technical advice, security measures, effect measurement and learning opportunities. The same unit also built shared platforms with an eye to overall optimisation and drove cross-unit change [Ref. 36]. On rules, the companies were developing usage policies and guidelines and evaluation schemes, limiting the range of data AI may consult through access permissions, and beginning to consider review structures at the planning and design stage [Ref. 36]. In the Japanese survey, the most common risk measure was "company-wide guidelines are in place" (41.1%), while in the US, Germany and China, expert review before introduction and periodic evaluation afterwards were more common than in Japan [Ref. 6].

Where standards and guidance fit

Titles, editions and dates were checked on each body's official page (Sourced; checked 15 October 2026). The standards' full texts are paid and were not read, so their detailed requirements are not described. The use column is our reading.

11. Cautions for manufacturing and R&D

Manufacturing sites and R&D differ from general office work. We organise what has been covered so far from the standpoint of a materials engineer (our reading).

AI-driven materials discovery and self-driving laboratories are covered in our explainers on materials informatics. For AI drug discovery, see our explainer on AI drug discovery and screening; for the power and cooling of AI data centers, see our explainers on AI data center power and cooling and on data centers and energy storage.

12. Limits, and what we could not confirm

(1) International comparison and definitions

The surveys by Eurostat, the Census Bureau and MIC cover different populations and use different questions, so we do not set adoption rates side by side to compare them. Japan's 86.4% uses answers excluding "don't know" as its denominator, which differs from the denominator of 68.9%.

(2) Companies' performance figures

The figures for Daikin Industries, Hitachi, Fairy Devices, Chugai Pharmaceutical and Nissin Foods Holdings are company announcements, or figures the white paper reports from material supplied by the companies. For some of them, the definition of detection accuracy or usage rate, the population and the measurement period could not be confirmed from the published material. We could not confirm completion of the pilot run, or an actual record of practical use and full-scale rollout.

(3) Actual cost-effectiveness

We could not confirm, within the scope of this survey, a primary source that gives the cost-effectiveness of an AI deployment as the actual record of a real company. The numbers in Chapter 7 are assumptions, not prices of any real product.

(4) Standards texts and the application of regulation

The full texts of ISO/IEC 42001, 23894, 38507, 42005, 42006 and 5259-1 are paid and were not read. We do not describe detailed requirements or the status of certification. Whether EU, US, Chinese or Japanese law applies is not covered.

(5) Current models and benchmark values

The τ-bench and SWE-bench numbers are from the original papers and we have not checked the level of current models. Evaluations such as GDPval are run by the developer, and we have not confirmed independent third-party replication.

13. How to read evaluations and standards

Enterprise AI deployment is assessed on four fronts: the reality of adoption (statistics), model capability (benchmarks), business outcomes (field effect studies) and structure and risk (standards and guidance). A benchmark asks how much a model can do, a field study asks what changed through deployment, and a standard or guideline asks how an organization manages it. These are separate questions. The table below lines up the evaluation methods and the standards or guidance for each, as far as sources could be confirmed. Where nothing could be confirmed, the cell says so.

Titles, editions and dates were checked on each body's official page, paper or official document (Sourced; checked 15 October 2026). The arrangement of this table is Our calculation (a reordering by this article), and the reading-note column is commentary.

Notes on reading

• Benchmark values depend on settings and on who ran them. Even under the same benchmark name, values cannot be compared if the version, tool use, number of attempts or scoring method differ. Evaluations run by a developer are distinguished from independent third-party evaluations. Public benchmarks can saturate (approach the ceiling so differences vanish) or be contaminated by training data.
• Check whether a standard or guideline is a recommendation or a requirement, and whether it has legal force. ISO/IEC 42001 is a set of requirements an organization adopts voluntarily, ISO/IEC 23894 and 42005 are guidance, the NIST AI RMF is a voluntary framework, and Japan's AI Guidelines for Business are non-binding soft law. These are separate from legislation (for example, the EU AI regulation), and conforming to a standard does not necessarily mean complying with a law (our reading).
• Distinguish stated values from measured ones, and trials from the field. An "accuracy" or "reduction rate" a company announces is the company's claim, and it can be evaluated only when the definition (what counts as correct), the number of cases and the measurement period are shown. As in the announcement of 91% at a training facility and 76% in the field, values change when conditions change [Ref. 11].

14. Glossary

15. References

1. Eurostat "Use of artificial intelligence in enterprises", Statistics Explained (2025 data, extracted December 2025; next update planned for December 2026) — ec.europa.eu/eurostat/statistics-explained/index.php?title=Use_of_artificial_intelligence_in_enterprises
2. U.S. Census Bureau "Large Firms With at Least 20 Employees Biggest AI Users" (America Counts: Stories, 26 May 2026; Business Trends and Outlook Survey data from 14 December 2025 to 3 May 2026) — www.census.gov/library/stories/2026/05/ai-use-businesses.html
3. Ministry of Internal Affairs and Communications (MIC), Japan MIC, 2026 White Paper on Information and Communications in Japan (in Japanese), Part I, Chapter 2, Section 1: (1) Policies on using generative AI (in Japanese) — www.soumu.go.jp/johotsusintokei/whitepaper/ja/r08/html/nd121110.html
4. MIC, Japan Same: (3) Business use of generative AI (in Japanese) — www.soumu.go.jp/johotsusintokei/whitepaper/ja/r08/html/nd121130.html
5. MIC, Japan Same: (5) Effects that generative AI may bring (in Japanese) — www.soumu.go.jp/johotsusintokei/whitepaper/ja/r08/html/nd121150.html
6. MIC, Japan Same: (4) Risks of concern and measures against them (in Japanese) — www.soumu.go.jp/johotsusintokei/whitepaper/ja/r08/html/nd121140.html
7. MIC, Japan Same: 3 (1) Case study of Chugai Pharmaceutical (from individual interviews; in Japanese) — www.soumu.go.jp/johotsusintokei/whitepaper/ja/r08/html/nd121310.html
8. MIC, Japan Same: 2 (3) Manufacturing and production (includes the Daikin Industries case; in Japanese) — www.soumu.go.jp/johotsusintokei/whitepaper/ja/r08/html/nd121230.html
9. MIC, Japan Same: 4 (2) Steps toward creating effects from AI deployment (in Japanese) — www.soumu.go.jp/johotsusintokei/whitepaper/ja/r08/html/nd121420.html
10. Daikin Industries and Hitachi "Daikin and Hitachi co-create and begin a pilot run toward practical use of an AI agent that supports equipment fault diagnosis in factories" (22 April 2025; in Japanese) — www.hitachi.com/ja-jp/press/articles/2025/04/0422/
11. Daikin Industries and Fairy Devices "AI agent developed by Daikin and Fairy Devices wins first place and the AI Agent Award at the GENIAC-PRIZE organised by METI and NEDO" (25 March 2026; in Japanese) — www.daikin.co.jp/press/2026/20260325
12. Nissin Foods Holdings "Initiatives on the use of generative AI" (thematic IR material for the fiscal year ended March 2024, 14 March 2024; in Japanese) — media.www.nissin.com/jp/company/ir/library/event/pdf/20240314_2.pdf
13. Ryseff J., De Bruhl B. F., Newberry S. J. (RAND) "The Root Causes of Failure for Artificial Intelligence Projects and How They Can Succeed: Avoiding the Anti-Patterns of AI", RAND Research Report RR-A2680-1 (13 August 2024) — www.rand.org/pubs/research_reports/RRA2680-1.html
14. Anthropic "Building effective agents" (19 December 2024; a developer's technical article. The page notes that the tooling landscape it describes has changed since December 2024) — www.anthropic.com/engineering/building-effective-agents
15. Brynjolfsson E., Li D., Raymond L. "Generative AI at Work", Q. J. Econ. 140(2), 889–942 (2025), DOI 10.1093/qje/qjae044 — doi.org/10.1093/qje/qjae044
16. Dell’Acqua F. et al. "Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of Artificial Intelligence on Knowledge Worker Productivity and Quality", Organization Science 37(2), 403–423 (2026), DOI 10.1287/orsc.2025.21838 (published version of HBS Working Paper 24-013, 2023) — doi.org/10.1287/orsc.2025.21838
17. Becker J., Rush N., Barnes E. et al. (METR) "Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity", arXiv:2507.09089 (v2) (first posted 12 July 2025) — arxiv.org/abs/2507.09089
18. Patwardhan T. et al. (OpenAI) "GDPval: Evaluating AI Model Performance on Real-World Economically Valuable Tasks", arXiv:2510.04374 (v1) — arxiv.org/abs/2510.04374
19. Liang P. et al. "Holistic Evaluation of Language Models", arXiv:2211.09110 (v2), Transactions on Machine Learning Research (2023) — arxiv.org/abs/2211.09110
20. Yao S. et al. "τ-bench: A Benchmark for Tool-Agent-User Interaction in Real-World Domains", arXiv:2406.12045 (v1) — arxiv.org/abs/2406.12045
21. Jimenez C. E. et al. "SWE-bench: Can Language Models Resolve Real-World GitHub Issues?", arXiv:2310.06770 (v3) — arxiv.org/abs/2310.06770
22. Mirza A. et al. "A framework for evaluating the chemical knowledge and reasoning abilities of large language models against the expertise of chemists", Nat. Chem. 17, 1027–1034 (2025), DOI 10.1038/s41557-025-01815-x — doi.org/10.1038/s41557-025-01815-x
23. Wilkinson M. D. et al. "The FAIR Guiding Principles for scientific data management and stewardship", Sci. Data 3, 160018 (2016), DOI 10.1038/sdata.2016.18 — doi.org/10.1038/sdata.2016.18
24. ISO/IEC "ISO/IEC 5259-1:2024" (Artificial intelligence — Data quality for analytics and machine learning (ML) — Part 1: Overview, terminology, and examples; edition 1) — www.iso.org/standard/81088.html
25. ISO/IEC "ISO/IEC 42001:2023" (Information technology — Artificial intelligence — Management system; edition 1, December 2023, JTC 1/SC 42) — www.iso.org/standard/81230.html
26. ISO/IEC "ISO/IEC 42006:2025" (Requirements for bodies providing audit and certification of artificial intelligence management systems; edition 1, July 2025) — www.iso.org/standard/42006
27. ISO/IEC "ISO/IEC 23894:2023" (Information technology — Artificial intelligence — Guidance on risk management; edition 1, February 2023) — www.iso.org/standard/77304.html
28. ISO/IEC "ISO/IEC 38507:2022" (Governance of IT — Governance implications of the use of artificial intelligence by organizations; edition 1) — www.iso.org/standard/56641.html
29. ISO/IEC "ISO/IEC 42005:2025" (Information technology — Artificial intelligence (AI) — AI system impact assessment; edition 1) — www.iso.org/standard/42005
30. ISO/IEC "ISO/IEC 22989:2022" (Artificial intelligence concepts and terminology; edition 1, July 2022. The page also lists the five-part ISO/IEC 5259 series as an "AI data quality management package") — www.iso.org/standard/74296.html
31. National Institute of Standards and Technology (NIST) "Artificial Intelligence Risk Management Framework (AI RMF 1.0)", NIST AI 100-1 (January 2023), DOI 10.6028/NIST.AI.100-1 — doi.org/10.6028/NIST.AI.100-1
32. NIST "Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile", NIST AI 600-1 (July 2024), DOI 10.6028/NIST.AI.600-1 — doi.org/10.6028/NIST.AI.600-1
33. NIST "AI Risk Management Framework" (NIST overview page. As checked on 15 October 2026, it states that AI RMF 1.0 is being revised as part of the White House AI Action Plan) — www.nist.gov/itl/ai-risk-management-framework
34. MIC and Ministry of Economy, Trade and Industry (METI), Japan "AI Guidelines for Business, Version 1.2" (31 March 2026; in Japanese) — www.meti.go.jp/shingikai/mono_info_service/ai_shakai_jisso/pdf/20260331_1.pdf
35. METI, Japan "Study Group on AI Guidelines for Business" page (list of publication dates for versions 1.0 to 1.2; last updated 24 August 2026; in Japanese) — www.meti.go.jp/shingikai/mono_info_service/ai_shakai_jisso/index.html
36. MIC, Japan MIC, 2026 White Paper on Information and Communications in Japan (in Japanese), Part I, Chapter 2, Section 1: 4 (1) Elements that matter for effective AI deployment and use (in Japanese) — www.soumu.go.jp/johotsusintokei/whitepaper/ja/r08/html/nd121410.html
37. MIC, Japan Same: (2) Organizational efforts (in Japanese) — www.soumu.go.jp/johotsusintokei/whitepaper/ja/r08/html/nd121120.html

16. Claim-to-source audit

Last updated: 15 October 2026 / Sources are limited to primary material (official pages of Eurostat, the US Census Bureau, MIC and METI; a RAND report; peer-reviewed papers; arXiv papers; official pages of standards bodies and NIST; and official company announcements). The full texts of standards are paid and were not read. The arrangement of deployment patterns, the cost assumptions and the readings of how to deploy are marked as commentary or our calculation to keep them apart from sourced facts. Companies' performance figures (detection accuracy, number of users and so on) are company-announced values; where the definition is not published we could not confirm it and do not compare them. We could not confirm the completion of pilot runs or records of practical use, actual cost-effectiveness, detailed requirements of the standards, or whether any national law applies, so they are not described. This article is not legal or investment advice and does not recommend deploying any particular product or company's offering. Figures 2 to 5 and 7 are vector drawings; Figures 1 and 6 are frames for AI-generated images (not yet created). None of them depicts a real factory, company, product or measured data.