Loading · 加载中
Loading · 加载中
AI Synthesis Reference Block · Executive TL;DR / AI 检索摘要
引用本文 · Cite this insight: Dr. Tong Yin (2026-08-31). Global AI in the Real Economy: From Adoption Hype to Industrial Value and Physical AI / 《实体经济中的全球 AI：从采用热潮到工业价值与物理 AI》. InsightBridge Global Intelligence. https://intelligence.insightbridge.global/articles/global-ai-in-the-real-economy-from-adoption-hype-to-industrial-value-and-physica — Series: technology
白皮书 · 2026 年 8 月 · 数据截至 2026 年 8 月 15 日 · InsightBridge Global LLC · 作者：Dr. Tong Yin · 覆盖 15 个国家与地区、12 个主要行业，含 31 张表、8 幅图、285 条来源（正文为英文原版）
Copyright and method InsightBridge Global LLC
How this white paper was assembled and how it should be quoted
1. Publisher and attribution This white paper was prepared and published by InsightBridge Global LLC and is attributed to InsightBridge Global LLC / Dr. Tong Yin. It was prepared in August 2026 and draws exclusively on material publicly available as at 15 August 2026.
2. Six evidence rules
— Single evidence base. Only figures verified in the underlying research file are used. No new numbers are introduced and no market-size forecasts are made.
— Unverifiable means n.a. Where no verifiable official figure could be obtained, the field reads n.a. Consumer data, telemetry estimates and industry guesses are never substituted.
— Never merge different bases. Enterprise adoption rates, population-level diffusion rates and industry surveys are three non-interchangeable measurement systems. They are presented in separate chapters and are never summed, divided or ranked together.
— Sources on every page. Numbered superscripts in the body text correspond to full, clickable URLs at the foot of the same page, so any figure can be checked page by page.
— Conflicts of interest are labelled. Vendor and vendor-commissioned research (NVIDIA, Thomson Reuters, Itron, Trimble, IBM–NRF and others) is flagged inside the tables and must not be read as an industry average.
— Self-assessment is distinguished from audit. All financial and productivity outcomes based on respondent judgement are labelled as self-reported and are not treated as audited financial data.
3. Comparability grades (A / B / C)
Table 1 Definition of comparability grades
| Grade | Meaning | Applies in this white paper to |
|---|---|---|
| A | Follows the Eurostat ICT enterprise survey framework (enterprises with 10 or more employees, NACE Rev.2 C–J, L–N and group 95.1, an itemised technology 1 list) and is therefore directly comparable | Germany, France, Denmark, EU-27 |
| B | Official national statistics whose size threshold, recall period, sector coverage or technology list differs from grade A; usable for orders of magnitude and trends, not for ranking | United States, United Kingdom, Canada, Korea, Singapore, Brazil |
| C | Authoritative institutional surveys, industry-body surveys or telemetry estimates with self-selected samples or a non-enterprise base; must not be ranked alongside official statistics | China (enterprise adoption), Japan, India, United Arab Emirates, Saudi Arabia |
4. Copyright and permitted use © 2026 InsightBridge Global LLC. All rights reserved. Tables and charts may be quoted provided the source, the definitional caveats and the disclaimer are retained in full. Individual figures must not be extracted for cross-country rankings or investment marketing. This white paper is not investment, legal or regulatory advice; the full disclaimer appears on the back cover.
Diffusion is fast; organisational value capture is not
Today's debate about AI rests on a category error: personal exposure, employee usage, corporate procurement,
production deployment and auditable financial return are all described as "adoption". This white paper
separates them into six measurable layers, covers 15 countries and territories and 12 major industries, and
treats the spread of software tools separately from industrial-grade, hardware-integrated physical AI.
On an official statistical basis (the enterprise as the unit of observation, 10 or more employees, an itemised list of AI
technologies), 20.0% of EU-27 enterprises used AI in 20252. In the same period, an executive sample reported that 88% of
organisations use AI regularly in at least one business function3. The fourfold gap is not a contradiction but a difference of
base and definition: the first is a census-style probability sample, the second a self-selected executive panel in which the
term "adoption" is explicitly "left undefined"4.
In the same 2025 McKinsey survey, only 23% of organisations had scaled agentic systems anywhere in the enterprise,
defined as "expanded deployment and adoption within at least one business function"3. BCG's global study finds 5% of
companies are "future-built", 35% are scaling and 60% obtain "almost no material value"5. The Stanford AI Index 2026,
citing McKinsey, reports that 36% of respondents see improved profitability and 33% improved organic revenue growth6.
Denmark: 75% of enterprises with 250 or more employees versus 37% of those with 10–49 (2025)7. France: 58% versus 15%
(2025)8. Germany: 57% versus 23% (2025)9. Korea: 66.9% versus 30.6% (2024)10. Singapore: non-SMEs 62.5% versus SMEs
14.5% (2024)11.
Strategy has moved to physical AI; enterprise penetration has not
On 16 March 2026 NVIDIA's founder Jensen Huang stated that "physical AI has arrived — every industrial company will
become a robotics company". FANUC, ABB Robotics, YASKAWA and KUKA (with more than two million robots installed
worldwide) are integrating Omniverse libraries and the Isaac simulation framework into virtual commissioning, and are
embedding Jetson modules in controllers for real-time inference at the edge12. Foxconn (Fii) used Omniverse digital twins
to cut "time to market, factory build and factory planning by roughly 50%" and PhysicsNeMo to accelerate CFD thermal
simulation by 150 times, from hours to minutes13.
Among German enterprises that use AI, only 6% use technologies for autonomous machine movement14. Denmark
recorded 2% in both 2023 and 2024 — the only AI category with no growth15. Brazil recorded 7% in 202516. The strategic
narrative and enterprise-level penetration must be presented together, or the reader will draw opposite conclusions.
CAICT reports that across 11 Chinese MaaS platforms the average call success rate for the DeepSeek-R1 service rose from
87.01% in February 2025 to 99.36% in September, output tokens per second rose from 17.86 to 26.76, and time to first
token fell from 3.07 seconds to 1.02 seconds17. This is the clearest official evidence in this study of industrial-grade
reliability metrics replacing benchmark scores.
"Personal exposure and tool-level usage are close to mainstream. Organisational workflow redesign and auditable
financial returns remain a minority achievement. The gap does not lie in model capability but in depth of
deployment, the data and governance foundation, and process change that can be booked in the profit and loss
account."
The central conclusion of this white paper; the supporting evidence is set out in chapter 4.
Three rules for readers First, never place 88% and 20% on the same chart. Second, never use population-level diffusion rates to rank countries for policy
purposes. Third, treat outcome figures from vendor surveys as sample experience, not industry averages. All three caveats are set
out in full in chapter 9, Data limitations and rules of use.
Each layer must always be quoted together with its base, or errors of an order of magnitude follow
Table 2 The six-layer framework: definitions, bases and risks of misreading
| Layer | Name | Definition | Base | Representative data | Risk of misreading |
|---|---|---|---|---|---|
| L1 | Exposure / individual use | An individual has used a generative AI product in a given period | Population, or working-age population (15–64) | Microsoft AI Economy Institute basis: roughly one in six people globally in H2 2025; 24.7% of the working-age population in the 18 global north, 14.1% in the global south | Never to be read as enterprise adoption. Built from aggregated anonymised telemetry adjusted for operating-system share, internet penetration and population; not a survey |
| L2 | Enterprise adoption | The enterprise answers whether it uses at least one AI technology from a list | All enterprises (typically a threshold of 10 or more employees) | EU-27 20.0% in 202519; US BTOS 19.8% for the collection period 20 ending 3 May 2026 | Different technology lists make figures incomparable; changes in question wording create breaks in series |
| L3 | Piloting | The share of organisations in experimentation or pilots | Surveyed organisations | India: 23% at the pilot stage21; commercial real estate: 88% of 22 investors and 92% of occupiers have launched AI pilots | A high pilot count is not capability; JLL notes implementation is broad but "most initiatives remain experimental with limited scaling" |
| L4 | Production deployment and scaling | Entry into real production processes or enterprise-wide rollout | Surveyed organisations or use cases | India: 47% have multiple use cases live in production21; 23; agents: construction: only 1% have scaled AI across projects23% 24 scaled | "Production" has no common definition; vendor surveys systematically overstate it |
| L5 | Financial value conversion | The share reporting revenue, cost, EBIT, ROI or productivity improvement | Organisations using AI | Cost 38%, profitability 36%, organic revenue growth 33%, market share 25%25; Denmark: seven in ten AI-using firms report more 26 efficient workflows | All figures are self-assessed perceptions, not audited financials; the same chart notes that as many respondents report improvement as report no impact |
| L6 | Physical deployment | AI embedded in robots, edge devices, industrial control and smart terminals | All enterprises, or manufacturing employment | Germany: only 6% of AI-using enterprises use autonomous machine movement27; Denmark 28; global robot 2% for two consecutive years 29 density 132 units per 10,000 employees in 2024 | Robot density measures the automation stock, not AI; the IFR describes it as a common basis for relating robot numbers to the size of an economy measured by its workforce |
Chapter 1 continued — Three-stage maturity and six methodological warnings The model to use consistently, and the caveats that must travel with it
1.2 The three-stage maturity model
Table 3 Three-stage maturity model
| Stage | Content | Layers | First-hand evidence |
|---|---|---|---|
| Stage 1 | Employees use general-purpose assistants | L1 and part of L2 | In BCG's five-level model of employee adoption, more than 85% 30 |
| Tool access and | individually; no process redesign | of employees remain at levels two and three | |
| experimentation | |||
| Stage 2 | AI enters existing business processes and | L3 → L4 → L5 | More than two-thirds of organisations use AI in more than one |
| Workflow integration and | spreads across functions | function and half in three or more31 | |
| scaling | |||
| Stage 3 | Determinism, latency, safety certification | L6 | NVIDIA's own term is "production-scale physical AI", delivered |
| Industrial / physical AI | and edge compute become the gating factors | through "an open, integrated platform to design, train, test and deploy"32 |
1.3 Six warnings that must be published with any figure
— The definition of "using AI" differs by country. The 2024 Eurostat questionnaire listed seven technologies and 33; in 2025 image, video and audio generation was added, taking France from seven to produced 13.48% for the EU-27 34. Extending the list raises the adoption rate by construction. eight technologies
— US BTOS contains a documented break. The original question asked about AI use "in producing goods or services in 3536. Figures before and after the last two weeks"; from November 2025 it asks about use "in any business function" cannot be joined into one trend line.
— One country at one point in time can show radically different rates depending on weighting. Late 2025 in the United States: BTOS firm-weighted about 18%; the Atlanta Fed's SBU employment-weighted about 78%, with an equal-weighted estimate of 69.4%. The Federal Reserve itself describes 78% as "a good upper bound on the range of workplace access to AI tools"
— The UK statistical office warns explicitly about its headline measure. The ONS states that the indicator "treats all reported use as equivalent and does not distinguish light use, user-level use and more deeply embedded production-grade adoption"
— France separates organised use from sporadic personal use. The INSEE survey asks about AI used "in an organised way within the enterprise's departments" and "in principle excludes sporadic, personal use by employees"
— Uncalibrated country rankings are not comparable. This white paper therefore assigns an A/B/C comparability grade to every country; see chapter 2.
Only enterprise-based official statistics, graded for comparability
The chart below includes only official statistics that use the enterprise as the unit of observation, each labelled A or B for
comparability. Grade A countries share the Eurostat framework and can be compared directly; grade B countries differ in
threshold, recall period or sector coverage and support only judgements of magnitude. China, Japan, India, the United Arab
Emirates and Saudi Arabia are excluded because no verifiable enterprise-based figure exists.
Figure 1 Enterprise AI adoption on an official statistical basis, latest available year
—The three largest advanced economies cluster tightly between 19% and 20%: the United States at 19.8%38, Canada at
19.2%39 and the EU-27 at 20.0%40.
—Dispersion within the European Union exceeds dispersion between continents: in 2025 Denmark led at 42.0%, followed
by Finland at 37.8% and Sweden at 35.0%, while Romania recorded 5.2%, Poland 8.4% and Bulgaria 8.5%40.
—The growth itself is real: the EU-27 moved from 8.1% in 2023 and 13.5% in 2024 to 20.0% in 2025; the United Kingdom
rose from about 12% in September 2023 to about 35% in June 202641; Canada moved from 6.1% in Q2 2024 through
12.2% in 2025 to 19.2% in Q2 202639.
The most reproducible finding in the entire dataset
Figure 2 Enterprise AI adoption, large firms versus small firms and SMEs, latest available year
French official data show that between 2023 and 2025 the gap between enterprises with fewer than 50 employees and those
with 250 or more widened from 16 to 43 percentage points42. In Denmark, firms with 10–49 employees moved from 12%
to 37% over three years while those with 250 or more moved from 51% to 75%43. The cause is not appetite: in Korea the
leading reason for not using AI is "the burden of economic cost" at 39.6%, followed by "lack of infrastructure and staff" at
34.9% and "no AI that meets our needs" at 26.8%44; in Canada 40.0% of businesses simply consider AI "not relevant" to
them, rising to 41.4% among firms with 1–4 employees45.
A structural point: AI-using firms already account for most economic activity French official data show that in 2025 enterprises using AI accounted for66% of turnoverand59% of total employmentwithin the
survey scope, up from 49% and 40% in 2024. INSEE states explicitly that 59% "does not mean that 59% of French employees directly
use AI". This is why the employment-weighted US figure (about 78%) and the firm-weighted figure (about 18%) can both be correct.
Table 4 Country data table 1 of 5
| Country / territory | Latest enterprise AI adoption (year, unit, definition) | Generative AI | Large firms vs SMEs | Production / scaling | Financial or productivity outcome | Leading sectors / main use cases | Grade |
|---|---|---|---|---|---|---|---|
| United States | 19.8% for the collection period ending 3 May 2026 ("used AI in any business function in the last 46; about 18% two weeks", BTOS) firm-weighted in December 2025 and about 78% employment-weighted in November 2025 47 (Atlanta Fed SBU) | Workforce level: about 41% of workers used generative AI at work in November 2025 (RPS); about 54% of firms had adopted LLMs (SBU, employment-weighted)47 | 250+ employees 37%; 100–249 32%; 4 or fewer below 20%46 | n.a. (BTOS does not separate pilots from production) | No official rate of financial conversion | Information, finance and insurance, and professional, scientific and technical services all above 30% in late 2025; information firms with 250+ employees about 73% in early 202648 | B |
| China | n.a. No verifiable official enterprise adoption rate was found; the 57th CNNIC report gives no enterprise rate and CAICT publishes industry scale and company counts rather than adoption49 | Individual basis: 602 million generative AI users as at December 2025, up 141.7% on end-2024, a penetration rate of 42.8% — this is L1 and cannot substitute for enterprise adoption49 | n.a.; CAICT describes lightweight, modular and distributed architectures rather than adoption rates50 | Industry-side proxy: as at Q1 2025 China had more than 30,000 basic-level, over 1,200 advanced-level and over 230 excellence-level smart factories, covering more than 80% of manufacturing categories51 | Smart-factory basis: R&D cycles shortened by 28.4% and production efficiency up 22.3% on average (the report does not state that AI is the sole cause)51; MaaS call success rate 87.01% 50 → 99.36% | Large models are being deployed fastest in electronics, raw materials and consumer goods, covering R&D, pilot validation, production and operations51; public-cloud model calls concentrate in text processing, role play, assistants, search, coding and 50 marketing, together more than half | C (industry -side indicators are B) |
| United | About 35% in June 2026 (BICS wave 159, | n.a. (BICS does not | 250+ employees 49% | Wave 159 added | n.a. (no official rate of financial | Information and communication 58% versus | B (the BICS |
| Kingdom | enterprises with 10+ employees reporting use of at least one AI technology; about 12% in September 2023)52 | publish a separate generative AI headline) | versus 0–9 employees 28% (June 2026)52 | questions on the breadth of AI use within the firm and the share of staff using AI in daily work; the ONS states the headline does not distinguish production-grade adoption52 | conversion) | construction 13% (June 2026)52; for financial services see chapter 3 (Bank of England / FCA 75%) | sample excludes finance and insurance) |
Table 5 Country data table 2 of 5
| Country / territory | Latest enterprise AI adoption (year, unit, definition) | Generative AI | Large firms vs SMEs | Production / scaling | Financial or productivity outcome | Leading sectors / main use cases | Grade |
|---|---|---|---|---|---|---|---|
| Germany | 26% in 2025 (share of all enterprises; the EU enterprise concept applies from the 2025 53 reporting year) | Destatis does not use the term generative AI; among AI-using enterprises, generation of natural language including program code 35% and generation of images, video or audio 52%53 | 250+ employees 57%; 50–249 36%; 10–49 23%53 | n.a. (the table does not cover pilots, production or scaling) | n.a. (the table does not cover financial or productivity outcomes) | Most common technologies among AI users: text mining 52%, image, video or audio generation 52%, speech recognition 42%; autonomous machine movement only 6%53 | A |
| France | 18% (reference period Q1 2025, enterprises with 10+ employees in the main market sectors, excluding agriculture and financial services; 8 points above 2024 and 12 points above 2023)54 | Image, video and audio generation was added in 2025, taking the list from seven to eight technologies54 | 250+ employees 58%; 50–249 31%; 10–49 15%; the gap widened from 16 to 43 percentage points54 | n.a. | Structural fact: AI-using enterprises accounted for 66% of turnover and 59% of total employment in scope (49% and 40% in 2024); INSEE states 59% does not mean 59% of employees use AI54 | n.a. (the INSEE page retrieved does not publish a sector ranking) | A |
| Denmark | 42% in 2025 (enterprises with 10+ employees in private non-financial urban industries; 15% in 2023 and 28% in 2024)55; Eurostat published 56 Denmark as the EU's highest at 42.0% | In 2024 the most common uses were generating written or spoken language 18% and text analysis 17%57 | 250+ 75%; 100–249 62%; 50–99 56%; 10–49 37% (N=4,071)55 | n.a. (ITAV does not separate pilots from production) | Official basis: among AI-using enterprises, seven in ten report more efficient workflows, more than half see improved products or services and one third report higher earnings58 | 2025: information and communication 79%; business services 51%; industry 39%; trade and transport 38%; construction 24%55 | A |
Table 6 Country data table 3 of 5
| Country / territory | Latest enterprise AI adoption (year, unit, definition) | Generative AI | Large firms vs SMEs | Production / scaling | Financial or productivity outcome | Leading sectors / main use cases | Grade |
|---|---|---|---|---|---|---|---|
| Canada | 19.2% in Q2 2026 ("used AI in producing goods or delivering services over the past 12 months", 59; 12.2% CSBC; sample 21,105, responses 9,251) in Q2 2025 and 6.1% in Q2 2024 | n.a. (the release does not publish a separate generative AI headline) | 100+ employees 27.8%; 59 1–4 employees 19.9% | n.a. | Official micro research links AI adoption and productivity using pooled SDTIU 2019 and 2021 data — the data years are 2019 and 2021 and cannot support conclusions about 2025–202660 | 40.0% (two in five) of businesses say AI is not relevant to them: 41.4% of firms with 1–4 employees and 21.3% of those with 100 or more59 | B (includes micro firms of 1–4 emp loyees; 12-month recall) |
| Japan | n.a. No official rate using the enterprise as the unit of observation, comparable with Eurostat, was found. The available official figure is 55.2% of respondents answering that generative AI is used in some part of their work61 | 55.2% as above; assistance with email, minutes and document preparation 47.3%61 | Qualitative: about half of SMEs report no clearly defined policy; the white paper states SMEs lag large firms61 | n.a. (this section covers no pilot, production or scaling shares) | n.a. (no revenue, cost, productivity or ROI figures). Qualitatively, Japanese respondents most often cite efficiency gains and relief of labour shortages, whereas the other three countries cite business expansion, new customers and innovation61 | Enterprises with a policy of active or targeted use: 49.7% in FY2024 and 42.7% in FY2023; the white paper notes this remains persistently below other countries61 | C (the base is informed employee s, not ente rprises; no sample size published) |
| Korea | 32.9% (reference year 2024, national establishments with 10+ employees, population of 211,615 units, "use of AI technologies and services"; 30.3% the previous year)62 | n.a. (the compendium does not publish a separate generative AI rate) | 250+ employees 66.9%; 10–49 30.6% (27.4% the previous year); incorporated firms 34.0% versus sole proprietorships 28.7%62 | Mode of use as a depth proxy: freeware 65.1%, purchased commercial software 21.8%, contracts with external suppliers 11.1% — a predominance of free tools indicates stage one62 | Digital-industry basis: 24.9% apply AI in decision-making and business activities (15.5% in 2023); 43.5% developed or acquired AI over the past three years (10,323 firms)63 | 2024: finance and insurance 63.1%, ICT 56.0%, education services 41.4%; construction 15.9% and real estate 16.1% trail62 | B (the ques tionnaire draws on Eurostat64) |
Table 7 Country data table 4 of 5
| Country / territory | Latest enterprise AI adoption (year, unit, definition) | Generative AI | Large firms vs SMEs | Production / scaling | Financial or productivity outcome | Leading sectors / main use cases | Grade |
|---|---|---|---|---|---|---|---|
| Singapore | Overall rate n.a. (IMDA publishes no single figure for all enterprises). By segment: non-SMEs 62.5% 65 (44.0% in 2023); SMEs 14.5% (4.2% in 2023) | Among AI-using enterprises, 84% use off-the-shelf generative AI tools, 52% adopt digital solutions with AI features and 44% have implemented custom or proprietary AI (pulse survey of about 500 firms)65 | As shown: 62.5% versus 14.5%; SME growth is driven mainly by micro and small firms65 | Depth proxy: average number of functions covered — SMEs 3, non-SMEs 565 | n.a. (the quantified benefits section of the report was not retrieved). Macro: the digital economy reached 18.6% of GDP in 2024, from 14.9% in 201966 | 2024: information and communication 35.9%, professional services 25.7%, finance and insurance 22.6%, manufacturing 18.0%, retail 8.9%, accommodation and food 4.7%; functions: IT 49%, customer service 43%, finance and accounting 40%65 | B (the AI definition is broader than Eurostat) |
| India | Official n.a. Authoritative survey: 47% of enterprises have multiple AI or generative AI use cases live in production and 23% are piloting (EY–CII; 200 organisations across more than 20 industries; C-suite and senior executives, 16 November 2025)67 | n.a. (AI and generative AI adoption are not separated) | n.a. | 47% in production and 23% piloting67 | Input constraint: more than 95% of organisations allocate less than 20% of the IT budget to AI and only 4% cross the 20% threshold; 64% report selective workforce transformation and 59% a persistent shortage of AI talent67 | 91% of leaders cite speed of deployment as the leading factor in build-versus-buy; nearly 60% co-create with start-ups and 78% use a hybrid model67; the NASSCOM 68 index page publishes no adoption figure | C (n=200, s elf-selecte d C-suite sample) |
| United Arab | n.a. No official enterprise rate was found. At the | 70.1% as above (use of | n.a. | n.a. | n.a. | n.a. (no official sector or use-case | C (a teleme |
| Emirates | individual level, 70.1% of the working-age population (15–64) has used a generative AI product (Q1 2026; 59.4% in H1 2025 and 64.0% in H2), described as the first economy to pass the 70% threshold69 | generative AI products) | breakdown) | try-derived population diffusion rate, not enterprise adoption7 0) |
Table 8 Country data table 5 of 5
| Country / territory | Latest enterprise AI adoption (year, unit, definition) | Generative AI | Large firms vs SMEs | Production / scaling | Financial or productivity outcome | Leading sectors / main use cases | Grade |
|---|---|---|---|---|---|---|---|
| Saudi Arabia | Enterprise basis n.a. Government basis: about 39% of participating entities currently use AI or are actively experimenting with it and about 56% plan future adoption (SDAIA readiness and adoption survey, July 2024, n=80)71 | n.a. | n.a. | 55% are developing AI solutions, 43% procure commercial solutions and 38% integrate AI into existing applications and operations (n=80)71 | Among entities using or piloting AI, 81% report significantly improved service delivery and 61% improved decision-making (n=80)71 | Leading use cases: optimising services and processes 71%, data analytics and BI 55%, forecasting 48%, chatbots and user support 45%, document processing 42%, content generation 16%71 | C (n=80, go vernment entities only, July 2024) |
| Brazil | 17% in 2025 (active enterprises with 10+ employees using some form of artificial intelligence, TIC Empresas 2025, sample 4,174)72. A second official basis covers industry: 41.9% in 2024 (10,167 extractive and manufacturing firms 73 — with 100+ employees; 16.9% in 2022)the two bases are entirely different and must not be merged | Cetic.br does not use the term generative AI; among AI users, natural language generation rose from 20% 72 in 2024 to 30% in 2025 | 2025: large firms 50%, medium 32%, small 15% (38%, 29% and 10% in 2024)7274 | n.a. (Cetic.br states that pilot, production and scaling data are not provided)74 | n.a. (no revenue, cost, productivity or ROI data). On the industrial side, AI used in production fell from 56.4% in 2022 to 52.0% in 2024, while management reached 87.9%, commercialisation 75.2% and product or process development 73.1%73 | 2025: information and communication 49%, professional activities 24%, services 19%, trade 17%, industry 14%, construction 14%, accommodation and food 10%; by type, workflow automation 68% and autonomous machine movement 7%72 | B (CATI survey; see the metho dology note) |
| EU-27 total | 20.0% in 2025 (enterprises with 10+ employees, NACE Rev.2 C–J, L–N and group 95.1; 13.5% in 2024, 8.1% in 2023 and 7.7% in 2021)75 | 2025: image, video or audio generation 9.5%; generation of written or spoken language 8.8%75 | 2024: large enterprises 41.17% versus 13.48% for all; of 1.54 million enterprises about 83% are small, 14% medium and 3% large, with a sample of 157,00076 | n.a. (the questionnaire does not measure pilots, production or scaling) | n.a. (the questionnaire does not measure financial outcomes) | 2025 most common uses: text mining 11.8%, image, video or audio generation 9.5%, generation of written or spoken language 8.8%, converting speech to machine-readable format 7.2%75 | A (the reference basis) |
kept separate Population-level diffusion and organisation-level survey series
It is currently the only time series covering a large number of countries on a single basis, but it measures individual use, not enterprise adoption
Table 9 Population-level AI diffusion, 13 of the top 30 economies
| Economy | H1 2025 | H2 2025 | Change |
|---|---|---|---|
| United Arab Emirates | 59.40% | 64.00% | +4.60% |
| Singapore | 58.60% | 60.90% | +2.30% |
| Norway | 45.30% | 46.40% | +1.10% |
| Ireland | 41.70% | 44.60% | +2.90% |
| France | 40.90% | 44.00% | +3.10% |
| United Kingdom | 36.40% | 38.90% | +2.50% |
| Canada | 33.50% | 35.00% | +1.50% |
| Sweden | 31.20% | 33.30% | +2.10% |
| Denmark | 26.60% | 28.70% | +2.10% |
| Germany | 26.50% | 28.60% | +2.10% |
| United States | 26.30% | 28.30% | +2.10% |
| Korea | 25.90% | 30.70% | +4.80% |
| Finland | 25.60% | 27.30% | +1.70% |
Table 10 Organisation-level series by region (McKinsey survey)
| Region | 2023 | 2024 | 2025 |
|---|---|---|---|
| North America | 61% | 77% | 91% |
| Europe | 58% | 75% | 90% |
| Asia-Pacific | 48% | 78% | 88% |
| Greater China | 57% | 80% | 82% |
| Developing markets | 49% | 72% | 88% |
| All regions | 49% | 78% | 88% |
Risk of misreading Table 9 must never appear on the same ranking chart as Tables 4 to 8. France ranks fifth on population diffusion at 44.0%, yet its official enterprise adoption rate is only 18%, below the EU-27 figure of 20.0%7879. The AI Index makes the same point in reverse: "despite leading in AI investment and model development, the United States fell to 24th with a population-level adoption rate of 28.3%", and it finds a strong, statistically significant positive correlation between diffusion and GDP per capita77. Table 10 is an executive-survey series; the AI Index notes in the same place that China and Europe posted the largest year-on-year gains, at 13 and 11 percentage points respectively.
One source, one sample, one basis for cross-sector comparison
The most common error in sector comparison is placing figures from different surveys side by side. This section therefore
begins with the only single-source, single-sample sector benchmark available: the McKinsey global survey of 1,491
participants fielded from 16 to 31 July 2024, measuring regular use of generative AI in at least one function80. The sample
dates from July 2024, is not contemporaneous with the 88% headline for 2025, and cannot be ranked alongside official
enterprise statistics.
Figure 3 Generative AI adoption by sector on a single consistent sample, July 2024
—Within the consistent sample the sector spread is 29 percentage points (technology 88% versus energy and materials
59%), whereas on official bases the spread can exceed 60 percentage points (Danish information and communication
79% versus Singaporean accommodation and food 4.7%)8182.
—Sector ceilings determine product strategy: finance and information and communication are the only two sectors in the
high-adoption band, with 75% of UK regulated financial firms already using AI83.
—The functional pattern is consistent across sectors: functions tied to information processing, software, customer
interaction and internal knowledge work adopt fastest, while strategy and corporate finance and risk and compliance
"remain low in most industries" — with financial services the sole exception84.
Lower levels, identical ordering
Figure 4 Sector AI adoption in four national statistical systems; comparable within panels only
Absolute levels on official bases are far below executive surveys, but the ordering is highly consistent: information and
communication leads, construction and accommodation and food trail. That ordering recurs across four mutually
independent official statistical systems — Denmark85, Korea86, Singapore87 and Brazil88 — which makes it the most robust
sector conclusion in this white paper. Sector taxonomies, size thresholds and AI definitions differ, so comparison is valid
only within each panel.
Table 11 Sector data table 1 of 4
| # | Sector | AI / generative AI adoption (year) | Production / scaling | Most common use cases | Quantified outcomes | Main deployment barriers | Source and sample notes |
|---|---|---|---|---|---|---|---|
| 1 | Software and ICT | Technology sector generative AI 88% (July 2024, 89; official bases: Denmark information and n=199) 90; United Kingdom communication 79% (2025) 91; US information firms with 58% (June 2026) 92; Korea 250+ employees about 73% (early 2026) 93 ICT 56.0% (2024) | Functions where agents are already scaled in the technology sector: software engineering 24%, IT 22%, service operations 21% (2025) — the highest anywhere, and still around one in five94 | Software engineering and code generation; IT operations; knowledge management; customer interaction | GitHub Copilot users completed 26% more pull requests (Cui et al. 2025); marketing teams increased multimodal ad output by 50% (Ju & Aral 2025); counter-evidence: METR found experienced open-source developers were 19% slower with AI, a result later work did not replicate94 | A counter-intuitive learning penalty: engineers who leaned heavily on AI to learn new libraries showed no statistically significant speed gain (Shen & Tamkin 2025)94 | Multiple sources; official statistics are enterprise-based, McKinsey is a self-selected executive sample |
| 2 | Financial services | 75% of UK regulated financial firms already use AI and a further 10% plan to within three years (58% plus 14% in 2022; 118 responses)95; McKinsey 89; basis, generative AI 63% (July 2024, n=193) 93; Singapore Korea finance and insurance 63.1% 96 finance and insurance 22.6% | Sub-sector divergence: insurance 95% and international banks 94% already use AI; financial market infrastructure firms are lowest at 57%95 | Risk and compliance (the only sector with high adoption in that function); fraud detection; customer service; marketing and sales (generative AI 29%) | n.a. (the Bank of England and FCA survey publishes no sector-level financial returns) | Regulation and model risk dominate; the 2024 survey added generative AI questions in response to its growth95 | Regulator questionnaire, n=118; the UK ONS BICS sample excludes finance and insurance, so the two are not complementary |
| 3 | Professional and legal services | Organisational use 22% in 2025, close to double the 12% of a year earlier (nearly 1,800 legal, tax, accounting, corporate risk and government professionals worldwide)97; McKinsey basis 68% 89 (July 2024, n=179) | Generative AI adoption in corporate tax departments rose to 75%, up 20 points in a year97 | Document-intensive workflows; research and retrieval; advisory work; knowledge management (58%) | Accountants increased throughput by 55% in weekly client support (Choi & Xie 2025)94 | 64% of professionals have had no training in the professional use of generative AI, and about one third of legal and risk professionals remain unsure whether their field should use it at all97 | Vendor survey (Thomson Reuters sells legal AI), n≈1,800 |
Table 12 Sector data table 2 of 4
| # | Sector | AI / generative AI adoption (year) | Production / scaling | Most common use cases | Quantified outcomes | Main deployment barriers | Source and sample notes |
|---|---|---|---|---|---|---|---|
| 4 | Media and telecom | McKinsey basis, generative AI 65% (July 2024, 98; NVIDIA's third annual telecom survey n=77) (450+ practitioners) reports that almost all respondents say their company is actively 99 deploying or evaluating AI projects | More than one third are investing or plan to invest in AI-RAN; 40% are deploying AI in network planning and operations99 | Customer experience optimisation (44%); network planning and operations (40%); field operations optimisation (33%) | 84% say AI helped raise annual revenue, of which 21% report revenue growth above 10% in specific areas; 77% say it reduced annual operating costs; 60% cite employee productivity as the largest benefit99 | n.a. (the survey does not quantify barriers) | Vendor survey; NVIDIA sells the related hardware, so these outcomes are not industry averages; n>450, self-selected |
| 5 | Manufacturing | AI/ML in use at facility or network scale 29% and generative AI deployed at the same scale 24%; piloting AI/ML 23% and piloting generative AI 38% (600 executives, revenue of $500m or more, more than 1,000 employees)100 ; official bases: Brazilian 101 industry with 100+ employees 41.9% (2024); 102 Danish industry 39%; Singapore 103 manufacturing 18.0% | Production is the weakest link: the share of Brazilian industrial firms applying AI in production fell from 56.4% in 2022 to 52.0% in 2024 while management reached 87.9%101 ; on the McKinsey basis the manufacturing function shows 98 only 7% generative AI use | Predictive maintenance; quality inspection; digital twins and virtual commissioning; production scheduling; R&D and design | Foxconn cut factory build and planning time by about 50% and accelerated CFD thermal simulation 150-fold104 ; on the Chinese smart-factory basis, R&D cycles −28.4% and production efficiency 105 +22.3% | An insufficient data and technology foundation: 78% of leaders direct more than 20% of their improvement budget to smart manufacturing to build it, and respondents rate their own maturity as barely at industry standard100 | Deloitte n=600 (skewed to large firms); IBGE is an official survey of 10,167 enterprises |
| 6 | Healthcare and life sciences | Individual use by physicians: 81% use AI in practice in 2026, more than double the 38% of 2023, with average use cases per physician rising from 1.1 to 2.3106 ; organisational level, McKinsey 98 basis 79% (July 2024, n=113) | Regulated products provide evidence of production deployment unique to this sector: the FDA maintains an AI-Enabled Medical Device List, stating these devices have met applicable premarket requirements while also noting the list "is not a comprehensive resource"107 | Summarising medical research; clinical documentation; diagnostic support; software engineering (37%) | More than three quarters of physicians say AI has improved their ability to care for patients (65% in 2023) and 70% see it as a way to automate burnout-inducing tasks106 | Robust validation of safety and efficacy (88%) and data privacy (86%) are seen as decisive for wider adoption; 88% worry about skill atrophy106 | The AMA survey is an annual industry-body survey (no sample size published); the FDA list is a regulatory register |
Table 13 Sector data table 3 of 4
| # | Sector | AI / generative AI adoption (year) | Production / scaling | Most common use cases | Quantified outcomes | Main deployment barriers | Source and sample notes |
|---|---|---|---|---|---|---|---|
| 7 | Retail and consumer goods | McKinsey basis 79% (July 2024, n=111); official bases are markedly lower: Singapore retail 109110 8.9% and Brazilian trade 17%; US BTOS places retail, accommodation and food services 111 among the laggards | no production or scaling share) | the sector's highest function); product search and recommendation; customer service; supply chain and inventory management | 45% of consumers use AI somewhere in the purchase journey, including researching products 41%, interpreting reviews 33% and finding deals 31%112 | cross-channel, cross-system and data challenges112 | sample is only 200; official statistics are more conservative |
| 8 | Transport and logistics | 44% of shippers already use AI in transport planning and optimisation; 42% of carriers deploy AI in pricing and route optimisation and 39% for real-time tracking (230+ European and North American executives)113 ; official bases: Singapore 109 transport and storage 10.3% and Brazilian 110 transport 16% | The report states explicitly that most companies remain at an early stage and provides no quantified split between pilots, production and scaling113 | Transport planning and optimisation; rate and route optimisation; real-time visibility and ETA monitoring (52%); freight procurement | n.a. (no quantified revenue or cost outcomes). On expectations: 86% of shippers expect AI to affect transport planning significantly within three to five years and 59% of carriers see pricing and route optimisation as the main source of value113 | Limited trust in automation: two thirds of shippers and more than half of carriers still see AI as augmenting rather than replacing human decisions113 | Vendor survey, n>230, Europe and North America |
| 9 | Hospitality and travel | 41% of European hotels use AI technologies, 43% use none and 16% plan to adopt soon (1,485 hotels across Austria, France, Germany, Greece, Italy and Switzerland)114 ; official bases: Singapore accommodation and food 4.7%, the lowest of any 109110 sector, and Brazil 10% | Very shallow: 29% of adopters implemented within the past two years and only 4% have more than three years of experience114 | The study lists application areas as operations, decision support and competitiveness (the full use-case ranking requires returning to the source) | n.a. | Structural barriers: 82% of the sample are independent hotels, 81% have fewer than 100 rooms and 49% employ fewer than 10 people114 ; in Korea the leading reason for non-adoption in accommodation and food is cost, 115 at 49.0% | Joint academic and industry-body survey, n=1,485, six countries only |
Table 14 Sector data table 4 of 4
| # | Sector | AI / generative AI adoption (year) | Production / scaling | Most common use cases | Quantified outcomes | Main deployment barriers | Source and sample notes |
|---|---|---|---|---|---|---|---|
| 1 | Energy and utilities | North American utilities: 41% of executives report | As shown: 41% fully integrated | Grid optimisation (57%); | The IEA finds that widespread use of | The IEA lists missing or | The IEA is an |
| 0 | AI-related technology is now "fully integrated" and a further 40% have made major investments with mature projects under way (500 US and 116 Canadian executives); McKinsey basis, energy and materials 59% (n=142, the lowest of any 117 sector) | versus only 27% of executives a year earlier who expected full integration within five years — reality ran ahead of expectations116 | safety and hazard identification (53%); demand forecasting (51%); predictive maintenance | existing AI applications to optimise industrial processes could save energy equivalent to more than the total energy consumption of Mexico today, and improve forecasting and integration of variable renewables118 | inaccessible data and digital infrastructure, skills shortages and persistent digital and physical security concerns that "often outweigh potential efficiency gains"; only 2% of energy start-up equity funding goes to companies with an AI value proposition118 | intergovernmental report with no adoption percentage; Itron is a vendor survey, n=500 | |
| 1 | Construction and | Construction: 45% of organisations report no AI | The most extreme | Design optioneering; | n.a. (RICS states it publishes no | Skills shortages and integration | RICS n>2,200; JLL |
| 1 | real estate | use at all (more than 2,200 professionals worldwide)119 ; commercial real estate: 88% of investors and 92% of occupiers have launched AI pilots (1,500+ senior decision-makers, 16 120 markets); official bases: UK construction 121122 13%, Danish construction 24%, Korean 123 construction 15.9% | adoption-to-value gap in this white paper: only 1% of construction firms have scaled AI across projects119 ; in commercial real estate, piloting is near universal yet only 5% report achieving all their AI goals and 120 47% achieved two or three | scheduling; risk management; cost control; JLL identifies 56 AI use cases across the CRE value chain | quantified revenue, cost, productivity, quality or risk outcomes). On expectations, nearly 70% of project managers and quantity surveyors believe AI will help deliver greater value119 | challenges: one quarter of firms plan to increase AI spending over the next 12 months, 28% have no plans and 22% are unsure119 ; more than 60% of property 120 investors are still not ready | n>1,500 across 16 markets; both are industry-body surveys |
| 1 | Public sector and | 67% of OECD countries use AI to improve the | n.a. (the OECD states it provides | Process automation; | Saudi Arabia: among entities using or | The OECD emphasises that how | The OECD is an |
| 2 | education | design and delivery of public services (2024)124 ; Saudi government entities about 39% use or are 125 piloting AI (n=80); Korean education services 123 41.4% (2024) | no pilot, production or scaling shares and publishes neither the sample size nor the method behind the 67%)124 | resource allocation such as staff scheduling; decision support; citizen engagement; personalised service pathways | piloting AI, 81% report significantly improved service delivery and 61% improved decision-making (n=80)125 ; the OECD publishes no quantified outcomes | people perceive, trust and interact with AI-driven decisions will determine effectiveness and legitimacy, and notes that civil servants are "taking matters into 124 their own hands" (shadow use) | intergovernmental study (no sample size for the 67%); SDAIA n=80 |
Chapter 3 continued — Four kinds of "use" that must be separated Why 81% and 1% can both be true of the same sector
A single sector can report 81% and 1% without contradiction, because the two figures measure different things. Any sector conclusion must first state which layer it belongs to.
Table 15 Four kinds of use, compared across sectors
| Layer | Representative figure 126 | Meaning and caveats Individual behaviour; implies nothing about institutional process, |
|---|---|---|
| Individual employee use | 81% of physicians use AI in practice (2026) | validation or governance |
| Shadow employee use | 54% of employees say they use AI tools even without formal authorisation127 | Shows tools are easy to obtain but not organised into products; also a governance risk |
| Organisational adoption | Organisational generative AI use in professional services is only 22%, although 89% of the same professionals see use cases128 | Formally deployed by the institution, but not necessarily in core processes |
| Production and scaling | Construction 1% scaled across projects; commercial real estate 5% achieved all AI goals; agentic systems scaled enterprise-wide 23%129130131 | The only layer that can be connected to the profit and loss account, and the narrowest point in the chain |
| Sector conclusions | ||
| First, the ordering of sectors is co | nsistent across official statistics and surveys and can be used to pr | ioritise market entry. Second, |
| absolute levels must never be quoted | across measurement systems. Third, the sectors with the widest gap — | construction and |
| commercial real estate — are also th | e sectors with the most pilots, which shows that pilot counts are not | a measure of capability. |
Five measures, five different bases
Figure 5 Five maturity measures on different bases; not one continuous conversion funnel
This must not be restated as "AI does not work". The correct formulation is that personal exposure and tool-level use are
close to mainstream while organisational workflow redesign and auditable financial returns remain a minority
achievement. The gap exists inside a single survey: 88% use AI regularly, 62% are at least experimenting with agents, and
only 23% have scaled agentic systems anywhere in the enterprise132. Fully scaled agent use is in single digits in almost
every business function, and even in IT and knowledge management about two thirds or more of respondents report no
use at all133.
A definitional discrepancy that must accompany any citation The web version of the AI Index 2026 Economy chapter states that generative AI is now used in at least one business function by70%
of organisations, whereas figure 4.3.1 in the PDF of the same chapter gives79%for generative AI in 2025 (and 88% for AI). This white
paper cites the PDF values and flags the discrepancy rather than merging the two.
Table 16 Value conversion table 1 of 3
| Indicator | Value | Year | Research body and sample | Definition and basis |
|---|---|---|---|---|
| Organisations using AI regularly in at | 88% (78% the previous year) | to 29 July 2025, weighted by each country's contribution to global | experimentation by a few employees through to embedding | |
| least one business function | GDP; 38% of respondents are at organisations with revenue above $1bn | across rewired business units135 | ||
| Organisations using generative AI | 79% (71% in 2024)136 | 2025 | As above (reproduced as AI Index figure 4.3.1) | Regular use of generative AI in at least one business function |
| Using AI in more than one function | More than two thirds; three or more functions half | 2025 | As above | — |
| At least experimenting with agents | 62%, of which 39% "have begun experimenting" | 2025 | As above | — |
| Agentic systems scaled somewhere in | 23% | 2025 | As above | "Expanded deployment and adoption within at least one business |
| the enterprise | function"; most report only one or two functions | |||
| Agents "fully scaled" by function | Single digits in almost every function; the highest in the technology sector are software engineering 24%, IT 22% and service operations 21% | 2025 | As above (AI Index figures 4.3.7 and 4.3.8) | In most functions, most respondents report no agent use at all; even in IT and knowledge management about two thirds or more report none |
| Executives describing their generative | 1%135 | 2024 | McKinsey (developed-market supplementary survey) | "Mature" |
| AI rollout as "mature" | ||||
| Reported improvement in | Innovation 64%, employee satisfaction 45%, customer | 2025 | McKinsey (AI Index figure 4.3.5) | Self-assessed as improved, no impact, worse or unknown; as |
| organisational metrics (organisations | satisfaction 45%, competitive differentiation 45%, cost | many respondents report improvement as report no impact, and | ||
| using AI regularly) | 38%, profitability 36%, organic revenue growth 33%, talent attraction and retention 33%, market share 25% | at most 7% say AI made cost metrics worse | ||
| EBIT impact | Organisations investing $25m or more in responsible AI are "far more likely to realise material AI benefits, including EBIT impact above 5%"137 | 2026 | McKinsey, 2026 AI Trust Maturity Survey; about 500 organisations, December 2025 to January 2026 | The page gives no share of respondents reporting EBIT impact, only this association |
Table 17 Value conversion table 2 of 3
| Indicator | Value | Year | Research body and sample | Definition and basis |
|---|---|---|---|---|
| Distribution of corporate AI maturity | Future-built 5%, scaling 35%, and the remaining 60% obtain "almost no material value, reporting minimal revenue and cost benefits despite significant 138 investment" | 2025 | BCG Build for the Future 2025, n=1,250 across 68 countries in Asia, Europe and North America | The four levels are scored on 41 underlying capabilities (0–25 stalled, >25–50 emerging, >50–75 scaling, >75–100 future-built)139 |
| Organisations with agents broadly | 13%140 | n.a. (no | BCG | "Broadly integrated these agents into workflows" |
| embedded in workflows | year on the BCG page) | |||
| Depth of employee adoption | More than 85% remain at stage two (task assistance) and stage three (delegation); fewer than 10% reach stage four (semi-autonomous collaboration) or beyond | n.a. | BCG survey of software developers at three different organisations (no sample size published) | Five stages: information assistance → task assistance → delegation → semi-autonomous collaboration → fully autonomous orchestration |
| "Pilot failure" | "For 95% of companies in the dataset, generative AI implementation fell short of expectations"; about 5% of AI pilots achieved rapid revenue acceleration141 | 2025 | MIT NANDA, The GenAI Divide, built on 150 leader interviews, 350 employee surveys and analysis of 300 public AI deployments | "Failure" is operationalised as falling short of expectations, stalling, or having almost no measurable effect on the profit and loss account; the reporting page does not state which part of the research the 95% base comes from, nor its geographic scope |
| Build versus buy | Purchasing AI tools from specialist vendors and building partnerships succeeds "about 67% of the time"; internal builds succeed only one third as often | 2025 | As above | The page gives no sample size for either route |
| Firm size and deployment stage | "Larger companies are the most likely to report that their AI projects have reached the scaling stage" | 2025 | McKinsey (AI Index figure 4.3.6, five revenue bands) | The figures cannot be aligned one-to-one with the bands; precise values require the original AI Index chart142 |
Table 18 Value conversion table 3 of 3
| Indicator | Value | Year | Research body and sample | Definition and basis |
|---|---|---|---|---|
| Governance capability and returns | About two thirds of respondents cite security and risk concerns as the leading barrier to fully scaling agentic 143 AI | 2026 | McKinsey AI Trust Maturity Survey, about 500 organisations | — |
| Data and digital foundation and value | 78% of manufacturing leaders direct more than 20% of their improvement budget to smart manufacturing to build the data and technology foundation144 | Fielded Aug ust–Septem ber 2024 | Deloitte, n=600 | — |
| Insufficient investment and value | More than 95% of Indian organisations allocate less than 20% of the IT budget to AI and only 4% cross the 20% threshold145 | 2025 | EY–CII, n=200, C-suite and senior executives | — |
| Distribution of value pools | 70% of AI value potential sits in core business functions (62% in 2024); the IT share rose 6 percentage points to 13%; R&D and innovation alone account for 15%146 | 2025 | BCG Build for the Future 2025, n=1,250 | This is the distribution of value potential, not adoption by function |
| Tool availability and maturity | In less mature industries fewer than 50% of employees have access to generative AI tools such as Copilot or ChatGPT; in more mature industries more than 70% do146 | 2025 | As above | — |
| Firm size and breadth of generative AI | Companies with revenue above $500m use generative AI more broadly across the organisation | 2024 | McKinsey147 | — |
studies that cannot be averaged Results range from −19% to +200% for comparable work
148
Table 19 Task-level productivity studies compiled in AI Index 2026, figure 4.4.27
| Study | Occupation | Application | Change in productivity | Who benefits most |
|---|---|---|---|---|
| Reimers & Waldfogel (2026) | Software engineers | LLM-generated content | +200% in output volume; publications tripled | New entrants drive volume; pre-AI authors maintain quality |
| Brynjolfsson et al. (2025) | Customer support agents | Conversational assistant | +14% to +15% in issues resolved per hour | Less experienced or lower-skilled agents, +30% to +35% |
| Cui et al. (2025) | Software developers | GitHub Copilot | +26% in pull requests completed | Junior and less experienced developers |
| Ju & Aral (2025) | Marketing teams | Multimodal ad creation | +50% in output per person | Human–machine teams |
| Choi & Xie (2025) | Accountants | AI bookkeeping | +55% in weekly client-support throughput | Experienced accountants using AI confidence scores to target review |
| Shen & Tamkin (2025) | Developers | Learning new libraries | 0%; the change in speed is not statistically significant | High scorers who use AI for conceptual enquiry and avoid the learning penalty |
| Becker et al. (2025) | Developers | Open-source tooling | −19%; slower when using AI | None; "a significant gap between perceived help and actual performance" |
Macroeconomic evidence is equally dispersed — Aldasoro et al. (2026), 12,000 European firms, 2019–2024: labour productivity +4%, with each 1% of spending on training adding 5.9 percentage points of gain.
— Yotzov et al. (2026), 6,000 executives in the United States, United Kingdom, Germany and Australia: expected productivity +1.4%, output +0.8% and employment −0.7% over three years; the AI Index summarises this as broad adoption with minimal realised productivity gains.
— Brynjolfsson (2026): US productivity grew 2.7% in 2025, close to double the 1.4% annual average of the preceding decade.
— Filippucci et al. (OECD, 2025), a ten-year horizon for the G7: annual labour productivity growth +0.4 to +1.3 percentage points for the United States and United Kingdom, and +0.2 to +0.8 for Italy and Japan.
— Penn Wharton Budget Model (2025): a contribution to total factor productivity of +0.01 percentage points, described as negligible.
— Brynjolfsson et al. (2025), US ADP payroll data to 2025: employment of early-career workers −15% to −16%.
The AI Index's own qualitative judgement is that the evidence is "neither settled nor uniformly positive", with smaller gains on tasks requiring deeper reasoning148 .
What supports the conclusion, and what must be said alongside it
1.The gap inside a single survey. 88% of organisations use AI regularly, but only 23% have scaled agentic systems
anywhere in the enterprise149, and fully scaled agent use is in single digits in almost every function150.
2.Self-reported financial outcomes run well below adoption. Only 36% see improved profitability and 33% improved
organic revenue growth, and as many respondents report improvement as report no impact150.
3.An independent second source reaches the same conclusion. BCG (n=1,250, 68 countries) finds 60% of companies
obtain almost no material value and only 5% are future-built151.
4.Extreme sector cases. Commercial real estate pilots at 88% to 92%, yet only 5% achieve all their AI goals152, and only
1% of construction firms scale across projects153.
5.The mismatch between individual and organisational use. French generative AI diffusion reached 44.0% of the
population in H2 2025 against official enterprise adoption of 18%; US population-level diffusion of 28.3% compares with
BTOS enterprise adoption of about 18% to 19.8%150154155.
The MIT NANDA figure of 95% is widely quoted as "95% of AI projects fail". The original wording is that "for 95% of
companies in the dataset, generative AI implementation fell short of expectations", where failure is operationalised as
stalling or having almost no measurable effect on the profit and loss account. The research rests on 150 interviews, 350
employee surveys and analysis of 300 public deployments; it is not a probability sample of the enterprise population and
its geographic scope is not published156. This white paper requires all three qualifications to be stated whenever the
number is used.
"Personal exposure and tool-level usage are close to mainstream. Organisational workflow redesign and auditable
financial returns remain a minority achievement. The gap does not lie in model capability but in depth of
deployment, the data and governance foundation, and process change that can be booked in the profit and loss
account."
The finding of chapter 4; quotable in full.
Dated first-hand statements and the division of labour between platforms
This chapter uses only dated, verifiable first-hand statements and deployed cases. It contains no market forecasts.
Table 20 NVIDIA's strategic statements, verbatim and dated
| Date | Source | Statement (extract) |
|---|---|---|
| 16 Mar 2026 | NVIDIA Newsroom | "Physical AI has arrived — every industrial company will become a robotics company." NVIDIA's full-stack platform, spanning computing, open models and software frameworks, is 157 described as the foundation for the robotics industry |
| 5 Jan 2026 | NVIDIA Newsroom (CES) | "The ChatGPT moment for robotics is here." "Breakthroughs in physical AI — models that understand the real world, reason and plan actions — are unlocking entirely new applications"158 |
| March 2026 (GTC) | Data Center Frontier, on-site report | "We have digital agents. Now we have physically embodied agents. We call them robots."; "Tokens are the new commodity. AI factories are the infrastructure that produces them."; "If you have the wrong architecture, even if it's free, it's not cheap enough."159 |
| 18 Mar 2025 | NVIDIA Newsroom (Rev Lebaredian, VP of Omniverse) | "Omniverse is an operating system that connects the world's physical data to the realm of physical AI."160 |
Why this is more than a marketing narrative Four robot manufacturers — FANUC, ABB Robotics, YASKAWA and KUKA, with more than two million robots installed worldwide — have integrated Omniverse libraries and the Isaac simulation framework into their virtual commissioning 157 offerings and are embedding Jetson modules in controllers for real-time inference on the line . That is a change to the production process chain, verifiable at the enterprise level, rather than a roadmap commitment.
What each platform does, and the quantified claims attached to it
Table 21 Division of labour across the physical AI stack
| Platform | Role (verbatim or near-verbatim) | Key quantified facts 161162 |
|---|---|---|
| Omniverse | "An operating system built on the OpenUSD framework that lets developers unify physical-world data and applications"; at GTC 2026 NVIDIA introduced the Omniverse DSX blueprint, allowing operators to create a physically accurate digital twin of an AI factory — layout, power distribution, cooling, network fabric and load behaviour — before construction begins | — |
| Isaac / Isaac Sim | Used "to create and train robots in simulation before deploying them in the real world"; Foxconn uses Isaac Sim for "simulation-driven design and evaluation ahead of physical deployment", simulating tasks such as screw driving and cable insertion | —163 |
| Isaac GR00T | A general-purpose robot foundation model. GR00T N2 previewed in March 2026, with availability planned for late 2026 | "Helps robots complete new tasks in new environments with more than twice the success rate of leading vision-language-action models" and currently ranks first on the MolmoSpaces and RoboArena generalist robot policy leaderboards; the CES 2026 release was GR00T N1.6164165 |
| Cosmos | A world foundation model. Cosmos 3 was announced in March 2026 as "the first world foundation model to unify synthetic world generation and vision" | The CES 2026 release comprised Cosmos Transfer 2.5, Cosmos Predict 2.5 and Cosmos Reason 2165 |
| Jetson / Jetson Thor | Robotics compute at the edge; "built for generative reasoning models", letting physical AI agents "run in real time at the edge while minimising reliance on the cloud" | 7.5x the AI compute, 3.1x the CPU performance and 2x the memory of the previous Jetson Orin; the Jetson T5000 module starts at $2,999 per 1,000 units and the developer kit at $3,499166 |
| Jetson T4000 / IGX Thor | Blackwell-architecture edge modules launched in January 2026; IGX Thor "extends robotics to the industrial edge with high-performance AI compute, enterprise software support and functional safety" | $1,999 per 1,000 units, 4x the performance of the previous generation, 1,200 FP4 TFLOPS, 64GB of memory, a configurable 70W power envelope, and a 4x improvement in energy efficiency and compute165 |
| Mega / Metropolis | Mega is "an Omniverse blueprint for testing multi-robot fleets at scale in industrial digital twins"; Metropolis underpins the video search and summarisation blueprint used to build agents that monitor activity across a facility | —161 |
| Nemotron / PhysicsNeMo | Nemotron 3 is "a family of open models, data and libraries to support transparent, efficient and specialised agentic AI development across industries"; PhysicsNeMo is a physics-informed machine learning framework | Foxconn used PhysicsNeMo to accelerate CFD simulation 150-fold against conventional methods167163 |
Verifiable deployments and their quantified outcomes
Table 22 Deployed physical AI cases
| Company | What was deployed | Quantified outcome | Source 168 |
|---|---|---|---|
| FANUC, ABB Robotics, YASKAWA, | More than two million robots installed worldwide; all four are integrating Omniverse libraries and the | 2 million robots installed | NVIDIA, 16 March 2026 |
| KUKA | Isaac simulation framework into their virtual commissioning offerings "to develop and validate complex robotic applications and entire production lines through physically accurate digital twins", and are embedding Jetson modules in controllers for real-time inference on the line | ||
| Foxconn (Fii) | The Fii Omniverse Digital Twin platform builds virtual replicas of factories using Omniverse libraries and OpenUSD; Isaac Sim supports simulation-driven design and evaluation; PhysicsNeMo handles thermal analysis; standardised USD assets allow entire lines to be assembled virtually and replicated across sites, for example from Taiwan to Mexico; the new Houston plant validated every mechanical, electrical and piping system virtually before construction | Leo Guo, general manager of Fii's robotics group: "We believe we can reduce [time to market, factory build and factory planning] by about 50%"; CFD thermal simulation 150x faster (minutes rather than hours); the FoxBrain large model was trained in four weeks | NVIDIA case study169 ; NVIDIA, 18 March 170 2025 |
| General Motors | "The largest US automaker, announced adoption of Omniverse to enhance its factories and train platforms for material handling, transport and precision welding" | — | NVIDIA, 18 March 2025170 |
| SAP and Unilever | SAP customers and partners can use Omniverse to develop their own virtual environments for warehouse management; Unilever adopted Omniverse and physically accurate digital twins to streamline and optimise marketing content creation for its products | — | NVIDIA, 18 March 2025170 |
| Agility Robotics and Boston Dynamics | Agility has integrated Jetson into the fifth generation of Digit and plans to use Jetson Thor as the onboard compute platform for the sixth; Boston Dynamics is integrating Jetson Thor into the Atlas humanoid | — | NVIDIA blog171 |
| Salesforce | Agents applied to incident resolution | Incident resolution time halved | NVIDIA, 5 January 2026172 |
| Automotive and autonomous driving | NVIDIA announced a further group of automotive partnerships including BYD, Hyundai, Nissan and Geely, alongside existing work with Mercedes-Benz and Toyota | The company states its platform will support "autonomous systems in tens of millions of vehicles a year" — a company claim, not independently verified | Data Center Frontier173 |
2024) Installations, operational stock, shares and density
Table 23 Installations and operational stock
| Indicator | Value |
|---|---|
| Global installations, 2024 | 542,076 units ("542,000" in the press release), the second-highest year on record and only 2% below the peak two years earlier; the fourth consecutive year above 500,000 |
| Historical series | 500,000 first passed in 2021; 552,946 units in 2022 (the record); 541,302 units in 2023 |
| Global operational stock, | 4,664,000 units, up 9% year on year |
| 2024 | |
| Regional shares, 2024 | Asia 74% of new installations (401,665 units, +5%); Europe 16% (85,006 units, −8%); the Americas 9% (50,077 units, −10%) |
| EU-27, 2024 | 67,819 units (−8%), or 80% of European installations |
| 2025 forecast | Global +6% to 575,000 units; Asia about 435,000; Europe below 79,000; North America about 43,500 |
| 2028 forecast | Above 700,000 units |
| China's operational stock exceed | s two million units, the largest in the world and roughly 4.5 times that of Japan, whose stock sta |
| and trails the EU average of 208 | . China, Japan, the United States, Korea and Germany together account for 80% (431,240 units) of g |
| manufacturers "climbed to 57% la | st year, against roughly 28% over the past decade"174175 |
| . |
Table 24 Major countries, 2024 (density in units per 10,000 employees)
| Country | 2024 installs | Change | Share | Density |
|---|---|---|---|---|
| China | 295,045 | +7% | 54% | 166 |
| Japan | 44,453 | −4% | 8% | 446 |
| United States | 34,164 | −9% | 6% | 307 |
| Korea | 30,596 | −3% | 6% | 1,220 |
| Germany | 26,982 | −5% | 5% | 449 |
| Singapore | n.a. | — | — | 818 |
| India | 9,100 | +7% | — | n.a. |
| Italy | 8,783 | −16% | — | top 20 |
| Mexico | 5,594 | −4% | — | 62 |
| Spain | 5,086 | +1% | — | top 20 |
| France | 4,900 | −24% | — | top 20 |
| Canada | 3,787 | −12% | — | 241 |
| United Kingdom | 2,500 | −35% | — | 112 |
The installation leader and the density leader are different countries
Figure 6 Industrial robot installations and global share by country, 2024
Figure 7 Robot density in units per 10,000 employees, 2024
Sources for Figures 6 and 7: IFR World Robotics 2025 press release and executive summary for installations, stock and shares177178, and
IFR robot density data for the 2024 reference year as published by The Robot Report179. The two charts must not be merged: installations
measure the annual addition, density measures the ratio of the installed stock to the workforce.
Concentration, localisation and the difference between volume and density
1.Concentration in five markets. China, Japan, the United States, Korea and Germany together accounted for 80%
(431,240 units) of global installations in 2024180.
2.Localisation in China. The domestic market share of Chinese manufacturers "climbed to 57% last year, against roughly
28% over the past decade"181.
3.Density and installations must be read separately. China leads installations (54%) but has a density of only 166,
ranking 22nd; Korea ranks fourth on installations but first on density at 1,220. Global average density reached 132 units
per 10,000 employees in 2024, having doubled between 2014 and 2024: EU-27 231, western Europe 267 (+3%), North
America 204 (+4%) and Asia 131 (+11%)182.
"The robot density metric provides a common basis for comparison by relating the total number of robots in use in
a country to the size of its economy as measured by its workforce."
Takayuki Ito, President of the IFR182
The other side of the picture must be shown at the same time: among enterprises that use AI, technologies for autonomous
machine movement remain in single digits — 6% in Germany183, 2% in Denmark for two consecutive years and the only
category with no growth184, and 7% in Brazil185. Robot density measures the automation stock; it is not a measure of AI
penetration.
The asymmetry between capital signals and installed penetration Between January and April 2025 China recorded140robotics-related financing deals,38.7%of all AI financing events, while
enterprise use of autonomous machine movement remains at2% to 7%. Capital is entering faster than the technology is
penetrating production lines, and this is the structural feature investors in physical AI most need to understand.
Why there is no universal "90% accuracy" bar
No authoritative body was found to set a uniform cross-industry accuracy threshold. Any claim that "90% is enough to
go live" has no source behind it. The evidence shows thresholds are set by the consequence of the task, across six separate
dimensions.
Figure 8 A six-threshold framework for industrial-grade AI
First-hand basis for each threshold in Figure 8: the FDA list of AI-enabled devices, where the route is regulatory review rather than a
percentage186; CAICT data on service success rate, TPS and time to first token187; NVIDIA Jetson T4000 and T5000 pricing, power envelope
and functional safety188189; the GR00T N2 task success claim190; the human-in-the-loop preference in logistics191; the link between
responsible-AI investment and EBIT impact192; and physicians' views on validation and skill atrophy193.
dimension Statements and metrics, not generalities
Table 25 Threshold dimensions for industrial and edge AI, with first-hand evidence
| Threshold dimension | First-hand evidence |
|---|---|
| Determinism and real-time | Jetson Thor is positioned to bring "high-speed sensor data processing and visual reasoning to the edge — workflows that |
| performance | were previously too slow to run in dynamic real-world environments", with "the low latency and high performance 194 real-world applications require" |
| Operation without connectivity | Jetson Thor lets next-generation physical AI agents "run in real time at the edge while minimising reliance on the cloud"194 |
| Functional safety certification | NVIDIA IGX Thor "extends robotics to the industrial edge with high-performance AI compute, enterprise software support and functional safety"195 |
| Unit cost and energy (a | Jetson T4000: $1,999 per 1,000 units, 1,200 FP4 TFLOPS, 64GB of memory and a configurable 70W envelope, positioned as |
| bill-of-materials constraint) | a cost-effective upgrade path195194 ; against Jetson T5000 at $2,999 per 1,000 units and a developer kit at $3,499. The economics of edge deployment are set by component cost and power envelope, not by benchmark scores. 196 |
| Architectural efficiency over "free" | Jensen Huang at GTC 2026: "If you have the wrong architecture, even if it's free, it's not cheap enough." |
| Pre-deployment validation instead of | The four major robot manufacturers use Omniverse and Isaac "to develop and validate complex robotic applications and |
| trial and error | entire production lines through physically accurate digital twins"; Foxconn assembles and validates every mechanical, electrical and piping system virtually before construction begins197198 |
| Reliability as an engineered metric | CAICT reports that across 11 Chinese MaaS platforms the average call success rate for DeepSeek-R1 rose from 87.01% in February 2025 to 99.36% in September, TPS from 17.86 to 26.76, and time to first token fell from 3.07 seconds to 1.02 seconds199 |
| The industrial significance of energy | The IEA states that "widespread adoption of existing AI applications to optimise industrial processes could deliver energy savings equivalent to more than the total energy consumption of Mexico today", while noting that missing or inaccessible data and digital infrastructure, skills shortages and persistent digital and physical security concerns "often outweigh potential efficiency gains"200 |
| Human-in-the-loop rather than full | In transport and logistics, two thirds of shippers and more than half of carriers still see AI as augmenting rather than |
| autonomy | replacing human decisions, and most prefer human-in-the-loop designs201 |
| Validation thresholds in healthcare | 88% of physicians regard robust validation of safety and efficacy, and 86% data privacy, as decisive for wider adoption202 ; the FDA's AI device list uses satisfaction of applicable premarket requirements, including focused review of overall safety and 203 effectiveness, as the entry standard |
and contracts How to set each threshold, with citable evidence
Table 26 Product threshold checklist for industrial-grade AI
| Threshold | How to set it, with citable first-hand evidence 204 |
|---|---|
| Task accuracy | Tier by consequence; set no single value. The medical-device route is regulatory review, not a percentage; robotics is measured by success rate: GR00T N2 achieves more than twice the success rate of leading VLA models on new tasks in new 205 environments — a relative multiple, not an absolute accuracy |
| Repeatability and determinism | Use the service call success rate as the operable metric: an average of 99.36% across 11 MaaS platforms in September 2025206 |
| Exception handling and human | Logistics evidence favours human-in-the-loop: two thirds of shippers and more than half of carriers see AI as augmenting |
| oversight | rather than replacing decisions207 |
| SLA and latency | Write TTFT and TPS into the contract: time to first token 1.02 seconds and TPS 26.76 (11-platform average, September 2025)206208 ; the edge requirement is "low latency and high performance" for dynamic real-world environments |
| Audit and governance | Directly related to returns: organisations investing $25m or more in responsible AI are more likely to report EBIT impact above 5%, and nearly two thirds cite security and risk as the leading barrier to scaling agents209 ; financial services are 210 monitored continuously by the Bank of England and FCA, whose 2024 survey drew 118 responses |
| Unit cost | Edge bill-of-materials cost is a hard constraint: Jetson T4000 at $1,999 per 1,000 units against T5000 at $2,999211 . On the inference side, 12 billion of 120 billion parameters are activated, and NVFP4 runs up to 4x faster than FP8 on Hopper with 212 no loss of accuracy |
| Energy | A configurable 70W envelope and a 4x improvement in energy efficiency (Jetson T4000)211 ; for sector-level savings potential 213 see the IEA (equivalent to more than Mexico's national energy consumption) |
| Operation without connectivity | An explicit product requirement: "run in real time at the edge while minimising reliance on the cloud"208 |
| Safety certification | Functional safety is an entry requirement at the industrial edge (NVIDIA IGX Thor)211 ; the energy sector cites digital and 213 physical security concerns as barriers that often outweigh efficiency gains |
| Lifecycle maintenance | "With NVIDIA CUDA ecosystem support across the full lifecycle, Jetson Thor is expected to deliver better throughput and faster response through future software releases"208 ; the long-run risk in healthcare is skill atrophy, a concern for 88% of 214 physicians |
model, but the more reliable one Three candidates, ranked by strength of evidence
Industry frequently asks which recent product represents the shift towards precision, low error rates, cost and speed rather than frontier capability. The available evidence does not point uniquely to one product, but three candidates can be ranked by strength of evidence.
Table 27 Candidate identification and assessment
| Candidate | Official positioning (verbatim or near-verbatim) | Assessment |
|---|---|---|
| Candidate 1: NVIDIA Nemotron 3 | Announced as "the most efficient family of open models, with leading | The strongest evidence. The description |
| (Nano / Super / Ultra), released 15 | accuracy" for building agentic AI applications, introducing a "hybrid latent | matches leading accuracy for its size class |
| December 2025 | mixture-of-experts architecture" that helps developers build and deploy reliable multi-agent systems; Jensen Huang emphasised "transparency and efficiency"; the ServiceNow CEO described "unmatched efficiency, speed and accuracy"215. Quantified claims for Nemotron 3 Super: 120 billion parameters with only 12 billion activated at inference; a 1 million token context window; up to 5x higher throughput and up to 2x higher accuracy than the previous generation; Mamba layers giving 4x memory and compute efficiency; multi-token prediction giving 3x faster inference; and NVFP4 216 running up to 4x faster than FP8 on Hopper with no loss of accuracy | combined with top-ranked efficiency and cost, throughput and latency as the core selling points. Note that NVIDIA also describes Nemotron 3 Ultra as "America's smartest open-weights AI model"; that wording appears in third-party reproductions and could not be verified on an official page, so it is not treated as evidence |
| Candidate 2: Isaac GR00T N2 | The official claim concerns task success rate rather than general | A partial match: success rates and real |
| (previewed March 2026) | capability: it "helps robots complete new tasks in new environments with more than twice the success rate of leading vision-language-action models" and ranks first on the MolmoSpaces and RoboArena leaderboards217 | physical tasks are emphasised. Against: it is claimed to rank first on its own leaderboards, which does not fit the description of not being the most advanced |
| Candidate 3: Jetson T4000 and IGX | The T4000's official selling points are entirely cost, energy efficiency and an | A partial match: cost, speed and reliability. |
| Thor (January 2026) | upgrade path: $1,999 per 1,000 units, 4x performance, 4x energy efficiency and a configurable 70W envelope; IGX Thor is sold on functional safety218 | Against: these are hardware modules, not models |
| Overall assessment: NVIDIA's product line f | rom late 2025 into early 2026 shows a clear engineering turn towards products that are no | t |
| the most capable but are more reliable, che | aper and faster. The clearest example is the Nemotron 3 family of open models, sold on | |
| efficiency and accuracy within a size class | , alongside Jetson T4000 and IGX Thor for the industrial edge, sold on cost, energy effic | iency and |
| functional safety. |
Claim-by-claim assessment
Claim under review: "OpenAI poached four to five hundred hardware staff from Apple, effectively an entire department, triggering litigation, and may build smart glasses to replace the phone." All status statements are as at 15 August 2026; the disputes are undecided.
Table 28 Claim-by-claim assessment, part 1 of 3
| # | Claim under review | Assessment | Evidence (verbatim or key points) |
|---|---|---|---|
| 1 | OpenAI has a transaction with Jony Ive and io | Confirmed | OpenAI's letter of 21 May 2025, updated 9 July 2025: "the io Products, Inc. team has officially merged with OpenAI"; "Jony Ive and LoveFrom remain independent and have assumed deep design and creative responsibilities across OpenAI"; io was co-founded by Jony Ive with Scott Cannon, Evans Hankey 219 and Tang Tan |
| 2 | The transaction was worth $6.5bn | Partly confirmed (press reporting, not official disclosure) | The official OpenAI page gives no figure219 . Reuters reported that "OpenAI announced the acquisition of Jony Ive's hardware start-up io Products in a deal 220 valued at $6.5bn", an all-stock transaction, and that according to sources OpenAI already held a 23% stake in io, reflecting an OpenAI valuation of $300bn; 221 MacRumors describes it as "an estimated $6.5bn" |
| 3 | The size of the io team | Not confirmed (no public figure) | OpenAI's letter describes assembling "the best hardware and software engineers, technologists, physicists, scientists, researchers and experts in product development and manufacturing" but gives no headcount219 ; none of the Reuters, MacRumors, 9to5Mac, Bloomberg Law or IPWatchdog pages reviewed provides a figure for the io team 222 |
| 4 | "Four to five hundred hardware staff poached from Apple" | Partly confirmed, but the figure must be recharacterised: it is "400+ former Apple employees" company-wide, not 400–500 hardware staff | Apple's complaint of 10 July 2026, as reproduced by 9to5Mac: "With over four hundred former Apple employees now working at OpenAI…"; Reuters 223 likewise reports "more than 400 former Apple employees now work for OpenAI"; Business Insider: "in the complaint, Apple says the total number of Apple alumni now at OpenAI is more than 400"224. Key qualification: this is Apple's own figure in its pleading, covers all former Apple employees and is not broken down by hardware function. |
Claim-by-claim assessment
Table 29 Claim-by-claim assessment, part 2 of 3
| # | Claim under review | Assessment | Evidence (verbatim or key points) |
|---|---|---|---|
| 5 | "Effectively an entire department" | Incorrect attribution (overstated) | No source supports "an entire department". What is verifiable is breadth rather than a wholesale move: Mark Gurman reported that OpenAI "brought in more than 40 people in the past month alone", spanning camera engineering, iPhone and Mac hardware, silicon, device test and reliability, industrial design, 225 manufacturing, audio, smartwatches, Vision Pro development, software and human factors; the Wall Street Journal, reviewing LinkedIn data, reported 226 that "dozens" of Apple engineers and designers had left for OpenAI in recent months; Bloomberg reports that "most" of OpenAI's Apple hires come from 227 the engineering organisation led by incoming CEO John Ternus |
| 6 | "It triggered litigation" — did Apple sue OpenAI directly | Confirmed (but the cause of action is trade secrets, not hiring as such) | On 10 July 2026 Apple sued OpenAI and two former employees in the US District Court for the Northern District of California, alleging misappropriation of its trade secrets to benefit the ChatGPT owner's move into consumer hardware; the defendants also include OpenAI Foundation, OpenAI Group PBC and io Products; the two former Apple employees are Chang Liu, a former senior systems electrical engineer, and Tang Yew Tan, a former vice-president of product design for iPhone and Apple Watch228 . TechCrunch adds that the claims are trade-secret theft and breach of contract, and notes that io is named in the 229 complaint while Ive is not a defendant |
| 7 | Was the litigation caused by hiring | Partly confirmed; the parties' positions conflict | Apple's central allegation is that "OpenAI orchestrated a broad campaign to systematically obtain and exploit Apple's confidential information through former employees, recruiting practices and supplier relationships in order to accelerate its entry into the consumer hardware business"228 . In its August 2026 motion OpenAI responded that "Apple should not be allowed to use a baseless and pretextual lawsuit to make up for its shortcomings in the talent market and in employee retention"230 |
| 8 | Whether other parties and disputes are involved | Confirmed — a second, unrelated line of litigation exists (iyO v OpenAI) | iyO, Inc. sued IO, OpenAI, Inc., OpenAI, LLC, Sam Altman and Jonathan Paul Ive, alleging that the IO mark is "confusingly similar" to its IYO mark; the Ninth Circuit affirmed the preliminary injunction in December 2025231 ; Bloomberg Law notes the injunction "does not bar all use of the 'io' trademark but 232233 restricts marketing and sales of products 'sufficiently similar' to iyO's AI audio computer"; iyO subsequently added trade-secret claims; on 27 April 2026 234 iyO issued a release stating that a federal court had granted a preliminary injunction and adding Tang Yew Tan to its amended claims; the trademark case was stayed on 28 July 2026 pending settlement talks |
Claim-by-claim assessment
Table 30 Claim-by-claim assessment, part 3 of 3
| # | Claim under review | Assessment | Evidence (verbatim or key points) |
|---|---|---|---|
| 9 | "May build smart glasses" | Partly confirmed: press reporting, not confirmed by OpenAI | On 20 February 2026 Reuters, citing The Information, reported that "OpenAI has more than 200 people working on a family of AI devices", including a smart 235 speaker; a related report lists "a smart speaker, smart glasses and a smart lamp", says the speaker will come first at an expected $200 to $300 and "will 236 not ship before February 2027 at the earliest", attributing this to "two people familiar with the project". No official OpenAI page confirms any specific 237 product category |
| 1 | "To replace the phone" | Not confirmed (speculation) | No source reviewed shows OpenAI officially claiming it will replace the smartphone. The strongest official statement is that the team is "focused on |
| 0 | developing products that inspire, empower and enable"237 ; Reuters describes the purpose of the transaction as capturing growing demand for physical AI 235 and augmented reality. The closest media formulation is Bloomberg's analysis headline "How Apple's Lawsuit Threatens to Disrupt OpenAI's Bid to Rival the iPhone", which is interpretation rather than an official statement | ||
| 1 | Current status of the litigation (as at 15 August | Confirmed: the disputes are | On 3 August 2026 OpenAI published an official response, "Apple is getting this wrong", updated on 6 August to note that "you can read our motion to dismiss |
| 1 | 2026) | undecided | here"; OpenAI argues that Apple's request for a preliminary injunction "is based on false information and is entirely unnecessary because we do not have, and do not want, any of their trade secrets"238; as at the Business Insider report, "Apple had not yet responded in court to OpenAI's motion and the judge had not ruled on the request to dismiss"; an Apple spokesperson said "significant evidence has emerged that individuals employed by OpenAI improperly 239 obtained Apple's confidential information about unreleased technology" |
A single paragraph that can be quoted in full
"On 10 July 2026 Apple sued OpenAI and two former Apple employees (Chang Liu and Tang Yew Tan) in the US
District Court for the Northern District of California. The causes of action are trade-secret misappropriation and
breach of contract, not hiring as such. In its complaint Apple states that more than 400 former Apple employees
now work at OpenAI, a company-wide figure that is not broken down by hardware function. OpenAI has filed a
motion to dismiss and publicly denies holding any Apple trade secrets. A separate line of litigation, unrelated to
Apple, involves the start-up iyO, which sued OpenAI, Sam Altman and Jony Ive over the 'io' trademark; the Ninth
Circuit affirmed a restrictive injunction in December 2025 and the case was stayed in July 2026 pending settlement
talks. Product categories such as smart glasses and smart speakers appear only in press reporting, notably by The
Information as relayed by Reuters; OpenAI has never confirmed them and has never officially claimed it will
replace the smartphone."
Recommended for verbatim use.Status: as at 15 August 2026 neither line of litigation has been finally decided by a court. This white
paper makes no prediction about the outcome.
—Talent mobility is not itself a cause of action. The verifiable causes are trade secrets and breach of contract, which
means due diligence should focus on onboarding procedures, revocation of system access and document segregation
rather than on hiring volumes240241.
—Trademark and brand risk exists independently. The injunction affirmed by the Ninth Circuit restricts marketing and
sales of products sufficiently similar to iyO's, rather than banning the mark outright242.
—Product categories should not enter an investment thesis. Hardware categories and timelines rest on anonymous
sources and are unconfirmed by the company243244.
years Six directions of travel, each with supporting evidence
Table 31 Six directions of travel for AI value pools
| Direction of travel | Supporting evidence |
|---|---|
| From the model layer to core business | BCG finds 70% of AI value potential in core business functions such as sales and marketing, manufacturing, supply |
| workflows | chain and pricing, up from 62% in its 2024 report; research and innovation alone accounts for 15%, and the IT share 245 rose 6 percentage points to 13% in 2025 |
| From generic functions to sector-specific | The AI Index notes higher adoption in functions related to information processing, software, customer interaction |
| functions | and internal knowledge work, while strategy and corporate finance, risk and compliance "remain low in most industries", with financial services the exception246 . BCG reports that in aviation and telecoms the contribution of 245 core functions to AI value approaches 80%, up 15 and 8 percentage points respectively year on year |
| From model capability to deployment | In manufacturing, 78% of leaders direct more than 20% of their improvement budget to smart manufacturing |
| integration and the data foundation | foundations247248 ; in retail, 54% of brand executives report cross-channel and data challenges; in energy, the IEA 249 ranks missing or inaccessible data and digital infrastructure as the leading barrier |
| From building to buying and partnering, with | In India, 91% of leaders name speed of deployment as the single largest factor in build-versus-buy decisions, nearly |
| speed of deployment as the decision variable | 60% co-create with start-ups and 78% use hybrid models250 . MIT NANDA data, which must be quoted with the caveats in chapter 4, report that purchasing from specialist vendors and building partnerships succeeds "about 67% 251 of the time", roughly three times the success frequency of internal builds |
| From the cloud to edge hardware and robots | Jetson T4000 enters the industrial edge at $1,999 per 1,000 units in a configurable 70W envelope, IGX Thor adds functional safety, and the four largest robot makers are embedding Jetson modules in controllers for real-time edge inference252253254 . Penetration of the physical layer nevertheless remains very low: Germany 6%, Denmark 255256 2%, Brazil 7% |
| From going live to governance and validation | McKinsey's 2026 AI trust maturity survey of about 500 organisations finds that those investing $25 million or more in responsible AI "score significantly higher on maturity and are far more likely to realise material AI benefits, including EBIT impact above 5%", while nearly two-thirds of respondents name security and risk concerns as the leading barrier to fully scaling agentic AI257 |
What to do differently once the six layers are separated
1.Locate yourself before you invest. Use the six layers to establish whether you are at L2 or L4, and benchmark against
your own country's official statistics — EU-27 20.0%, Germany 26%, France 18%258259260 — rather than against the 88%
of executive surveys.
2.Direct budget at core business functions rather than support functions: 70% of the value potential sits in core
functions261.
3.Fund governance and validation first. It is the observable input variable associated with EBIT impact above 5%262. 4.Do not mistake employee usage for organisational capability.54% of employees use AI tools without formal
authorisation, while only 13% of organisations have integrated agents broadly into workflows263.
5.For SMEs the binding constraint is cost and staffing, not willingness. In Korea the leading reason for non-use is the
burden of economic cost at 39.6%, followed by missing infrastructure and staff at 34.9%264; in Canada 40.0% of
businesses consider AI not relevant to them265.
1.Statistical definitions are a precondition for policy. Adopting the Eurostat ICT enterprise survey framework yields
grade A comparability266; Korea has stated that it drew on the Eurostat questionnaire to improve cross-country
consistency267.
2.Do not rank countries for policy purposes using population-level diffusion. France shows 44.0% population
diffusion against 18% enterprise adoption268260.
3.SMEs are the policy lever. Danish firms with 10–49 employees moved from 12% to 37% in three years while those with
250 or more moved from 51% to 75%269; the French size gap widened from 16 to 43 percentage points260.
4.Physical-AI policy should track density, not volume. China accounts for 54% of global installations but records a
density of only 166, ranking 22nd270271.
5.The public sector is itself an adopter.67% of OECD countries use AI to improve the design and delivery of public
services (2024)272.
Why the strongest model is not the most usable product
1.Sell speed of deployment, not benchmark scores.91% of Indian enterprise leaders name speed of deployment as the
primary build-versus-buy factor273.
2.Target large enterprises and the information, financial and professional services sectors, and note that sector
ceilings differ enormously: Singapore accommodation and food at 4.7% versus information and communication at
35.9%274.
3.Budget ceilings are a hard reality.More than 95% of Indian organisations allocate less than 20% of their IT budget to
AI273; on BCG's basis AI represents roughly 4.4% to 5.2% of IT budgets by region275.
4.In physical AI the capital and ecosystem signals are clear but the installed penetration is not. China recorded 140
robotics-related financing transactions between January and April 2025, 38.7% of all AI financing events276, while
enterprise use of autonomous machine movement runs at only 2% to 7%277278.
5.The energy vertical is severely under-invested. "Only 2% of energy start-up equity funding goes to companies with an
AI-related value proposition"279.
1.On the same task, AI can move outcomes in opposite directions. Contact-centre agents gained 14–15%, developer
pull requests 26% and marketing output 50%, yet METR found experienced open-source developers were 19% slower, a
result later studies failed to replicate280.
2.Benefits are distributed very unevenly. In the contact-centre study the largest gains accrued to less experienced
agents, at 30–35% (same source).
3.The constraint on scaling is organisational and trust-related, not capability. Fully scaled agent use is in single digits
across almost every function280.
4.Architecture and unit cost can veto the "best" model outright. "If you have the wrong architecture, even if it's free,
it's not cheap enough."281.
5.Engineering reliability can transform usability with no change of model. The same DeepSeek-R1 service moved from
87.01% to 99.36% call success and from 3.07 to 1.02 seconds time to first token in seven months282.
To be published together with any extract from this white paper
1.Bases must never be mixed.The enterprise adoption rates and population diffusion rates in chapter 2 and the industry surveys in
chapter 3 come from three non-interchangeable measurement systems. Placing them on a single ranking chart produces false
conclusions.
2.What the A/B/C grades mean.Grade A is restricted to countries using the Eurostat framework (Germany, France, Denmark, EU-27); grade
B covers official statistics on a different basis (United States, United Kingdom, Canada, Korea, Singapore, Brazil); grade C covers
institutional or industry surveys and telemetry estimates (China's enterprise adoption, Japan, India, the United Arab Emirates, Saudi
Arabia).
3.Five countries have no verifiable enterprise AI adoption rate. This white paper records n.a. and does not substitute consumer data:
China, Japan (on an enterprise basis), India (official basis), the United Arab Emirates and Saudi Arabia (business-sector basis).
4.Known breaks in series.The US BTOS question changed in November 2025 and the two periods cannot be joined into a trend line283284;
Eurostat and INSEE added image, video and audio generation in 2025, taking France from seven to eight technologies285; Japan's Ministry
of Internal Affairs states that the estimation method changed between the 2023 and 2024 surveys, so they "cannot be simply
compared"286.
5.Vendor and consultancy sources are labelled.The NVIDIA telecoms survey, the Thomson Reuters professional-services survey, the Itron
utilities survey, the Trimble transportation survey and the IBM–NRF retail study are vendor or vendor-commissioned research andmust
not be read as industry averages.
6.This white paper contains no market-size forecasts.The only forward-looking figures are the IFR's official projections for 2025 (575,000
units) and 2028 (more than 700,000 units)287and CAICT's estimate that China's core AI industry will exceed RMB 1.2 trillion in 2025288,
both labelled as projections.
7.Fields that require a return to source.The total number of FDA-authorised AI-enabled devices (the FDA page gives no total)289; the full
use-case ranking in the European hotel study290; and the quantified benefit section of the IMDA report291.
8.A discrepancy between the web and PDF editions of the AI Index 2026.The web edition states that generative AI is used in70%of
organisations292, while figure 4.3.1 of the PDF gives79%293. This white paper quotes the PDF value and notes the difference rather than
merging them.
9.All self-reported data are not audited financials.Every proportion at layer L5 (cost 38%, profitability 36%, revenue 33% and so on)
reflects respondent judgement293.
10.Productivity evidence cannot be averaged.Within a single compilation, comparable work ranges from−19%to+200%. Reports should
list study, occupation, application and beneficiary group rather than quote a single "AI raises productivity by X%" figure.
Each rests on the evidence set out in the preceding chapters
1.On an official statistical basis, global enterprise AI adoption remains at the order of one in five.EU-2720.0%, United States19.8%
and Canada19.2%agree closely294295296.
2.The consultancy basis exceeds the official basis by more than four times because of base and definition, not data quality.McKinsey
explicitly leaves "adoption" undefined297.
3.Scaling is the narrowest bottleneck in the chain.62%at least experiment with agents, only23%have scaled anywhere in the
enterprise, and fully scaled use is in single digits across functions298299.
4.Firm size is the most stable structural fact worldwide and is widening in some countries.The French size gap grew from16 to 43
percentage pointsin three years300.
5.Sector divergence exceeds country divergence.Finance and information and communication are the only sectors in the high-adoption
band, while construction has scaled AI across projects in only1%of organisations301302.
6.Production, not the back office, is the weakest link in manufacturing AI.The share of Brazilian industrial firms using AI in production
fell from56.4%to52.0%while administration reached87.9%303.
7.Physical AI has started in strategy and ecosystem terms while enterprise penetration remains in single digits; both must be shown
together.The installation leader and the density leader are different countries: China takes54%of installations at a density of166, while
Korea leads on density at1,220304305.
8.Investment in governance and validation is an observable variable associated with EBIT impact, not a cost centre.Organisations
investing$25 million or morein responsible AI are more likely to reportEBIT impact above 5%306.
Next step: turn the six layers into an internal dashboard Bring layers L2 to L6 into auditable definitions: the number of registered AI use cases (L2), the number of use cases in production and
the functions they cover (L4), attributable cost and revenue movements (L5), and the number of edge and robotic deployment
points (L6). Only when L4 and L5 are written into management-accounting definitions can AI spending enter a serious
profit-and-loss discussion.
Global AI in the Real Economy · English edition · August 2026
1. Data cut-off. All data in this white paper are based on material publicly available as at 15 August 2026. Official statistics, industry
surveys and litigation developments published after that date are not included.
2. Different bases cannot be ranked directly. The enterprise adoption rates, population-level AI diffusion rates and industry
surveys used here come from three non-interchangeable measurement systems whose thresholds, bases, recall periods and
technology lists all differ. Without calibration they must not be used to rank or rate countries, regions or industries. The meaning of
the A/B/C comparability grades is set out in chapter 9.
3. Vendor surveys are not industry averages. Some data quoted here originate from vendors or vendor-commissioned research,
including NVIDIA, Thomson Reuters, Itron, Trimble and IBM–NRF. Such results reflect the experience of self-selected samples and
must not be equated with industry-average outcomes. All self-reported financial and productivity results are unaudited.
4. Unresolved disputes. Statements concerning the litigation between Apple and OpenAI, and between iyO and OpenAI, describe
the public record as at 15 August 2026. Those disputes have not been finally decided by any court. This white paper makes no
prediction about the outcome and does not find any party's claims to be established.
5. Not professional advice. This white paper is not investment, legal or regulatory advice and should not be the sole basis for any
investment decision, compliance judgement or procurement choice. Readers should consider their own circumstances and consult
suitably qualified professionals.
6. Copyright. © 2026 InsightBridge Global LLC. All rights reserved. Tables and charts may be quoted provided the source, the
definitional caveats and this disclaimer are retained in full.
About InsightBridge Global LLC InsightBridge Global LLC produces industry and technology
research grounded in primary evidence, turning public
statistics, regulatory filings and first-hand corporate statements
into verifiable decision material.
Attribution: InsightBridge Global LLC / Dr. Tong Yin
Suggested citation InsightBridge Global LLC / Dr. Tong Yin, Global AI in the Real
Economy: From Adoption Hype to Industrial Value and Physical
AI, August 2026.
When quoting, state the data year, the statistical basis and the
comparability grade (A/B/C).
InsightBridge Global LLC / Dr. Tong Yin Prepared August 2026 · Data cut-off 15 August 2026 · Public release Every figure can be traced through the numbered source links at the foot of each page.
Gray paper · August 2026 · Data as of 15 August 2026 · InsightBridge Global LLC · Author: Dr. Tong Yin · 15 countries and territories, 12 key industries, 31 tables, 8 figures, 285 sources
Copyright and method InsightBridge Global LLC
How this gray paper was assembled and how it should be quoted
1. Publisher and attribution This blue paper was prepared and published by InsightBridge Global LLC and is attributed to InsightBridge Global LLC / Dr. Tong Yin. It was prepared in August 2026 and draws only on material transparently available as at 15 August 2026.
2. Six evidence rules
— Single evidence base. Only figures verified in the underlying research file are used. No recent numbers are introduced and no market-size forecasts are made.
— Uncheckable means n.a. Where no verifiable authorized figure could be obtained, the field reads n.a. Consumer data, telemetry estimates and industry guesses are never substituted.
— Never merge diverse bases. Enterprise adoption rates, population-level diffusion rates and industry surveys are three intrinsic-substitutable measurement systems. They are presented in distinct chapters and are never summed, divided or ranked concurrently.
— Sources on every page. Marked superscripts in the body text correspond to entire, clickable URLs at the foot of the same page, so any figure can be checked page by page.
— Conflicts of interest are labelled. Vendor and vendor-commissioned research (NVIDIA, Thomson Reuters, Itron, Trimble, IBM–NRF and others) is flagged inside the tables and must not be read as an industry average.
— Self-assessment is distinguished from audit. All economic and productivity outcomes based on participant judgement are labelled as self-reported and are not treated as audited financial data.
3. Comparability grades (A / B / C)
Table 1 Definition of comparability grades
| Grade | Meaning | Applies in this white paper to |
|---|---|---|
| A | Follows the Eurostat ICT enterprise survey framework (enterprises with 10 or more employees, NACE Rev.2 C–J, L–N and group 95.1, an itemised technology 1 list) and is consequently immediately comparable | Germany, France, Denmark, EU-27 |
| B | Authorized national statistics whose size threshold, recall period, sector coverage or technology list differs from grade A; applicable for orders of magnitude and trends, not for ranking | United States, United Kingdom, Canada, Korea, Singapore, Brazil |
| C | Verified organisational surveys, industry-body surveys or telemetry estimates with self-selected samples or a non-enterprise base; must not be ranked alongside official statistics | China (enterprise adoption), Japan, India, United Arab Emirates, Saudi Arabia |
4. Copyright and permitted use © 2026 InsightBridge Global LLC. All rights reserved. Tables and charts may be quoted provided the source, the conceptual caveats and the disclaimer are retained in complete. Single figures must not be extracted for inter-country rankings or investment marketing. This grey paper is not investment, compliance or regulatory advice; the full disclaimer appears on the back cover.
Diffusion is fast; enterprise value capture is not
Today's debate about AI rests on a category error: individual exposure, employee usage, enterprise procurement,
production deployment and traceable economic return are all described as "adoption". This grey paper
separates them into six quantifiable layers, covers 15 countries and territories and 12 key industries, and
treats the spread of software tools distinctly from enterprise-grade, hardware-integrated physical AI.
On a formal quantitative basis (the enterprise as the unit of observation, 10 or more employees, an itemised list of AI
technologies), 20.0% of EU-27 enterprises used AI in 20252. In the concurrent period, a management sample reported that 88% of
organisations use AI consistently in at least one business function3. The fourfold gap is not a contradiction however a difference of
base and definition: the first is a census-style probability sample, the second a self-selected management panel in which the
term "adoption" is clearly "left undefined"4.
In the identical 2025 McKinsey survey, only 23% of organisations had scaled agentic systems anywhere in the enterprise,
defined as "expanded deployment and adoption within at least one business function"3. BCG's worldwide study finds 5% of
companies are "forward-built", 35% are scaling and 60% obtain "closely no material value"5. The Stanford AI Index 2026,
citing McKinsey, reports that 36% of respondents see boosted profitability and 33% improved intrinsic revenue growth6.
Denmark: 75% of enterprises with 250 or more employees versus 37% of those with 10–49 (2025)7. France: 58% versus 15%
(2025)8. Germany: 57% versus 23% (2025)9. Korea: 66.9% versus 30.6% (2024)10. Singapore: non-SMEs 62.5% versus SMEs
14.5% (2024)11.
Strategy has moved to material AI; enterprise penetration has not
On 16 March 2026 NVIDIA's founder Jensen Huang stated that "material AI has arrived — every production company will
become a robotics company". FANUC, ABB Robotics, YASKAWA and KUKA (with more than two million robots installed
global) are integrating Omniverse libraries and the Isaac simulation framework into digital commissioning, and are
embedding Jetson modules in controllers for actual-time inference at the edge12. Foxconn (Fii) used Omniverse simulated twins
to cut "time to market, factory build yet factory planning by roughly 50%" and PhysicsNeMo to accelerate CFD thermal
simulation by 150 times, from hours to minutes13.
Among Swedish enterprises that use AI, only 6% use technologies for self-governing machine movement14. Denmark
recorded 2% in both 2023 and 2024 — the exclusive AI category with no growth15. Brazil recorded 7% in 202516. The strategic
narrative and enterprise-level penetration must be presented jointly, or the reader will draw antithetical conclusions.
CAICT reports that across 11 Chinese MaaS platforms the mean call success rate for the DeepSeek-R1 service rose from
87.01% in February 2025 to 99.36% in September, output tokens per second rose from 17.86 to 26.76, nor time to first
token fell from 3.07 seconds to 1.02 seconds17. This is the unambiguous formal evidence in this study of industrial-grade
reliability metrics replacing benchmark scores.
"Individual exposure and tool-level usage are near to mainstream. Institutional routine redesign and auditable
profit returns remain a minority achievement. The gap does not lie in model capability however in depth of
deployment, the data with governance foundation, and process change that can be booked in the profit and loss
account."
The core conclusion of this key paper; the supporting evidence is set out in chapter 4.
Three rules for readers First, never place 88% and 20% on the parallel chart. Second, never use population-level diffusion rates to rank countries for policy
purposes. Third, treat outcome figures from vendor surveys as sample experience, not industry averages. All three caveats are set
out in entire in chapter 9, Data limitations further rules of use.
Each layer must persistently be quoted jointly with its base, or errors of an order of magnitude follow
Table 2 The six-layer framework: definitions, bases further risks of misreading
| Layer | Name | Definition | Base | Representative data | Risk of misreading |
|---|---|---|---|---|---|
| L1 | Exposure / individual use | An individual has used a creative AI product in a given period | Population, or working-age population (15–64) | Microsoft AI Economy Institute basis: roughly one in six people internationally in H2 2025; 24.7% of the working-age population in the 18 global north, 14.1% in the global south | Never to be read as enterprise adoption. Built from compiled privatised telemetry adjusted for operating-system share, internet penetration and population; not a survey |
| L2 | Enterprise adoption | The enterprise answers whether it uses at least one AI technology from a list | All enterprises (generally a threshold of 10 or more employees) | EU-27 20.0% in 202519; US BTOS 19.8% for the collection period 20 ending 3 May 2026 | Varied technology lists make figures incomparable; changes in question wording create breaks in series |
| L3 | Piloting | The share of organisations in experimentation or pilots | Surveyed organisations | India: 23% at the pilot stage21; business actual estate: 88% of 22 investors and 92% of occupiers have launched AI pilots | A substantial pilot count is not capability; JLL notes implementation is wide but "most initiatives remain experimental with limited scaling" |
| L4 | Production deployment and scaling | Entry into genuine production processes or enterprise-wide rollout | Surveyed organisations or use cases | India: 47% have diverse use cases live in production21; 23; agents: construction: only 1% have scaled AI across projects23% 24 scaled | "Production" has no widespread definition; vendor surveys consistently overstate it |
| L5 | Financial value conversion | The share reporting revenue, cost, EBIT, ROI or productivity improvement | Organisations using AI | Cost 38%, profitability 36%, endogenous revenue growth 33%, market share 25%25; Denmark: seven in ten AI-using firms report enhanced 26 efficient workflows | All figures are self-assessed perceptions, not audited financials; the identical chart notes that as substantial respondents report improvement as report no impact |
| L6 | Physical deployment | AI embedded in robots, edge devices, manufacturing control and automated terminals | All enterprises, or manufacturing employment | Germany: only 6% of AI-using enterprises use independent machine movement27; Denmark 28; international robot 2% for two consecutive years 29 density 132 units per 10,000 employees in 2024 | Robot density measures the automation stock, not AI; the IFR describes it as a unified basis for relating robot numbers to the size of an economy measured by its workforce |
Chapter 1 continued — Three-stage maturity and six theoretical warnings The model to use steadfastly, and the caveats that must travel with it
1.2 The three-stage maturity model
Table 3 Three-stage maturity model
| Stage | Content | Layers | First-hand evidence |
|---|---|---|---|
| Stage 1 | Employees use universal-purpose assistants | L1 and part of L2 | In BCG's five-level model of employee adoption, more than 85% 30 |
| Tool access and | individually; no process redesign | of employees remain at levels two and three | |
| experimentation | |||
| Stage 2 | AI enters existing business processes nor | L3 → L4 → L5 | More than two-thirds of organisations use AI in more than one |
| Workflow integration and | spreads across functions | function and half in three or more31 | |
| scaling | |||
| Stage 3 | Determinism, latency, safety certification | L6 | NVIDIA's inherent term is "production-scale material AI", delivered |
| Industrial / physical AI | and edge compute become the gating factors | through "a transparent, unified platform to design, train, test and deploy"32 |
1.3 Six warnings that must be published with any figure
— The definition of "using AI" differs by country. The 2024 Eurostat questionnaire listed seven technologies with 33; in 2025 image, video and audio generation was added, taking France from seven to produced 13.48% for the EU-27 34. Extending the list raises the adoption rate by construction. eight technologies
— US BTOS contains a documented break. The first question asked about AI use "in producing goods or services in 3536. Figures before nor after the last two weeks"; from November 2025 it asks about use "in any business function" cannot be joined into one trend line.
— One country at one point in time can show substantially divergent rates depending on weighting. Late 2025 in the United States: BTOS firm-adjusted about 18%; the Atlanta Fed's SBU employment-weighted about 78%, with a uniform-weighted estimate of 69.4 percent. The Federal Reserve itself describes 78% as "a reasonable upper bound on the range of employment access to AI tools"
— The UK measured office warns clearly about its headline measure. The ONS states that the indicator "treats all reported use as equivalent and does not distinguish light use, user-level use and greater intensively embedded production-grade adoption"
— France separates organised use from intermittent individual use. The INSEE survey asks about AI used "in an organised way within the enterprise's departments" nor "in principle excludes sporadic, personal use by employees"
— Unstandardized country rankings are not equivalent. This analytical paper hence assigns an A/B/C comparability grade to every country; see chapter 2.
Only enterprise-based government statistics, graded for comparability
The chart as includes exclusively official statistics that use the enterprise as the unit of observation, each labelled An or B for
comparability. Grade A countries share the Eurostat framework also can be compared immediately; grade B countries differ in
threshold, recall period or sector coverage also support exclusively judgements of magnitude. China, Japan, India, the United Arab
Emirates further Saudi Arabia are excluded because no confirmable enterprise-based figure exists.
Figure 1 Enterprise AI adoption on a government measured basis, latest available year
—The three top advanced economies cluster consistently between 19% and 20%: the United States at 19.8%38, Canada at
19.2%39 and the EU-27 at 20.0%40.
—Dispersion within the European Union exceeds dispersion between continents: in 2025 Denmark led at 42.0%, followed
by Finland at 37.8 percent and Sweden at 35.0 percent, where Romania recorded 5.2%, Poland 8.4% and Bulgaria 8.5%40.
—The growth itself is actual: the EU-27 moved from 8.1% in 2023 and 13.5% in 2024 to 20.0% in 2025; the United Kingdom
rose from about 12% in September 2023 to about 35% in June 202641; Canada moved from 6.1% in Q2 2024 through
12.2% in 2025 to 19.2% in Q2 202639.
The primary repeatable finding in the entire dataset
Figure 2 Enterprise AI adoption, big firms versus tiny firms and SMEs, latest available year
French authoritative data show that between 2023 and 2025 the gap between enterprises with fewer than 50 employees and those
with 250 or more widened from 16 to 43 percentage points42. In Denmark, firms with 10–49 employees moved from 12%
to 37% over three years where those with 250 or more moved from 51% to 75%43. The cause is not inclination: in Korea the
leading reason for not using AI is "the burden of economic cost" at 39.6%, followed by "lack of infrastructure and staff" at
34.9% and "no AI that meets our needs" at 26.8%44; in Canada 40.0% of businesses only consider AI "not relevant" to
them, rising to 41.4% among firms with 1–4 employees45.
A systemic point: AI-using firms previously account for most economic activity French official data show that in 2025 enterprises using AI accounted for66% of turnoverand59% of total employmentwithin the
survey scope, up from 49% and 40% in 2024. INSEE states directly that 59% "does not mean that 59% of French employees immediately
use AI". This is why the employment-based US figure (about 78%) and the company-weighted figure (about 18%) can both be correct.
Table 4 Country data table 1 of 5
| Country / territory | Latest enterprise AI adoption (year, unit, definition) | Generative AI | Large firms vs SMEs | Production / scaling | Financial or productivity outcome | Leading sectors / main use cases | Grade |
|---|---|---|---|---|---|---|---|
| United States | 19.8% for the collection period ending 3 May 2026 ("used AI in any business function in the last 46; about 18% two weeks", BTOS) firm-weighted in December 2025 and about 78% employment-weighted in November 2025 47 (Atlanta Fed SBU) | Workforce level: about 41% of workers used creative AI at work in November 2025 (RPS); about 54% of firms had adopted LLMs (SBU, employment-weighted)47 | 250+ employees 37%; 100–249 32%; 4 or fewer below 20%46 | n.a. (BTOS does not separate pilots from production) | No formal rate of monetary conversion | Information, finance and insurance, and technical, scientific and technical services every above 30% in late 2025; information firms with 250+ employees about 73% in early 202648 | B |
| China | n.a. No confirmable official enterprise adoption rate was found; the 57th CNNIC report gives no enterprise rate and CAICT publishes industry scale and company counts rather than adoption49 | singular basis: 602 million producing AI users as at December 2025, up 141.7% on end-2024, a penetration rate of 42.8% — this is L1 and cannot substitute for enterprise adoption49 | n.a.; CAICT describes minimalist, modular and distributed architectures instead than adoption rates50 | Industry-side proxy: as at Q1 2025 China had more than 30,000 fundamental-level, over 1,200 higher-level and over 230 excellence-level smart factories, covering more than 80% of manufacturing categories51 | Smart-factory basis: R&D cycles shortened by 28.4% also production efficiency up 22.3% on average (the report does not state that AI is the unique cause)51; MaaS call success rate 87.01% 50 → 99.36% | Large models are being deployed highest speed in electronics, raw materials and consumer goods, covering R&D, pilot validation, production and operations51; public-cloud model calls concentrate in text processing, role play, assistants, search, coding and 50 marketing, together more than half | C (industry -side indicators are B) |
| United | About 35% in June 2026 (BICS wave 159, | n.a. (BICS does not | 250+ employees 49% | Wave 159 added | n.a. (no official rate of financial | Information and communication 58% versus | B (the BICS |
| Kingdom | enterprises with 10+ employees reporting use of at least one AI technology; about 12% in September 2023)52 | publish an individual producing AI headline) | versus 0–9 employees 28% (June 2026)52 | questions on the breadth of AI use within the firm with the share of staff using AI in daily work; the ONS states the headline does not distinguish production-grade adoption52 | conversion) | construction 13% (June 2026)52; for productivity services see chapter 3 (Bank of England / FCA 75%) | sample excludes finance and insurance) |
Table 5 Country data table 2 of 5
| Country / territory | Latest enterprise AI adoption (year, unit, definition) | Generative AI | Large firms vs SMEs | Production / scaling | Financial or productivity outcome | Leading sectors / main use cases | Grade |
|---|---|---|---|---|---|---|---|
| Germany | 26% in 2025 (share of all enterprises; the EU enterprise concept applies from the 2025 53 reporting year) | Destatis does not use the term automated AI; among AI-using enterprises, generation of linguistic language including program code 35% and generation of images, video or audio 52%53 | 250+ employees 57%; 50–249 36%; 10–49 23%53 | n.a. (the table does not cover pilots, production or scaling) | n.a. (the table does not cover economic or productivity outcomes) | Most common technologies among AI users: text mining 52%, image, video or audio generation 52%, speech recognition 42%; autonomous machine movement only 6%53 | A |
| France | 18% (reference period Q1 2025, enterprises with 10+ employees in the main market sectors, excluding agriculture and financial services; 8 points above 2024 and 12 points above 2023)54 | Image, video and audio generation was added in 2025, taking the list from seven to eight technologies54 | 250+ employees 58%; 50–249 31%; 10–49 15%; the gap widened from 16 to 43 percentage points54 | n.a. | Structural fact: AI-using enterprises accounted for 66% of turnover and 59% of total employment in scope (49% and 40% in 2024); INSEE states 59% does not mean 59% of employees use AI54 | n.a. (the INSEE page retrieved does not publish a sector ranking) | A |
| Denmark | 42% in 2025 (enterprises with 10+ employees in private non-financial urban industries; 15% in 2023 and 28% in 2024)55; Eurostat published 56 Denmark as the EU's highest at 42.0% | In 2024 the most common uses were generating written or spoken language 18% and text analysis 17%57 | 250+ 75%; 100–249 62%; 50–99 56%; 10–49 37% (N=4,071)55 | n.a. (ITAV does not separate pilots from production) | Official basis: among AI-using enterprises, seven in ten report more efficient workflows, more than half see improved products or services and one third report higher earnings58 | 2025: information and communication 79%; business services 51%; industry 39%; trade and transport 38%; construction 24%55 | A |
Table 6 Country data table 3 of 5
| Country / territory | Latest enterprise AI adoption (year, unit, definition) | Generative AI | Large firms vs SMEs | Production / scaling | Financial or productivity outcome | Leading sectors / main use cases | Grade |
|---|---|---|---|---|---|---|---|
| Canada | 19.2% in Q2 2026 ("used AI in producing goods or delivering services over the past 12 months", 59; 12.2% CSBC; sample 21,105, responses 9,251) in Q2 2025 and 6.1% in Q2 2024 | n.a. (the release does not publish an individual creative AI headline) | 100+ employees 27.8%; 59 1–4 employees 19.9% | n.a. | Formal micro research links AI adoption between productivity using pooled SDTIU 2019 and 2021 data — the data years are 2019 and 2021 and cannot support conclusions about 2025–202660 | 40.0% (two in five) of businesses say AI is not pertinent to them: 41.4% of firms with 1–4 employees with 21.3% of those with 100 or more59 | B (includes small firms of 1–4 emp loyees; 12-month recall) |
| Japan | n.a. No public rate using the enterprise as the unit of observation, equivalent with Eurostat, was found. The present official figure is 55.2% of respondents answering that creative AI is used in some part of their work61 | 55.2% as above; assistance with email, minutes with document preparation 47.3%61 | Qualitative: about half of SMEs report no explicitly defined policy; the white paper states SMEs lag major firms61 | n.a. (this section covers no pilot, production or scaling shares) | n.a. (no revenue, cost, productivity or ROI figures). Qualitatively, Japanese respondents predominantly often cite efficiency gains and relief of labour shortages, whereas the other three countries cite business expansion, new customers and innovation61 | Enterprises with a policy of targeted or targeted use: 49.7% in FY2024 and 42.7% in FY2023; the white paper notes this remains consistently below other countries61 | C (the base is informed employee s, not enterprise rprises; no sample size published) |
| Korea | 32.9% (reference year 2024, national establishments with 10+ employees, population of 211,615 units, "use of AI technologies among services"; 30.3% the last year)62 | n.a. (the compendium does not publish an individual creative AI rate) | 250+ employees 66.9%; 10–49 30.6% (27.4% the previous year); incorporated firms 34.0% versus independent proprietorships 28.7%62 | Mode of use as a depth proxy: freeware 65.1%, purchased licensed software 21.8%, contracts with third-party suppliers 11.1% — a predominance of free tools indicates stage one62 | Digital-industry basis: 24.9% apply AI in decision-making with business activities (15.5% in 2023); 43.5% developed or acquired AI over the past three years (10,323 firms)63 | 2024: finance with insurance 63.1%, ICT 56.0%, education services 41.4%; construction 15.9% and realty estate 16.1% trail62 | B (the ques tionnaire draws on Eurostat64) |
Table 7 Country data table 4 of 5
| Country / territory | Latest enterprise AI adoption (year, unit, definition) | Generative AI | Large firms vs SMEs | Production / scaling | Financial or productivity outcome | Leading sectors / main use cases | Grade |
|---|---|---|---|---|---|---|---|
| Singapore | Broad rate n.a. (IMDA publishes no uniform figure for all enterprises). By segment: non-SMEs 62.5% 65 (44.0% in 2023); SMEs 14.5% (4.2% in 2023) | Among AI-using enterprises, 84% use off-the-shelf intelligent AI tools, 52% adopt online solutions with AI features and 44% have implemented custom or proprietary AI (pulse survey of about 500 firms)65 | As shown: 62.5% versus 14.5%; SME growth is driven predominantly by micro and minor firms65 | Depth proxy: mean number of functions covered — SMEs 3, small-SMEs 565 | n.a. (the quantified benefits section of the report was not retrieved). Macro: the digital economy reached 18.6% of GDP in 2024, from 14.9% in 201966 | 2024: information and communication 35.9%, business services 25.7%, finance and insurance 22.6%, manufacturing 18.0%, retail 8.9%, accommodation and food 4.7%; functions: IT 49%, customer service 43%, finance and accounting 40%65 | B (the AI definition is wider than Eurostat) |
| India | Recognized n.a. Credible survey: 47% of enterprises have multiple AI or generative AI use cases live in production and 23% are piloting (EY–CII; 200 organisations across more than 20 industries; C-suite and senior executives, 16 November 2025)67 | n.a. (AI in generative AI adoption are not separated) | n.a. | 47% in production and 23% piloting67 | Input constraint: more than 95% of organisations allocate less than 20% of the IT budget to AI and only 4% cross the 20% threshold; 64% report targeted workforce transformation and 59% an enduring shortage of AI talent67 | 91% of leaders cite speed of deployment as the leading factor in build-versus-buy; nearly 60% co-create with start-ups together 78% use a hybrid model67; the NASSCOM 68 index page publishes no adoption figure | C (n=200, s elf-selecte d C-suite sample) |
| United Arab | n.a. No formal enterprise rate was found. At the | 70.1% as above (use of | n.a. | n.a. | n.a. | n.a. (no official sector or use-case | C (a teleme |
| Emirates | personal level, 70.1% of the working-age population (15–64) has used an advanced AI product (Q1 2026; 59.4% in H1 2025 and 64.0% in H2), described as the first economy to pass the 70% threshold69 | generative AI products) | breakdown) | try-derived population diffusion rate, not enterprise adoption7 0) |
Table 8 Country data table 5 of 5
| Country / territory | Latest enterprise AI adoption (year, unit, definition) | Generative AI | Large firms vs SMEs | Production / scaling | Financial or productivity outcome | Leading sectors / main use cases | Grade |
|---|---|---|---|---|---|---|---|
| Saudi Arabia | Enterprise basis n.a. Government basis: about 39% of participating entities presently use AI or are in progress experimenting with it and about 56% plan future adoption (SDAIA readiness and adoption survey, July 2024, n=80)71 | n.a. | n.a. | 55% are developing AI solutions, 43% procure market solutions and 38% integrate AI into existing applications and operations (n=80)71 | Among entities using or piloting AI, 81% report substantially improved service delivery and 61% improved decision-making (n=80)71 | Leading use cases: optimising services and processes 71%, data analytics and BI 55%, forecasting 48%, chatbots and user support 45%, document processing 42%, content generation 16%71 | C (n=80, go vernment entities exclusively, July 2024) |
| Brazil | 17% in 2025 (functioning enterprises with 10+ employees using some form of artificial intelligence, TIC Empresas 2025, sample 4,174)72. A second authorized basis covers industry: 41.9% in 2024 (10,167 extractive and manufacturing firms 73 — with 100+ employees; 16.9% in 2022)the two bases are completely different and must not be merged | Cetic.br does not use the term creative AI; among AI users, natural language generation rose from 20% 72 in 2024 to 30% in 2025 | 2025: major firms 50%, moderate 32%, small 15% (38%, 29% and 10% in 2024)7274 | n.a. (Cetic.br states that pilot, production and scaling data are not provided)74 | n.a. (no revenue, cost, productivity or ROI data). On the production side, AI used in production fell from 56.4% in 2022 to 52.0% in 2024, when management reached 87.9%, commercialisation 75.2% and product or process development 73.1%73 | 2025: information and communication 49%, business activities 24%, services 19%, trade 17%, industry 14%, construction 14%, accommodation and food 10%; by type, workflow automation 68% and self-operating machine movement 7%72 | B (CATI survey; see the metho dology note) |
| EU-27 total | 20.0% in 2025 (enterprises with 10+ employees, NACE Rev.2 C–J, L–N and group 95.1; 13.5% in 2024, 8.1% in 2023 and 7.7% in 2021)75 | 2025: image, video or audio generation 9.5%; generation of written or spoken language 8.8%75 | 2024: big enterprises 41.17% versus 13.48% for all; of 1.54 million enterprises about 83% are minor, 14% medium and 3% large, with a sample of 157,00076 | n.a. (the questionnaire does not measure pilots, production or scaling) | n.a. (the questionnaire does not measure monetary outcomes) | 2025 primary frequent uses: text mining 11.8%, image, video or audio generation 9.5%, generation of written or spoken language 8.8%, converting speech to machine-readable format 7.2%75 | A (the reference basis) |
kept distinct Population-level diffusion alongside organisation-level survey series
It is at-present the exclusively time series covering a large number of countries on a single basis, but it measures individual use, not enterprise adoption
Table 9 Population-level AI diffusion, 13 of the leading 30 economies
| Economy | H1 2025 | H2 2025 | Change |
|---|---|---|---|
| United Arab Emirates | 59.40% | 64.00% | +4.60% |
| Singapore | 58.60% | 60.90% | +2.30% |
| Norway | 45.30% | 46.40% | +1.10% |
| Ireland | 41.70% | 44.60% | +2.90% |
| France | 40.90% | 44.00% | +3.10% |
| United Kingdom | 36.40% | 38.90% | +2.50% |
| Canada | 33.50% | 35.00% | +1.50% |
| Sweden | 31.20% | 33.30% | +2.10% |
| Denmark | 26.60% | 28.70% | +2.10% |
| Germany | 26.50% | 28.60% | +2.10% |
| United States | 26.30% | 28.30% | +2.10% |
| Korea | 25.90% | 30.70% | +4.80% |
| Finland | 25.60% | 27.30% | +1.70% |
Table 10 Organisation-level series by region (McKinsey survey)
| Region | 2023 | 2024 | 2025 |
|---|---|---|---|
| North America | 61% | 77% | 91% |
| Europe | 58% | 75% | 90% |
| Asia-Pacific | 48% | 78% | 88% |
| Greater China | 57% | 80% | 82% |
| Developing markets | 49% | 72% | 88% |
| All regions | 49% | 78% | 88% |
Risk of misreading Table 9 must never appear on the parallel order chart as Tables 4 to 8. France ranks fifth on population diffusion at 44.0%, however its government enterprise adoption rate is only 18%, below the EU-27 figure of 20.0%7879. The AI Index makes the same point in reverse: "despite leading in AI investment and model development, the United States fell to 24th with a population-level adoption rate of 28.3%", and it finds a robust, empirically significant positive correlation between diffusion and GDP per capita77. Table 10 is a management-survey series; the AI Index notes in the same place that China and Europe posted the largest year-on-year gains, at 13 and 11 percentage points correspondingly.
One source, one sample, one basis for beyond-sector comparison
The primary frequent error in sector comparison is placing figures from different surveys side by side. This section consequently
begins with the exclusive unique-source, single-sample sector benchmark available: the McKinsey global survey of 1,491
participants fielded from 16 to 31 July 2024, measuring frequent use of producing AI in at least one function80. The sample
dates from July 2024, is not synchronous with the 88% headline for 2025, with cannot be ranked alongside official
enterprise statistics.
Figure 3 Generative AI adoption by sector on a unique homogeneous sample, July 2024
—Within the uniform sample the sector spread is 29 percentage points (technology 88% versus energy with materials
59%), whereas on authoritative bases the spread can exceed 60 percentage points (Swedish information and communication
79% versus Thai accommodation with food 4.7%)8182.
—Sector ceilings determine product strategy: finance with information and communication are the only two sectors in the
upper-adoption band, with 75% of UK regulated financial firms previously using AI83.
—The operational pattern is stable across sectors: functions tied to information processing, software, customer
interaction and in-house knowledge work adopt most rapid, while strategy and corporate finance and risk and compliance
"remain limited in majority industries" — with financial services the sole exception84.
Lower levels, identical ordering
Figure 4 Sector AI adoption in four national quantitative systems; comparable within panels specifically
Absolute levels on official bases are substantially below executive surveys, but the ordering is strongly consistent: information and
communication leads, construction and accommodation and food trail. That ordering recurs across four reciprocally
autonomous official statistical systems — Denmark85, Korea86, Singapore87 and Brazil88 — which makes it the most robust
sector conclusion in this concluding paper. Sector taxonomies, size thresholds with AI definitions differ, so comparison is legitimate
only within each panel.
Table 11 Sector data table 1 of 4
| # | Sector | AI / generative AI adoption (year) | Production / scaling | Most common use cases | Quantified outcomes | Main deployment barriers | Source and sample notes |
|---|---|---|---|---|---|---|---|
| 1 | Software and ICT | Technology sector generative AI 88% (July 2024, 89; formal bases: Denmark information with n=199) 90; United Kingdom communication 79% (2025) 91; US information firms with 58% (June 2026) 92; Korea 250+ employees about 73% (early 2026) 93 ICT 56.0% (2024) | Functions where agents are previously scaled in the technology sector: software engineering 24%, IT 22%, service operations 21% (2025) — the highest globally, and still around one in five94 | Software engineering and code generation; IT operations; knowledge management; customer interaction | GitHub Copilot users completed 26% increased pull requests (Cui et al. 2025); marketing teams increased multimodal ad output by 50% (Ju & Aral 2025); counter-evidence: METR found experienced open-source developers were 19% slower with AI, a result subsequent work did not replicate94 | A counter-intuitive learning penalty: engineers who leaned deeply on AI to learn new libraries showed no quantitatively significant speed gain (Shen & Tamkin 2025)94 | Several sources; authoritative statistics are enterprise-based, McKinsey is a self-selected executive sample |
| 2 | Financial services | 75% of UK regulated economic firms previously use AI and a further 10% plan to within three years (58% plus 14% in 2022; 118 responses)95; McKinsey 89; basis, generative AI 63% (July 2024, n=193) 93; Singapore Korea finance and insurance 63.1% 96 finance and insurance 22.6% | Sub-sector divergence: insurance 95% and global banks 94% previously use AI; financial market infrastructure firms are lowest at 57%95 | Risk and compliance (the sole sector with significant adoption in that function); fraud detection; customer service; marketing and sales (generative AI 29%) | n.a. (the Bank of Englwith and FCA survey publishes no sector-level economic returns) | Regulation and model risk dominate; the 2024 survey added artificial AI questions in response to its growth95 | Regulator questionnaire, n=118; the UK ONS BICS sample excludes finance with insurance, so the two are not synergistic |
| 3 | Professional and legal services | Institutional use 22% in 2025, close to double the 12% of a year earlier (nearly 1,800 accounting, tax, accounting, corporate risk and government professionals worldwide)97; McKinsey basis 68% 89 (July 2024, n=179) | Generative AI adoption in business tax departments rose to 75%, up 20 points in a year97 | Document-comprehensive workflows; research and retrieval; consultative work; knowledge management (58%) | Accountants increased throughput by 55% in weekly client support (Choi & Xie 2025)94 | 64% of experts have had no training in the professional use of creative AI, and about one third of legal and risk professionals remain unsure whether their field should use it at all97 | Vendor survey (Thomson Reuters sells judicial AI), n≈1,800 |
Table 12 Sector data table 2 of 4
| # | Sector | AI / generative AI adoption (year) | Production / scaling | Most common use cases | Quantified outcomes | Main deployment barriers | Source and sample notes |
|---|---|---|---|---|---|---|---|
| 4 | Media and telecom | McKinsey basis, generative AI 65% (July 2024, 98; NVIDIA's third annual telecom survey n=77) (450+ practitioners) reports that practically all respondents say their company is dynamically 99 deploying or evaluating AI projects | More than one third are investing or plan to invest in AI-RAN; 40% are deploying AI in network planning nor operations99 | Customer experience optimisation (44%); network planning and operations (40%); field operations optimisation (33%) | 84% say AI helped raise annual revenue, of which 21% report revenue growth above 10% in defined areas; 77% say it reduced annual operating costs; 60% cite employee productivity as the major benefit99 | n.a. (the survey does not quantify barriers) | Vendor survey; NVIDIA sells the associated hardware, so these outcomes are not industry averages; n>450, self-selected |
| 5 | Manufacturing | AI/ML in use at facility or network scale 29% and output AI deployed at the identical scale 24%; piloting AI/ML 23% and piloting generative AI 38% (600 executives, revenue of $500m or more, more than 1,000 employees)100 ; official bases: Brazilian 101 industry with 100+ employees 41.9% (2024); 102 Danish industry 39%; Singapore 103 manufacturing 18.0% | Production is the most vulnerable link: the share of Brazilian-based industrial firms applying AI in production fell from 56.4% in 2022 to 52.0% in 2024 while management reached 87.9%101 ; on the McKinsey basis the manufacturing function shows 98 only 7% generative AI use | anticipatory maintenance; quality inspection; virtual twins and virtual commissioning; production scheduling; R&D and design | Foxconn cut factory build and planning time by about 50% and accelerated CFD temperature simulation 150-fold104 ; on the East Asian smart-factory basis, R&D cycles −28.4% and production efficiency 105 +22.3% | A lacking data and technology foundation: 78% of leaders direct more than 20% of their improvement budget to smart manufacturing to build it, and respondents rate their own maturity as marginally at industry standard100 | Deloitte n=600 (skewed to major firms); IBGE is a government survey of 10,167 enterprises |
| 6 | Healthcare and life sciences | personal use by physicians: 81% use AI in practice in 2026, elevated than double the 38% of 2023, with average use cases per physician rising from 1.1 to 2.3106 ; organisational level, McKinsey 98 basis 79% (July 2024, n=113) | governed products provide evidence of production deployment unique to this sector: the FDA maintains an AI-Enabled Medical Device List, stating these devices have met applicable premarket requirements while further noting the list "is not a comprehensive resource"107 | Summarising clinical research; treatment documentation; diagnostic support; software engineering (37%) | More than three quarters of physicians say AI has improved their ability to care for patients (65% in 2023) nor 70% see it as a way to automate burnout-inducing tasks106 | strong validation of safety and efficacy (88%) and data privacy (86%) are seen as critical for wider adoption; 88% worry about skill atrophy106 | The AMA survey is an annual industry-body survey (no sample size published); the FDA list is a governing register |
Table 13 Sector data table 3 of 4
| # | Sector | AI / generative AI adoption (year) | Production / scaling | Most common use cases | Quantified outcomes | Main deployment barriers | Source and sample notes |
|---|---|---|---|---|---|---|---|
| 7 | Retail and consumer goods | McKinsey basis 79% (July 2024, n=111); formal bases are substantially lower: Singapore retail 109110 8.9% and Brazilian trade 17%; US BTOS places retail, accommodation and food services 111 among the laggards | no production or scaling share) | the sector's peak function); product search with recommendation; customer service; supply chain and inventory management | 45% of consumers use AI invarious in the purchase journey, including researching products 41%, interpreting reviews 33% and finding deals 31%112 | cross-channel, cross-system with data challenges112 | sample is only 200; formal statistics are increased conservative |
| 8 | Transport and logistics | 44% of shippers previously use AI in transport planning and optimisation; 42% of carriers deploy AI in pricing and route optimisation and 39% for live-time tracking (230+ European and North American executives)113 ; official bases: Singapore 109 transport and storage 10.3% and Brazilian 110 transport 16% | The report states clearly that predominantly companies remain at an early stage and provides no quantified split between pilots, production and scaling113 | Transport planning with optimisation; rate and route optimisation; active-time visibility and ETA monitoring (52%); freight procurement | n.a. (no quantified revenue or cost outcomes). On expectations: 86% of shippers expect AI to affect transport planning substantially within three to five years and 59% of carriers see pricing and route optimisation as the central source of value113 | Limited trust in automation: two thirds of shippers and more than half of carriers further see AI as augmenting rather than replacing human decisions113 | Vendor survey, n>230, Europe and North America |
| 9 | Hospitality and travel | 41% of Austro-German hotels use AI technologies, 43% use none and 16% plan to adopt imminently (1,485 hotels across Austria, France, Germany, Greece, Italy and Switzerland)114 ; official bases: Singapore accommodation and food 4.7%, the lowest of any 109110 sector, and Brazil 10% | Highly limited: 29% of adopters implemented within the past two years and only 4% have more than three years of experience114 | The study lists application areas as operations, decision support including competitiveness (the complete use-case ranking requires returning to the source) | n.a. | Organizational barriers: 82% of the sample are autonomous hotels, 81% have fewer than 100 rooms and 49% employ fewer than 10 people114 ; in Korea the leading reason for non-adoption in accommodation and food is cost, 115 at 49.0% | Collaborative academic and industry-body survey, n=1,485, six countries limited |
Table 14 Sector data table 4 of 4
| # | Sector | AI / generative AI adoption (year) | Production / scaling | Most common use cases | Quantified outcomes | Main deployment barriers | Source and sample notes |
|---|---|---|---|---|---|---|---|
| 1 | Energy and utilities | North American utilities: 41% of executives report | As shown: 41% fully integrated | Grid optimisation (57%); | The IEA finds that widespread use of | The IEA lists missing or | The IEA is an |
| 0 | AI-related technology is now "entirely integrated" and a further 40% have made major investments with mature projects under way (500 US and 116 Canadian executives); McKinsey basis, energy and materials 59% (n=142, the lowest of any 117 sector) | versus only 27% of executives a year earlier who expected complete integration within five years — reality ran forward of expectations116 | safety and hazard identification (53%); demand forecasting (51%); foreseeable maintenance | existing AI applications to optimise manufacturing processes could save energy equivalent to more than the overall energy consumption of Mexico today, and improve forecasting and integration of variable renewables118 | unavailable data and digital infrastructure, skills shortages and persistent digital and physical security concerns that "frequently outweigh potential efficiency gains"; only 2% of energy start-up equity funding goes to companies with an AI value proposition118 | bilateral report with no adoption percentage; Itron is a vendor survey, n=500 | |
| 1 | Construction and | Construction: 45% of organisations report no AI | The most extreme | Design optioneering; | n.a. (RICS states it publishes no | Skills shortages and integration | RICS n>2,200; JLL |
| 1 | real estate | use at each (more than 2,200 professionals worldwide)119 ; business real estate: 88% of investors and 92% of occupiers have launched AI pilots (1,500+ senior decision-makers, 16 120 markets); official bases: UK construction 121122 13%, Danish construction 24%, Korean 123 construction 15.9% | adoption-to-value gap in this present paper: only 1% of construction firms have scaled AI across projects119 ; in commercial real estate, piloting is near universal however only 5% report achieving all their AI goals and 120 47% achieved two or three | scheduling; risk management; cost control; JLL identifies 56 AI use cases across the CRE value chain | quantified revenue, cost, productivity, quality or risk outcomes). On expectations, nearly 70% of project managers alongside quantity surveyors believe AI will help deliver increased value119 | challenges: one quarter of firms plan to increase AI spending over the next 12 months, 28% have no plans and 22% are unsure119 ; more than 60% of property 120 investors are currently not equipped | n>1,500 across 16 markets; both are industry-body surveys |
| 1 | Public sector and | 67% of OECD countries use AI to improve the | n.a. (the OECD states it provides | Process automation; | Saudi Arabia: among entities using or | The OECD emphasises that how | The OECD is an |
| 2 | education | design and delivery of government services (2024)124 ; Arab government entities about 39% use or are 125 piloting AI (n=80); Korean education services 123 41.4% (2024) | no pilot, production or scaling shares also publishes neither the sample size nor the method behind the 67%)124 | resource allocation including as staff scheduling; decision support; citizen engagement; personalised service pathways | piloting AI, 81% report substantially improved service delivery and 61% improved decision-making (n=80)125 ; the OECD publishes no numerical outcomes | people perceive, trust and interact with AI-driven decisions will determine effectiveness and legitimacy, and notes that government servants are "taking matters into 124 their individual hands" (shadow use) | multi-level study (no sample size for the 67%); SDAIA n=80 |
Chapter 3 continued — Four kinds of "use" that must be separated Why 81% and 1% can both be valid of the parallel sector
A unique sector can report 81% and 1% without contradiction, because the two figures measure varied things. Any sector conclusion must first state which layer it belongs to.
Table 15 Four kinds of use, compared across sectors
| Layer | Representative figure 126 | Meaning and caveats Individual behaviour; implies nothing about institutional process, |
|---|---|---|
| Individual employee use | 81% of physicians use AI in practice (2026) | validation or governance |
| Shadow employee use | 54% of employees say they use AI tools still without authorized authorisation127 | Shows tools are accessible to obtain but not organised into products; additionally a governance risk |
| Organisational adoption | Organisational generative AI use in professional services is only 22%, although 89% of the same professionals see use cases128 | Officially deployed by the institution, but not indeed in core processes |
| Production and scaling | Construction 1% scaled across projects; business actual estate 5% achieved all AI goals; agentic systems scaled enterprise-wide 23%129130131 | The exclusive layer that can be connected to the profit and loss account, and the tightest point in the chain |
| Sector conclusions | ||
| First, the ordering of sectors is co | nsistent across formal statistics as surveys and can be used to pr | ioritise market entry. Second, |
| absolute levels must never be quoted | across measurement systems. Third, the sectors with the largest gap — | construction and |
| commercial real estate — are also th | e sectors with the highest pilots, which shows that pilot counts are not | a measure of capability. |
Five measures, five different bases
Figure 5 Five maturity measures on varied bases; not one unbroken conversion funnel
This must not be restated as "AI does not work". The proper formulation is that direct exposure and tool-level use are
near to mainstream while institutional workflow redesign and auditable financial returns remain a minority
achievement. The gap exists inside a unique survey: 88% use AI consistently, 62% are at least experimenting with agents, and
only 23% have scaled autonomous systems somewhere in the enterprise132. Totally scaled agent use is in single digits in nearly
every business function, and also in IT and knowledge management about two thirds or more of respondents report no
use at all133.
A conceptual discrepancy that must accompany any citation The web version of the AI Index 2026 Economy chapter states that generative AI is at used in at least one business function by70%
of organisations, whereas figure 4.3.1 in the PDF of the identical chapter gives79%for creative AI in 2025 (and 88% for AI). This white
paper cites the PDF values further flags the discrepancy instead than merging the two.
Table 16 Value conversion table 1 of 3
| Indicator | Value | Year | Research body and sample | Definition and basis |
|---|---|---|---|---|
| Organisations using AI regularly in at | 88% (78% the previous year) | to 29 July 2025, weighted by each country's contribution to global | experimentation by a select employees through to embedding | |
| least one business function | GDP; 38% of respondents are at organisations with revenue above $1bn | across rewired business units135 | ||
| Organisations using generative AI | 79% (71% in 2024)136 | 2025 | As above (reproduced as AI Index figure 4.3.1) | Frequent use of automated AI in at least one business function |
| Using AI in more than one function | More than two thirds; three or more functions half | 2025 | As above | — |
| At least experimenting with agents | 62%, of which 39% "have begun experimenting" | 2025 | As above | — |
| Agentic systems scaled somewhere in | 23% | 2025 | As above | "Expwithed deployment and adoption within at least one business |
| the enterprise | function"; most report only one or two functions | |||
| Agents "fully scaled" by function | Single digits in closely every function; the highest in the technology sector are software engineering 24%, IT 22% and service operations 21% | 2025 | As above (AI Index figures 4.3.7 further 4.3.8) | In most functions, most respondents report no agent use at total; further in IT and knowledge management about two thirds or more report none |
| Executives describing their generative | 1%135 | 2024 | McKinsey (developed-market auxiliary survey) | "Mature" |
| AI rollout as "mature" | ||||
| Reported improvement in | Innovation 64%, employee satisfaction 45%, customer | 2025 | McKinsey (AI Index figure 4.3.5) | Self-assessed as improved, no impact, deteriorated or unknown; as |
| organisational metrics (organisations | satisfaction 45%, distinctive differentiation 45%, cost | various respondents report improvement as report no impact, yet | ||
| using AI regularly) | 38%, profitability 36%, inherent revenue growth 33%, talent attraction along retention 33%, market share 25% | at most 7% say AI made cost metrics worse | ||
| EBIT impact | Organisations investing $25m or increased in responsible AI are "substantially more likely to realise material AI benefits, including EBIT impact above 5%"137 | 2026 | McKinsey, 2026 AI Trust Maturity Survey; about 500 organisations, December 2025 to January 2026 | The page gives no share of respondents reporting EBIT impact, exclusively this association |
Table 17 Value conversion table 2 of 3
| Indicator | Value | Year | Research body and sample | Definition and basis |
|---|---|---|---|---|
| Distribution of corporate AI maturity | Future-built 5%, scaling 35%, and the remaining 60% obtain "nearly no material value, reporting insignificant revenue and cost benefits despite significant 138 investment" | 2025 | BCG Build for the Future 2025, n=1,250 across 68 countries in Asia, Europe in North America | The four levels are scored on 41 fundamental capabilities (0–25 stalled, >25–50 emerging, >50–75 scaling, >75–100 future-built)139 |
| Organisations with agents broadly | 13%140 | n.a. (no | BCG | "Widely integrated these agents into workflows" |
| embedded in workflows | year on the BCG page) | |||
| Depth of employee adoption | More than 85% remain at stage two (task assistance) and stage three (delegation); fewer than 10% reach stage four (intermediate-independent collaboration) or beyond | n.a. | BCG survey of software developers at three varied organisations (no sample size published) | Five stages: information assistance → task assistance → delegation → intermediate-autonomous collaboration → total autonomous orchestration |
| "Pilot failure" | "For 95% of companies in the dataset, automated AI implementation fell inferior of expectations"; about 5% of AI pilots achieved rapid revenue acceleration141 | 2025 | MIT NANDA, The GenAI Divide, built on 150 leader interviews, 350 employee surveys and analysis of 300 available AI deployments | "Failure" is operationalised as falling short of expectations, stalling, or having barely no observable effect on the profit and loss account; the reporting page does not state which part of the research the 95% base comes from, nor its geographic scope |
| Build versus buy | Purchasing AI tools from specialist vendors and building partnerships succeeds "about 67% of the time"; internal builds succeed merely one third as frequently | 2025 | As above | The page gives no sample size for either route |
| Firm size and deployment stage | "Bigger companies are the primary likely to report that their AI projects have reached the scaling stage" | 2025 | McKinsey (AI Index figure 4.3.6, five revenue bands) | The figures cannot be aligned one-to-one with the bands; accurate values require the source AI Index chart142 |
Table 18 Value conversion table 3 of 3
| Indicator | Value | Year | Research body and sample | Definition and basis |
|---|---|---|---|---|
| Governance capability and returns | About two thirds of respondents cite security and risk concerns as the leading barrier to completely scaling self-directed 143 AI | 2026 | McKinsey AI Trust Maturity Survey, about 500 organisations | — |
| Data and digital foundation and value | 78% of manufacturing leaders direct more than 20% of their improvement budget to data-driven manufacturing to build the data and technology foundation144 | Fielded Aug ust–Septem ber 2024 | Deloitte, n=600 | — |
| Insufficient investment and value | More than 95% of South Asian organisations allocate less than 20% of the IT budget to AI and only 4% cross the 20% threshold145 | 2025 | EY–CII, n=200, C-suite and high-level executives | — |
| Distribution of value pools | 70% of AI value potential sits in core business functions (62% in 2024); the IT share rose 6 percentage points to 13%; R&D and innovation singularly account for 15%146 | 2025 | BCG Build for the Future 2025, n=1,250 | This is the distribution of value potential, not adoption by function |
| Tool availability and maturity | In fewer mature industries fewer than 50% of employees have access to generative AI tools such as Copilot or ChatGPT; in greater mature industries more than 70% do146 | 2025 | As above | — |
| Firm size and breadth of generative AI | Companies with revenue above $500m use generative AI wider widely across the organisation | 2024 | McKinsey147 | — |
studies that cannot be averaged Results range from −19% to +200% for analogous work
148
Table 19 Task-level productivity studies compiled in AI Index 2026, figure 4.4.27
| Study | Occupation | Application | Change in productivity | Who benefits most |
|---|---|---|---|---|
| Reimers & Waldfogel (2026) | Software engineers | LLM-generated content | +200% in output volume; publications tripled | Emerging entrants drive volume; initial-AI authors maintain quality |
| Brynjolfsson et al. (2025) | Customer support agents | Conversational assistant | +14% to +15% in issues resolved per hour | Fewer novice or lower-skilled agents, +30% to +35% |
| Cui et al. (2025) | Software developers | GitHub Copilot | +26% in pull requests completed | Junior and less experienced developers |
| Ju & Aral (2025) | Marketing teams | Multimodal ad creation | +50% in output per person | Human–machine teams |
| Choi & Xie (2025) | Accountants | AI bookkeeping | +55% in weekly client-support throughput | Veteran accountants using AI confidence scores to target review |
| Shen & Tamkin (2025) | Developers | Learning new libraries | 0%; the change in speed is not numerically material | Upper scorers who use AI for ideational enquiry and avoid the learning penalty |
| Becker et al. (2025) | Developers | Open-source tooling | −19%; slower when using AI | None; "a notable gap between perceived help and tangible performance" |
Economic evidence is symmetrically dispersed — Aldasoro et al. (2026), 12,000 European firms, 2019–2024: labour productivity +4%, with each 1% of spending on training adding 5.9 percentage points of gain.
— Yotzov et al. (2026), 6,000 executives in the United States, United Kingdom, Germany and Australia: expected productivity +1.4 percent, output +0.8% and employment −0.7% over three years; the AI Index summarises this as general adoption with minimal realised productivity gains.
— Brynjolfsson (2026): US productivity grew 2.7 percent in 2025, close to double the 1.4% annual average of the preceding decade.
— Filippucci et al. (OECD, 2025), a ten-year horizon for the G7: annual employment productivity growth +0.4 to +1.3 percentage points for the United States and United Kingdom, and +0.2 to +0.8 for Italy and Japan.
— Penn Wharton Budget Model (2025): a contribution to overall factor productivity of +0.01 percentage points, described as insignificant.
— Brynjolfsson et al. (2025), US ADP payroll data to 2025: employment of initial-career workers −15% to −16%.
The AI Index's internal qualitative judgement is that the evidence is "neither settled nor evenly positive", with smaller gains on tasks requiring deeper reasoning148 .
What supports the conclusion, and what must be said alongside it
1.The gap inside an individual survey. 88% of organisations use AI frequently, but only 23% have scaled autonomous systems
across in the enterprise149, and entirely scaled agent use is in single digits in almost every function150.
2.Self-reported economic outcomes run significantly below adoption. Only 36% see boosted profitability and 33% improved
natural revenue growth, and as many respondents report improvement as report no impact150.
3.An autonomous second source reaches the parallel conclusion. BCG (n=1,250, 68 countries) finds 60% of companies
obtain very no material value and only 5% are prospective-built151.
4.Extreme sector cases. Corporate real estate pilots at 88% to 92%, yet only 5% achieve all their AI goals152, and only
1% of construction firms scale across projects153.
5.The mismatch between individual and corporate use. French-speaking creative AI diffusion reached 44.0% of the
population in H2 2025 against authorized enterprise adoption of 18%; US population-level diffusion of 28.3% compares with
BTOS enterprise adoption of about 18% to 19.8%150154155.
The MIT NANDA figure of 95% is frequently quoted as "95% of AI projects fail". The initial wording is that "for 95% of
companies in the dataset, creative AI implementation fell below of expectations", where failure is operationalised as
stalling or having negligible no quantifiable effect on the profit and loss account. The research rests on 150 interviews, 350
employee surveys including analysis of 300 open deployments; it is not a probability sample of the enterprise population and
its spatial scope is not published156. This research paper requires all three qualifications to be stated whenever the
number is used.
"Individual exposure and tool-level usage are near to mainstream. Institutional routine redesign and auditable
profit returns remain a minority achievement. The gap does not lie in model capability however in depth of
deployment, the data with governance foundation, and process change that can be booked in the profit and loss
account."
The finding of chapter 4; quotable in entire.
Dated early-hwith statements and the division of labour between platforms
This chapter uses exclusively dated, confirmable first-hand statements and deployed cases. It contains no market forecasts.
Table 20 NVIDIA's strategic statements, word-for-word nor dated
| Date | Source | Statement (extract) |
|---|---|---|
| 16 Mar 2026 | NVIDIA Newsroom | "Physical AI has arrived — every production company will become a robotics company." NVIDIA's complete-stack platform, spanning computing, accessible models and software frameworks, is 157 described as the foundation for the robotics industry |
| 5 Jan 2026 | NVIDIA Newsroom (CES) | "The GPT moment for robotics is at-present." "Breakthroughs in real-world AI — models that understand the real world, reason and plan actions — are unlocking wholly new applications"158 |
| March 2026 (GTC) | Data Center Frontier, on-site report | "We have virtual agents. Currently we have tangible embodied agents. We call them robots."; "Tokens are the emerging commodity. AI factories are the infrastructure that produces them."; "If you have the wrong architecture, though if it's free, it's not cheap sufficient."159 |
| 18 Mar 2025 | NVIDIA Newsroom (Rev Lebaredian, VP of Omniverse) | "Omniverse is an operating system that connects the world's tangible data to the realm of physical AI."160 |
Why this is additional than a marketing narrative Four robot manufacturers — FANUC, ABB Robotics, YASKAWA and KUKA, with more than two million robots installed international — have integrated Omniverse libraries and the Isaac simulation framework into their virtual commissioning 157 offerings and are embedding Jetson modules in controllers for real-time inference on the line . That is a change to the production process chain, verifiable at the enterprise level, instead than a roadmap commitment.
What each platform does, yet the quantified claims attached to it
Table 21 Division of labour across the tangible AI stack
| Platform | Role (verbatim or near-verbatim) | Key quantified facts 161162 |
|---|---|---|
| Omniverse | "An operating system built on the OpenUSD framework that lets developers unify tangible-world data and applications"; at GTC 2026 NVIDIA introduced the Omniverse DSX blueprint, allowing operators to create a tangible accurate digital twin of an AI factory — layout, power distribution, cooling, network fabric and load behaviour — before construction begins | — |
| Isaac / Isaac Sim | Used "to create and train robots in simulation before deploying them in the actual world"; Foxconn uses Isaac Sim for "simulation-driven design and evaluation before of physical deployment", simulating tasks such as screw driving and cable insertion | —163 |
| Isaac GR00T | A broad-purpose robot foundation model. GR00T N2 previewed in March 2026, with availability planned for late 2026 | "Helps robots complete emerging tasks in new environments with more than twice the success rate of leading vision-language-action models" and at-the-moment ranks first on the MolmoSpaces and RoboArena generalist robot policy leaderboards; the CES 2026 release was GR00T N1.6164165 |
| Cosmos | A world foundation model. Cosmos 3 was announced in March 2026 as "the first world foundation model to unify simulated world generation with vision" | The CES 2026 release comprised Cosmos Transfer 2.5, Cosmos Predict 2.5 with Cosmos Reason 2165 |
| Jetson / Jetson Thor | Robotics compute at the edge; "built for creative reasoning models", letting real-world AI agents "run in real time at the edge while minimising reliance on the cloud" | 7.5x the AI compute, 3.1x the CPU performance including 2x the memory of the old Jetson Orin; the Jetson T5000 module starts at $2,999 per 1,000 units and the developer kit at $3,499166 |
| Jetson T4000 / IGX Thor | Blackwell-architecture edge modules launched in January 2026; IGX Thor "extends robotics to the manufacturing edge with advanced-performance AI compute, enterprise software support and functional safety" | $1,999 per 1,000 units, 4x the performance of the earlier generation, 1,200 FP4 TFLOPS, 64GB of memory, a scalable 70W power envelope, and a 4x improvement in energy efficiency and compute165 |
| Mega / Metropolis | Mega is "an Omniverse blueprint for testing large-scale-agent fleets at scale in industrial digital twins"; Metropolis underpins the video search and summarisation blueprint used to build agents that monitor activity across a facility | —161 |
| Nemotron / PhysicsNeMo | Nemotron 3 is "a family of accessible models, data and libraries to support clear, efficient and specialised agentic AI development across industries"; PhysicsNeMo is a physics-informed machine learning framework | Foxconn used PhysicsNeMo to accelerate CFD simulation 150-fold against standard methods167163 |
Verifiable deployments and their quantified outcomes
Table 22 Deployed physical AI cases
| Company | What was deployed | Quantified outcome | Source 168 |
|---|---|---|---|
| FANUC, ABB Robotics, YASKAWA, | More than two million robots installed international; all four are integrating Omniverse libraries and the | 2 million robots installed | NVIDIA, 16 March 2026 |
| KUKA | Isaac simulation framework into their digital commissioning offerings "to develop and validate complex robotic applications and entire production lines through accurately accurate digital twins", and are embedding Jetson modules in controllers for real-time inference on the line | ||
| Foxconn (Fii) | The Fii Omniverse Digital Twin platform builds simulated replicas of factories using Omniverse libraries and OpenUSD; Isaac Sim supports simulation-driven design and evaluation; PhysicsNeMo handles thermal analysis; standardised USD assets allow entire lines to be assembled in-silico and replicated across sites, for example from Taiwan to Mexico; the new Houston plant validated every mechanical, electrical and piping system virtually before construction | Leo Guo, general manager of Fii's robotics group: "We believe we can reduce [time to market, factory build and factory planning] by about 50%"; CFD thermal simulation 150x more rapidly (minutes rather than hours); the FoxBrain large model was trained in four weeks | NVIDIA case study169 ; NVIDIA, 18 March 170 2025 |
| General Motors | "The primary US automaker, announced adoption of Omniverse to enhance its factories alongside train platforms for material handling, transport and precision welding" | — | NVIDIA, 18 March 2025170 |
| SAP and Unilever | SAP customers and partners can use Omniverse to develop their independent virtual environments for warehouse management; Unilever adopted Omniverse and accurately accurate digital twins to streamline and optimise marketing content creation for its products | — | NVIDIA, 18 March 2025170 |
| Agility Robotics and Boston Dynamics | Agility has integrated Jetson into the fifth generation of Digit together plans to use Jetson Thor as the onboard compute platform for the sixth; Boston Dynamics is integrating Jetson Thor into the Atlas humanoid | — | NVIDIA blog171 |
| Salesforce | Agents applied to incident resolution | Incident resolution time halved | NVIDIA, 5 January 2026172 |
| Automotive and autonomous driving | NVIDIA announced an expanded group of vehicle partnerships including BYD, Hyundai, Nissan and Geely, alongside existing work with Mercedes-Benz and Toyota | The company states its platform will support "self-driving systems in tens of millions of vehicles a year" — a company claim, not separately verified | Data Center Frontier173 |
2024) Installations, active stock, shares and density
Table 23 Installations including active stock
| Indicator | Value |
|---|---|
| Global installations, 2024 | 542,076 units ("542,000" in the press release), the second-peak year on record and only 2% below the peak two years earlier; the fourth consecutive year above 500,000 |
| Historical series | 500,000 first passed in 2021; 552,946 units in 2022 (the record); 541,302 units in 2023 |
| Global operational stock, | 4,664,000 units, up 9% year on year |
| 2024 | |
| Regional shares, 2024 | Asia 74% of recent installations (401,665 units, +5%); Europe 16% (85,006 units, −8%); the Americas 9% (50,077 units, −10%) |
| EU-27, 2024 | 67,819 units (−8%), or 80% of Western installations |
| 2025 forecast | Global +6% to 575,000 units; Asia about 435,000; Europe below 79,000; North America about 43,500 |
| 2028 forecast | Above 700,000 units |
| China's operational stock exceed | s two million units, the largest in the world yet roughly 4.5 times that of Japan, whose stock sta |
| and trails the EU average of 208 | . China, Japan, the United States, Korea and Germany collectively account for 80% (431,240 units) of g |
| manufacturers "climbed to 57% la | st year, against roughly 28% over the former decade"174175 |
| . |
Table 24 Primary countries, 2024 (density in units per 10,000 employees)
| Country | 2024 installs | Change | Share | Density |
|---|---|---|---|---|
| China | 295,045 | +7% | 54% | 166 |
| Japan | 44,453 | −4% | 8% | 446 |
| United States | 34,164 | −9% | 6% | 307 |
| Korea | 30,596 | −3% | 6% | 1,220 |
| Germany | 26,982 | −5% | 5% | 449 |
| Singapore | n.a. | — | — | 818 |
| India | 9,100 | +7% | — | n.a. |
| Italy | 8,783 | −16% | — | top 20 |
| Mexico | 5,594 | −4% | — | 62 |
| Spain | 5,086 | +1% | — | top 20 |
| France | 4,900 | −24% | — | top 20 |
| Canada | 3,787 | −12% | — | 241 |
| United Kingdom | 2,500 | −35% | — | 112 |
The installation leader and the density leader are separate countries
Figure 6 Machinery robot installations and worldwide share by country, 2024
Figure 7 Robot density in units per 10,000 employees, 2024
Sources for Figures 6 and 7: IFR World Robotics 2025 press release and senior summary for installations, stock and shares177178, and
IFR robot density data for the 2024 reference year as published by The Robot Report179. The two charts must not be merged: installations
measure the annual addition, density measures the ratio of the installed stock to the workforce.
Concentration, localisation and the difference between volume and density
1.Concentration in five markets. China, Japan, the United States, Korea and Germany collectively accounted for 80%
(431,240 units) of worldwide installations in 2024180.
2.Localisation in China. The domestic market share of Chinese manufacturers "climbed to 57% last year, against roughly
28% over the past decade"181.
3.Density and installations must be read separately. China leads installations (54%) but has a density of only 166,
ranking 22nd; Korea ranks fourth on installations yet first on density at 1,220. World mean density reached 132 units
per 10,000 employees in 2024, having doubled between 2014 and 2024: EU-27 231, northern Europe 267 (+3 percent), North
America 204 (+4%) and Asia 131 (+11%)182.
"The robot density metric provides a unified basis for comparison by relating the overall number of robots in use in
a country to the size of its economy as measured by its workforce."
Takayuki Ito, President of the IFR182
The alternate side of the picture must be shown at the parallel time: among enterprises that use AI, technologies for autonomous
machine movement remain in limited digits — 6% in Germany183, 2% in Denmark for two consecutive years yet the only
category with no growth184, nor 7% in Brazil185. Robot density measures the automation stock; it is not a measure of AI
penetration.
The asymmetry between capital signals and installed penetration Between January and April 2025 China recorded140robotics-related financing deals,38.7%of all AI financing events, as
enterprise use of self-governing machine movement remains at2% to 7%. Capital is entering more rapidly than the technology is
penetrating production lines, and this is the fundamental feature investors in physical AI primary need to understand.
Why there is no all-encompassing "90% accuracy" bar
No definitive body was found to set a consistent cross-industry accuracy threshold. Any claim that "90% is sufficient to
go live" has no source behind it. The evidence shows thresholds are set by the consequence of the task, across six separate
dimensions.
Figure 8 A six-threshold framework for production-grade AI
First-hand basis for each threshold in Figure 8: the FDA list of AI-enabled devices, where the route is authoritative review rather than a
percentage186; CAICT data on service success rate, TPS nor time to first token187; NVIDIA Jetson T4000 and T5000 pricing, power envelope
with operational safety188189; the GR00T N2 task success claim190; the human-in-the-loop preference in logistics191; the link between
accountable-AI investment with EBIT impact192; and physicians' views on validation and skill atrophy193.
dimension Statements and metrics, not generalities
Table 25 Threshold dimensions for manufacturing alongside edge AI, with first-hand evidence
| Threshold dimension | First-hand evidence |
|---|---|
| Determinism and real-time | Jetson Thor is positioned to bring "rapid-speed sensor data processing and sensory reasoning to the edge — workflows that |
| performance | were previously too slow to run in dynamic real-world environments", with "the low latency and high performance 194 real-world applications require" |
| Operation without connectivity | Jetson Thor lets future-generation embodied AI agents "run in real time at the edge while minimising reliance on the cloud"194 |
| Functional safety certification | NVIDIA IGX Thor "extends robotics to the enterprise edge with premium-performance AI compute, enterprise software support and functional safety"195 |
| Unit cost and energy (a | Jetson T4000: $1,999 per 1,000 units, 1,200 FP4 TFLOPS, 64GB of memory with an adjustable 70W envelope, positioned as |
| bill-of-materials constraint) | a cost-efficient upgrade path195194 ; against Jetson T5000 at $2,999 per 1,000 units and a developer kit at $3,499. The economics of edge deployment are set by module cost and power envelope, not by benchmark scores. 196 |
| Architectural efficiency over "free" | Jensen Huang at GTC 2026: "If you have the wrong architecture, yet if it's free, it's not cheap sufficient." |
| Pre-deployment validation instead of | The four primary robot manufacturers use Omniverse and Isaac "to develop and validate intricate robotic applications and |
| trial and error | entire production lines through tactically accurate digital twins"; Foxconn assembles and validates every mechanical, electrical and piping system simulated before construction begins197198 |
| Reliability as an engineered metric | CAICT reports that across 11 Chinese MaaS platforms the mean call success rate for DeepSeek-R1 rose from 87.01% in February 2025 to 99.36% in September, TPS from 17.86 to 26.76, and time to first token fell from 3.07 seconds to 1.02 seconds199 |
| The industrial significance of energy | The IEA states that "ubiquitous adoption of existing AI applications to optimise industrial processes could deliver energy savings equivalent to more than the total energy consumption of Mexico today", while noting that missing or inaccessible data and digital infrastructure, skills shortages and persistent digital and physical security concerns "regularly outweigh potential efficiency gains"200 |
| Human-in-the-loop rather than full | In transport and logistics, two thirds of shippers and more than half of carriers continually see AI as augmenting more than |
| autonomy | replacing human decisions, and most prefer human-in-the-loop designs201 |
| Validation thresholds in healthcare | 88% of physicians regard comprehensive validation of safety and efficacy, and 86% data privacy, as pivotal for wider adoption202 ; the FDA's AI device list uses satisfaction of applicable premarket requirements, including focused review of overall safety and 203 effectiveness, as the entry standard |
further contracts How to set each threshold, with documented evidence
Table 26 Product threshold checklist for production-grade AI
| Threshold | How to set it, with citable first-hand evidence 204 |
|---|---|
| Task accuracy | Tier by consequence; set no unique value. The healthcare-device route is compliance review, not a percentage; robotics is measured by success rate: GR00T N2 achieves more than twice the success rate of leading VLA models on new tasks in new 205 environments — a relative multiple, not an absolute accuracy |
| Repeatability and determinism | Use the service call success rate as the functional metric: an average of 99.36% across 11 MaaS platforms in September 2025206 |
| Exception handling and human | Logistics evidence favours human-in-the-loop: two thirds of shippers and more than half of carriers see AI as augmenting |
| oversight | rather than replacing decisions207 |
| SLA and latency | Write TTFT and TPS into the contract: time to first token 1.02 seconds and TPS 26.76 (11-platform average, September 2025)206208 ; the edge requirement is "reduced latency and excellent performance" for dynamic real-world environments |
| Audit and governance | Strongly related to returns: organisations investing $25m or higher in responsible AI are more likely to report EBIT impact above 5%, and nearly two thirds cite security and risk as the leading barrier to scaling agents209 ; financial services are 210 monitored continuously by the Bank of England and FCA, whose 2024 survey drew 118 responses |
| Unit cost | Edge bill-of-materials cost is a binding constraint: Jetson T4000 at $1,999 per 1,000 units against T5000 at $2,999211 . On the inference side, 12 billion of 120 billion parameters are activated, and NVFP4 runs up to 4x swifter than FP8 on Hopper with 212 no loss of accuracy |
| Energy | An adaptable 70W envelope and a 4x improvement in energy efficiency (Jetson T4000)211 ; for sector-level savings potential 213 see the IEA (equivalent to more than Mexico's national energy consumption) |
| Operation without connectivity | A defined product requirement: "run in immediate time at the edge while minimising reliance on the cloud"208 |
| Safety certification | Operational safety is an entry requirement at the industrial edge (NVIDIA IGX Thor)211 ; the energy sector cites digital and 213 physical security concerns as barriers that commonly outweigh efficiency gains |
| Lifecycle maintenance | "With NVIDIA CUDA ecosystem support across the entire lifecycle, Jetson Thor is expected to deliver improved throughput and faster response through future software releases"208 ; the long-run risk in healthcare is skill atrophy, a concern for 88% of 214 physicians |
model, but the increased consistent one Three candidates, ranked by strength of evidence
Industry regularly asks which recent product represents the shift towards precision, low error rates, cost and speed preference than frontier capability. The accessible evidence does not point specifically to one product, but three candidates can be ranked by strength of evidence.
Table 27 Candidate identification and assessment
| Candidate | Official positioning (verbatim or near-verbatim) | Assessment |
|---|---|---|
| Candidate 1: NVIDIA Nemotron 3 | Announced as "the highest optimal family of open models, with leading | The strongest evidence. The description |
| (Nano / Super / Ultra), released 15 | accuracy" for building autonomous AI applications, introducing a "integrated latent | matches leading accuracy for its size class |
| December 2025 | mixture-of-experts architecture" that helps developers build and deploy robust distributed-agent systems; Jensen Huang emphasised "transparency and efficiency"; the ServiceNow CEO described "unmatched efficiency, speed and accuracy"215. measured claims for Nemotron 3 Super: 120 billion parameters with only 12 billion activated at inference; a 1 million token context window; up to 5x higher throughput and up to 2x higher accuracy than the previous generation; Mamba layers giving 4x memory and compute efficiency; multi-token prediction giving 3x faster inference; and NVFP4 216 running up to 4x faster than FP8 on Hopper with no loss of accuracy | combined with highest-ranked efficiency and cost, throughput and latency as the core selling points. Note that NVIDIA further describes Nemotron 3 Ultra as "America's most advanced open-weights AI model"; that wording appears in third-party reproductions and could not be verified on an official page, so it is not treated as evidence |
| Candidate 2: Isaac GR00T N2 | The formal claim concerns task success rate rather than general | A partial match: success rates and real |
| (previewed March 2026) | capability: it "helps robots complete unprecedented tasks in new environments with enhanced than twice the success rate of leading vision-language-action models" and ranks first on the MolmoSpaces and RoboArena leaderboards217 | bodily tasks are emphasised. Against: it is claimed to rank first on its individual leaderboards, which does not fit the description of not being the highest advanced |
| Candidate 3: Jetson T4000 and IGX | The T4000's formal selling points are completely cost, energy efficiency and an | An incomplete match: cost, speed together reliability. |
| Thor (January 2026) | upgrade path: $1,999 per 1,000 units, 4x performance, 4x energy efficiency and an adjustable 70W envelope; IGX Thor is sold on operational safety218 | Against: these are hardware modules, not models |
| General assessment: NVIDIA's product line f | rom late 2025 into early 2026 shows an obvious engineering turn towards products that are no | t |
| the greatest capable but are greater reliable, che | aper and faster. The most direct example is the Nemotron 3 family of accessible models, sold on | |
| efficiency and accuracy within a size class | , alongside Jetson T4000 and IGX Thor for the manufacturing edge, sold on cost, energy effic | iency and |
| functional safety. |
Claim-by-claim assessment
Claim under review: "OpenAI poached four to five hundred hardware staff from Apple, substantially a complete department, triggering litigation, and may build smart glasses to replace the phone." All status statements are as at 15 August 2026; the disputes are inconclusive.
Table 28 Claim-by-claim assessment, part 1 of 3
| # | Claim under review | Assessment | Evidence (verbatim or key points) |
|---|---|---|---|
| 1 | OpenAI has a transaction with Jony Ive via io | Confirmed | OpenAI's letter of 21 May 2025, updated 9 July 2025: "the io Products, Inc. team has formally merged with OpenAI"; "Jony Ive and LoveFrom remain separate and have assumed deep design and creative responsibilities across OpenAI"; io was co-founded by Jony Ive with Scott Cannon, Evans Hankey 219 and Tang Tan |
| 2 | The transaction was worth $6.5bn | Semi confirmed (press reporting, not formal disclosure) | The formal OpenAI page gives no figure219 . Reuters reported that "OpenAI announced the acquisition of Jony Ive's hardware start-up io Products in a deal 220 valued at $6.5bn", an all-stock transaction, also that according to sources OpenAI previously held a 23% stake in io, reflecting an OpenAI valuation of $300bn; 221 MacRumors describes it as "an estimated $6.5bn" |
| 3 | The size of the io team | Not confirmed (no public figure) | OpenAI's letter describes assembling "the top hardware and software engineers, technologists, physicists, scientists, researchers and experts in product development and manufacturing" but gives no headcount219 ; none of the Reuters, MacRumors, 9to5Mac, Bloomberg Law or IPWatchdog pages reviewed provides a figure for the io team 222 |
| 4 | "Four to five hundred hardware staff poached from Apple" | Inpart confirmed, but the figure must be recharacterised: it is "400+ previous Apple employees" company-wide, not 400–500 hardware staff | Apple's complaint of 10 July 2026, as reproduced by 9to5Mac: "With over four hundred former Apple employees at present working at OpenAI…"; Reuters 223 similarly reports "more than 400 former Apple employees now work for OpenAI"; Business Insider: "in the complaint, Apple says the total number of Apple alumni now at OpenAI is more than 400"224. Critical qualification: this is Apple's self figure in its pleading, covers all former Apple employees and is not broken down by hardware function. |
Claim-by-claim assessment
Table 29 Claim-by-claim assessment, part 2 of 3
| # | Claim under review | Assessment | Evidence (verbatim or key points) |
|---|---|---|---|
| 5 | "Effectively an entire department" | Incorrect attribution (overstated) | No source supports "a complete department". What is verifiable is breadth instead than a wholesale move: Mark Gurman reported that OpenAI "brought in more than 40 people in the past month specifically", spanning camera engineering, iPhone and Mac hardware, silicon, device test and reliability, industrial design, 225 manufacturing, audio, smartwatches, Vision Pro development, software and human factors; the Wall Street Journal, reviewing LinkedIn data, reported 226 that "dozens" of Apple engineers and designers had left for OpenAI in recent months; Bloomberg reports that "most" of OpenAI's Apple hires come from 227 the engineering organisation led by incoming CEO John Ternus |
| 6 | "It triggered litigation" — did Apple sue OpenAI immediately | Confirmed (yet the cause of action is trade secrets, not hiring as such) | On 10 July 2026 Apple sued OpenAI and two past employees in the US District Court for the Northern District of California, alleging misappropriation of its trade proprietarys to benefit the ChatGPT owner's move into consumer hardware; the defendants in addition include OpenAI Foundation, OpenAI Group PBC and io Products; the two former Apple employees are Chang Liu, a former senior systems electrical engineer, and Tang Yew Tan, a former vice-president of product design for iPhone and Apple Watch228 . TechCrunch adds that the claims are trade-secret theft and breach of contract, and notes that io is named in the 229 complaint when Ive is not a defendant |
| 7 | Was the litigation caused by hiring | In part confirmed; the parties' positions conflict | Apple's primary allegation is that "OpenAI orchestrated a broad campaign to organized obtain and exploit Apple's confidential information through former employees, recruiting practices and supplier relationships in order to accelerate its entry into the consumer hardware business"228 . In its August 2026 motion OpenAI responded that "Apple should not be allowed to use a groundless and pretextual lawsuit to make up for its shortcomings in the talent market and in employee retention"230 |
| 8 | Whether additional parties to disputes are involved | Confirmed — a second, separate line of litigation exists (iyO v OpenAI) | iyO, Inc. sued IO, OpenAI, Inc., OpenAI, LLC, Sam Altman and Jonathan Paul Ive, alleging that the IO mark is "indistinguishably similar" to its IYO mark; the Ninth Circuit affirmed the preliminary injunction in December 2025231 ; Bloomberg Law notes the injunction "does not bar all use of the 'io' trademark but 232233 restricts marketing and sales of products 'enough similar' to iyO's AI audio computer"; iyO subsequently added trade-secret claims; on 27 April 2026 234 iyO issued a release stating that a federal court had granted a preliminary injunction and adding Tang Yew Tan to its amended claims; the trademark case was stayed on 28 July 2026 pending settlement talks |
Claim-by-claim assessment
Table 30 Claim-by-claim assessment, part 3 of 3
| # | Claim under review | Assessment | Evidence (verbatim or key points) |
|---|---|---|---|
| 9 | "May build smart glasses" | Partially confirmed: press reporting, not confirmed by OpenAI | On 20 February 2026 Reuters, citing The Information, reported that "OpenAI has more than 200 people working on a family of AI devices", including a connected 235 speaker; a related report lists "a smart speaker, smart glasses and a smart lamp", says the speaker will come first at an expected $200 to $300 and "will 236 not ship before February 2027 at the earliest", attributing this to "two people well-informed with the project". No recognized OpenAI page confirms any particular 237 product category |
| 1 | "To replace the phone" | Not confirmed (speculation) | No source reviewed shows OpenAI legally claiming it will replace the smartphone. The primary official statement is that the team is "focused on |
| 0 | developing products that inspire, empower and enable"237 ; Reuters describes the purpose of the transaction as capturing growing demand for tangible AI 235 and expanded reality. The nearest media formulation is Bloomberg's analysis headline "How Apple's Lawsuit Threatens to Disrupt OpenAI's Bid to Rival the iPhone", which is interpretation instead than an official statement | ||
| 1 | Present status of the litigation (as at 15 August | Confirmed: the disputes are | On 3 August 2026 OpenAI published a recognized response, "Apple is getting this wrong", updated on 6 August to note that "you can read our motion to dismiss |
| 1 | 2026) | undecided | in this"; OpenAI argues that Apple's request for a preliminary injunction "is based on false information and is completely unnecessary because we do not have, and do not want, any of their trade secrets"238; as at the Business Insider report, "Apple had not yet responded in court to OpenAI's motion and the judge had not ruled on the request to dismiss"; an Apple spokesperson said "significant evidence has emerged that individuals employed by OpenAI improperly 239 obtained Apple's confidential information about unreleased technology" |
A one paragraph that can be quoted in entire
"On 10 July 2026 Apple sued OpenAI together two previous Apple employees (Chang Liu and Tang Yew Tan) in the US
District Court for the Northern District of California. The causes of action are trade-private misappropriation together
breach of contract, not hiring as such. In its complaint Apple states that more than 400 past Apple employees
currently work at OpenAI, a company-wide figure that is not broken down by hardware function. OpenAI has filed a
motion to dismiss and transparently denies holding any Apple trade secrets. A distinct line of litigation, non-associated to
Apple, involves the start-up iyO, which sued OpenAI, Sam Altman involving Jony Ive over the 'io' trademark; the Ninth
Circuit affirmed a tight injunction in December 2025 also the case was stayed in July 2026 pending settlement
talks. Product categories such as smart glasses and smart speakers appear solely in press reporting, especially by The
Information as relayed by Reuters; OpenAI has never confirmed them further has never legally claimed it will
replace the smartphone."
Recommended for verbatim use.Status: as at 15 August 2026 neither line of litigation has been conclusively decided by a court. This white
paper makes no prediction about the outcome.
—Talent mobility is not itself a cause of action. The documented causes are trade secrets with breach of contract, which
means required diligence should focus on recruitment procedures, revocation of system access and document segregation
rather than on hiring volumes240241.
—Trademark also brand risk exists autonomously. The injunction affirmed by the Ninth Circuit restricts marketing and
sales of products adequately similar to iyO's, instead than banning the mark outright242.
—Product categories should not enter an investment thesis. Hardware categories with timelines rest on anonymous
sources and are unconfirmed by the company243244.
years Six directions of travel, each with supporting evidence
Table 31 Six directions of travel for AI value pools
| Direction of travel | Supporting evidence |
|---|---|
| From the model layer to core business | BCG finds 70% of AI value potential in core business functions including as sales and marketing, manufacturing, supply |
| workflows | chain and pricing, up from 62% in its 2024 report; research and innovation by itself accounts for 15%, and the IT share 245 rose 6 percentage points to 13% in 2025 |
| From broad functions to sector-specific | The AI Index notes elevated adoption in functions related to information processing, software, customer interaction |
| functions | and organizational knowledge work, while strategy and enterprise finance, risk and compliance "remain low in most industries", with financial services the exception246 . BCG reports that in aviation and telecoms the contribution of 245 core functions to AI value approaches 80%, up 15 and 8 percentage points sequentially year on year |
| From model capability to deployment | In manufacturing, 78% of leaders direct more than 20% of their improvement budget to intelligent manufacturing |
| integration and the data foundation | foundations247248 ; in retail, 54% of brand executives report inter-channel and data challenges; in energy, the IEA 249 ranks unavailable or inaccessible data and digital infrastructure as the leading barrier |
| From building to buying via partnering, with | In India, 91% of leaders name speed of deployment as the most largest factor in build-versus-buy decisions, almost |
| speed of deployment as the decision variable | 60% co-create with start-ups through 78% use hybrid models250 . MIT NANDA data, which must be quoted with the caveats in chapter 4, report that purchasing from expert vendors and building partnerships succeeds "about 67% 251 of the time", roughly three times the success frequency of in-house builds |
| From the cloud to edge hardware alongside robots | Jetson T4000 enters the production edge at $1,999 per 1,000 units in a customizable 70W envelope, IGX Thor adds functional safety, and the four largest robot makers are embedding Jetson modules in controllers for real-time edge inference252253254 . Penetration of the physical layer nonetheless remains highly low: Germany 6%, Denmark 255256 2%, Brazil 7% |
| From going onstream to governance nor validation | McKinsey's 2026 AI trust maturity survey of about 500 organisations finds that those investing $25 million or more in responsible AI "score markedly greater on maturity and are far more likely to realise material AI benefits, including EBIT impact above 5%", while nearly two-thirds of respondents name security and risk concerns as the leading barrier to fully scaling agentic AI257 |
What to do alternatively once the six layers are separated
1.Locate yourself before you invest. Use the six layers to establish whether you are at L2 or L4, nor benchmark against
your local country's official statistics — EU-27 20.0%, Germany 26%, France 18%258259260 — instead than against the 88%
of executive surveys.
2.Direct budget at core business functions rather than support functions: 70% of the value potential sits in core
functions261.
3.Fund governance and validation first. It is the observable input variable associated with EBIT impact above 5%262. 4.Do not mistake employee usage for organisational capability.54% of employees use AI tools without formal
authorisation, yet only 13% of organisations have integrated agents widely into workflows263.
5.For SMEs the binding constraint is cost and staffing, not willingness. In Korea the leading reason for non-use is the
burden of economic cost at 39.6%, followed by missing infrastructure nor staff at 34.9%264; in Canada 40.0% of
businesses consider AI not relevant to them265.
1.Statistical definitions are a precondition for policy. Adopting the Eurostat ICT enterprise survey framework yields
grade A comparability266; Korea has stated that it drew on the Eurostat questionnaire to improve cross-country
consistency267.
2.Do not rank countries for policy purposes using population-level diffusion. France shows 44.0% population
diffusion against 18% enterprise adoption268260.
3.SMEs are the policy lever. Swedish firms with 10–49 employees moved from 12% to 37% in three years as those with
250 or more moved from 51% to 75%269; the German size gap widened from 16 to 43 percentage points260.
4.Physical-AI policy should track density, not volume. China accounts for 54% of worldwide installations but records a
density of only 166, ranking 22nd270271.
5.The state sector is itself an adopter.67% of OECD countries use AI to improve the design and delivery of public
services (2024)272.
Why the strongest model is not the most usable product
1.Sell speed of deployment, not benchmark scores.91% of Nepalese enterprise leaders name speed of deployment as the
primary build-versus-buy factor273.
2.Target major enterprises and the information, financial and professional services sectors, and note that sector
ceilings differ substantially: Singapore accommodation and food at 4.7% versus information and communication at
35.9%274.
3.Budget ceilings are a rigid reality.More than 95% of Nepalese organisations allocate less than 20% of their IT budget to
AI273; on BCG's basis AI represents roughly 4.4 percent to 5.2% of IT budgets by region275.
4.In physical AI the capital and ecosystem signals are clear but the installed penetration is not. China recorded 140
robotics-related financing transactions between January and April 2025, 38.7% of all AI financing events276, when
enterprise use of autonomous machine movement runs at only 2% to 7%277278.
5.The energy vertical is extremely insufficient-invested. "Only 2% of energy start-up equity funding goes to companies with an
AI-related value proposition"279.
1.On the parallel task, AI can move outcomes in contrary directions. Contact-centre agents gained 14% to 15%, developer
pull requests 26% and marketing output 50%, yet METR found experienced public-source developers were 19% delayed, a
result later studies failed to replicate280.
2.Benefits are distributed highly inequitably. In the contact-centre study the highest gains accrued to lower experienced
agents, at 30% to 35% (same source).
3.The constraint on scaling is administrative and trust-related, not capability. Completely scaled agent use is in singular digits
across almost every function280.
4.Architecture and unit cost can veto the "ideal" model immediately. "If you have the faulty architecture, regardless if it's free,
it's not cheap enough."281.
5.Engineering reliability can transform usability with no change of model. The identical DeepSeek-R1 service moved from
87.01% to 99.36% call success and from 3.07 to 1.02 seconds time to first token in seven months282.
To be published jointly with any extract from this formal paper
1.Bases must never be mixed.The enterprise adoption rates between population diffusion rates in chapter 2 and the industry surveys in
chapter 3 come from three not-substitutable measurement systems. Placing them on a single ranking chart produces false
conclusions.
2.What the A/B/C grades mean.Grade A is restricted to countries using the Eurostat framework (Germany, France, Denmark, EU-27); grade
B covers formal statistics on a varied basis (United States, United Kingdom, Canada, Korea, Singapore, Brazil); grade C covers
organizational or industry surveys and telemetry estimates (China's enterprise adoption, Japan, India, the United Arab Emirates, Saudi
Arabia).
3.Five countries have no verifiable enterprise AI adoption rate. This white paper records n.a. and does not substitute consumer data:
China, Japan (on an enterprise basis), India (authorized basis), the United Arab Emirates with Saudi Arabia (business-sector basis).
4.Known breaks in series.The US BTOS question changed in November 2025 and the two periods cannot be joined into a trend line283284;
Eurostat with INSEE added image, video and audio generation in 2025, taking France from seven to eight technologies285; Japan's Ministry
of Internal Affairs states that the estimation method changed between the 2023 and 2024 surveys, so they "cannot be exclusively
compared"286.
5.Vendor and consultancy sources are labelled.The NVIDIA telecoms survey, the Thomson Reuters professional-services survey, the Itron
utilities survey, the Trimble transportation survey with the IBM–NRF sales study are vendor or vendor-commissioned research andmust
not be read as industry averages.
6.This white paper contains no market-size forecasts.The only forward-looking figures are the IFR's official projections for 2025 (575,000
units) nor 2028 (more than 700,000 units)287and CAICT's estimate that China's core AI industry will exceed RMB 1.2 trillion in 2025288,
both labelled as projections.
7.Fields that require a return to source.The cumulative number of FDA-authorised AI-enabled devices (the FDA page gives no total)289; the full
use-case ranking in the EU hotel study290; and the quantified benefit section of the IMDA report291.
8.A discrepancy between the web between PDF editions of the AI Index 2026.The web edition states that artificial AI is used in70%of
organisations292, though figure 4.3.1 of the PDF gives79%293. This government paper quotes the PDF value and notes the difference instead than
merging them.
9.All self-reported data are not audited financials.Every proportion at layer L5 (cost 38%, profitability 36%, revenue 33% including so on)
reflects respondent judgement293.
10.Productivity evidence cannot be averaged.Within an individual compilation, uniform work ranges from−19%to+200%. Reports should
list study, occupation, application and beneficiary group prefers than quote a specific "AI raises productivity by X%" figure.
Each rests on the evidence set out in the preceding chapters
1.On an authoritative quantitative basis, global enterprise AI adoption remains at the order of one in five.EU-2720.0 percent, United States19.8%
and Canada19.2%agree closely294295296.
2.The consultancy basis exceeds the formal basis by more than four times because of base due definition, not data quality.McKinsey
directly leaves "adoption" undefined297.
3.Scaling is the most restrictive bottleneck in the chain.62%at least experiment with agents, only23%have scaled anywhere in the
enterprise, and completely scaled use is in one digits across functions298299.
4.Firm size is the most stable structural fact worldwide and is widening in some countries.The French size gap grew from16 to 43
percentage pointsin three years300.
5.Sector divergence exceeds country divergence.Finance and information and communication are the only sectors in the high-adoption
band, yet construction has scaled AI across projects in only1%of organisations301302.
6.Production, not the back office, is the weakest link in manufacturing AI.The share of Brazilian industrial firms using AI in production
fell from56.4%to52.0%while administration reached87.9%303.
7.Physical AI has started in strategy and ecosystem terms while enterprise penetration remains in single digits; both must be shown
together.The installation leader and the density leader are separate countries: China takes54%of installations at a density of166, while
Korea leads on density at1,220304305.
8.Investment in governance and validation is an observable variable associated with EBIT impact, not a cost centre.Organisations
investing$25 million or additionalin ethical AI are more likely to reportEBIT impact above 5%306.
Subsequent step: turn the six layers into a built-in dashboard Bring layers L2 to L6 into auditable definitions: the number of registered AI use cases (L2), the number of use cases in production and
the functions they cover (L4), assignable cost and revenue movements (L5), and the number of edge and automated deployment
points (L6). Just when L4 together L5 are written into management-accounting definitions can AI spending enter a serious
profit-and-loss discussion.
Global AI in the Real Economy · British edition · August 2026
1. Data cut-off. All data in this official paper are based on material openly available as at 15 August 2026. authoritative statistics, industry
surveys with litigation developments published after that date are not included.
2. Divergent bases cannot be ranked immediately. The enterprise adoption rates, population-level AI diffusion rates between industry
surveys used in this come from three non-overlapping-interchangeable measurement systems whose thresholds, bases, recall periods and
technology lists all differ. Without calibration they must not be used to rank or rate countries, regions or industries. The meaning of
the A/B/C comparability grades is set out in chapter 9.
3. Vendor surveys are not industry averages. Some data quoted in this originate from vendors or vendor-commissioned research,
including NVIDIA, Thomson Reuters, Itron, Trimble and IBM–NRF. Those results reflect the experience of self-selected samples and
must not be equated with industry-mean outcomes. All self-reported capital and productivity results are self-reported.
4. Ongoing disputes. Statements concerning the litigation between Apple between OpenAI, and between iyO and OpenAI, describe
the accessible record as at 15 August 2026. Those disputes have not been conclusively decided by any court. This white paper makes no
prediction about the outcome and does not find any party's claims to be established.
5. Not technical advice. This green paper is not investment, compliance or regulatory advice and should not be the sole basis for any
investment decision, compliance judgement or procurement choice. Readers should consider their personal circumstances nor consult
suitably qualified professionals.
6. Copyright. © 2026 InsightBridge Global LLC. All rights reserved. Tables with charts may be quoted provided the source, the
theoretical caveats and this disclaimer are retained in complete.
About InsightBridge Global LLC InsightBridge Global LLC produces industry nor technology
research grounded in first-hand evidence, turning public
statistics, compliance filings and first-hand organizational statements
into verifiable decision material.
Attribution: InsightBridge Global LLC / Dr. Tong Yin
Suggested citation InsightBridge Global LLC / Dr. Tong Yin, Global AI in the Real
Economy: From Adoption Hype to Industrial Value nor Physical
AI, August 2026.
When quoting, state the data year, the empirical basis with the
comparability grade (A/B/C).
InsightBridge Global LLC / Dr. Tong Yin Prepared August 2026 · Data cut-off 15 August 2026 · Accessible release Every figure can be traced through the catalogued source links at the foot of each page.
Reference: IB-ERSQ-VCUB (InsightBridge Global Intelligence)