Australia, China, Europe (EU27), Canada and the United States, measured against the most recent official statistics published as at 6 September 2026.
Verified 6 September 2026 Every figure in this report has been checked against a named source. Eighteen of the nineteen cited URLs were fetched and confirmed to contain the claim attributed to them; the exception is the Stanford AI Index chapter page, which returns only navigation to an automated fetch, so its figures were confirmed through independent secondary reporting instead. Section 10 records the status of each claim individually.
Two verification passes were run. The first found one load-bearing error (the survey scope for Australia and Canada was stated wrongly), which had inverted a conclusion. That is corrected, and Sections 1, 4, 8 and 9 have been re-derived from the corrected frames. The second pass confirmed the Department of Employment and Workplace Relations figures against two independent outlets and restored three claims that had been provisionally deleted.
Two caveats travel with this document:
Subject to those two caveats being carried with any republication, this document is fit to be published under the COAD name.
Verification note, 14 September 2026 A New York Times feature of 6 September 2026 characterised China as adopting artificial intelligence faster than almost any other country. That characterisation was tested and could not be verified. The article gives no source for it, and no national Chinese AI adoption measure is available from any primary statistical or ministerial source; searching returns only commercial aggregator content, which this project’s sourcing standard excludes. The Ministry of Industry and Information Technology figure held in this comparison covers manufacturers above a turnover threshold and is expressly not a national rate. This is recorded as corroboration of the comparison’s central finding, that cross-jurisdiction AI adoption rates are not comparable and should not be ranked, rather than as a criticism of the article. (Note added to this report on 30 September 2026. The article’s wording is described rather than quoted, because the newspaper’s site cannot be reached to check it.)
There is no internationally harmonised measure of business AI adoption. The five jurisdictions each publish a rate, and those rates are not measured on the same basis. The differences are large enough to reverse a country ranking:
Every chart in this report states the survey frame beneath it. Section 9 sets out the full comparability matrix. Cross-jurisdiction rankings drawn from Figure 1 alone should be treated as indicative, not as measured differences.
Each bar is the most recent figure published by the responsible national or supranational statistical authority. The frame beneath each bar is what that authority actually measured.
Australia, Canada and the United States all measure employing businesses, so their three rates (12.0, 19.2 and 19.8 per cent) sit on broadly the same frame. The roughly seven percentage points separating Australia from the other two is therefore substantially a real difference in behaviour, not an artefact of how the surveys are drawn. The European Union's 20.0 per cent is the outlier in frame terms, because it excludes every enterprise with fewer than 10 employees and so sits above what an all-employing-business measure would return. Revised 6 September 2026 An earlier version of this report stated that the Australian and Canadian surveys covered non-employing businesses. That was wrong; both cover employing businesses only. The correction narrows the frame explanation and widens the measured Australian shortfall.
Three separate Australian measures were published within twelve months of one another, and they do not agree, because they measure different populations and different thresholds of use:
They are three points on a depth-of-use gradient rather than contradictions to be settled by picking one, and the gradient is itself a finding. Section 4 sets it out.
In every jurisdiction that publishes a size breakdown, adoption rises sharply with employment size. The width of that gap is a better cross-country signal than the headline rate, because it is less sensitive to how the survey frame is drawn.
Europe shows the widest gradient: 55.0 per cent of large enterprises use AI against 17.0 per cent of small enterprises, a range of 38 percentage points within one harmonised survey. Australia's large-business rate of 35 per cent sits well above its headline figure and rose from 9 per cent in 2021–22, while medium businesses moved from 3 to 22 per cent over the same period. Canada shows the narrowest gradient of the four, at 27.8 per cent for firms with 100 or more employees against 19.9 per cent for firms with one to four employees.
For a policy audience the implication is direct. Adoption is concentrated in the firms that employ the most people per enterprise and hold the most capital, and it is thinnest among the small and micro businesses that make up the overwhelming majority of Australian enterprises by count. Any distributional consequence of AI-driven productivity gains will therefore be felt first through large employers.
Only two of the five jurisdictions publish a repeated series on a stable definition long enough to plot a trend. Both approximately tripled in three to four years.
Where the trend can be measured, it is steep and it is recent. The European Union moved from 7.7 per cent in 2021 to 13.5 per cent in 2024 and 20.0 per cent in 2025: a gain of 6.5 percentage points in a single year, which is larger than the entire increase over the preceding three years combined. Canada tripled from 6.1 per cent to 19.2 per cent across three annual observations of the same quarter.
The absence of an equivalent Australian series is itself a finding. Australia cannot presently produce a defensible year-on-year adoption trend on an unchanged definition, which limits the evidence base available to Australian policy makers relative to their European and Canadian counterparts.
The most consequential finding in this report is not how many businesses have adopted AI, but how few have deployed it at a depth capable of changing output or headcount.
The only comparable depth measure for China comes from a private survey, not an official one. The Deloitte China and University of Hong Kong China AI Adoption Index 2026, published February 2026, surveyed a little over 100 senior executives across mainland China and Hong Kong and reported that more than one-third of organisations remain in exploration, 56 per cent are in limited implementation, 23 per cent report measurable financial impact and 4 per cent describe their deployment as transformational.
Confidence: low A sample of roughly 100 executives, self-selected and skewed to large organisations, cannot support a national rate. The figures are reproduced here because no better source exists, and should be cited as an executive-perception survey rather than as a measure of Chinese enterprise behaviour.
The individual numbers are each frame-weak, but the shape recurs in every measure that reports depth, and that recurrence is the finding. Adoption in the sense of ‘someone in the business uses an AI tool’ is now common. Adoption in the sense of ‘AI is embedded in how the business produces its output’ remains rare: under 10 per cent in Australia on Treasury's measure, 8 per cent in Canada, and 7 per cent of Australian small and medium enterprises making broad use of AI on the Assistant Minister's figure of 15 May 2026. Four sources, four different instruments, four deep-use figures between 7 and 10 per cent.
The any-use figures are where the frames diverge sharply and where care is required. Australia has three: 12 per cent of all employing businesses (ABS), 43 per cent of small and medium enterprises (National AI Centre and Fifth Quadrant), and about two-thirds of businesses (Treasury). The first two are documented and reconcilable, because a national all-business rate and an SME-panel rate measure different things. Treasury's two-thirds is neither documented nor reconcilable with either, and it is hatched in Figure 4 for that reason. The safe reading is that the deep-use ceiling is well established across sources, while the any-use denominator is not.
Within the adopting group there is a countervailing movement. The National AI Centre records broad adoption among small and medium enterprises reaching its highest level in seven months during the December 2025 to February 2026 quarter, while one-off, limited use declines. Those who have committed are deepening; the majority have not committed.
Both the productivity dividend that funds a citizen dividend and the labour displacement that justifies one flow from deep deployment, not from trial use. Deep deployment sits between 7 and 10 per cent in every Australian and Canadian measure available. Neither effect has therefore been generated at scale. This supports two propositions that need to be held together rather than traded off: the fiscal base for a dividend is not yet present at the size sometimes assumed, and the displacement it is designed to answer is still ahead rather than behind. The policy question is one of sequencing and readiness, not of whether the transition has already occurred. Note that this argument rests on the deep-use figures, which are corroborated across four instruments, and not on Treasury's undocumented two-thirds.
Inference, not measurement This paragraph is analytical interpretation by the author of this report. It is not a finding of any cited statistical authority.
Adoption rates converge across the five jurisdictions. The capacity to supply AI does not: it is concentrated to a degree that adoption statistics conceal entirely.
Two figures are confirmed on an annual 2025 basis by the Stanford Institute for Human-Centered Artificial Intelligence AI Index Report 2026:
Unresolved Equivalent annual 2025 private-investment figures for Australia, Canada and the European Union could not be confirmed on the same basis during preparation of this report, and are therefore not charted. The nearest confirmed figures are cumulative across 2013–2024 (AI Index Report 2025): United States US$471 billion, China US$119 billion, United Kingdom US$28 billion, Canada US$15 billion, Germany US$13 billion, France US$11 billion, Australia US$4 billion. Annual and cumulative figures must not be placed on the same axis. Confirming the annual series for the remaining three jurisdictions requires the underlying AI Index data appendix.
The AI Index itself notes that China's private-investment figure understates total Chinese AI spending, because the Chinese state channels resources through guidance funds and state-initiated investment vehicles that do not appear in private-capital databases. The 23-to-1 ratio is a ratio of private capital, not of national effort.
Australia's position is the significant one for COAD purposes. Australia sits within a regional grouping holding 1.6 gigawatts of the world's 122.2 gigawatts of installed data-centre capacity, and recorded approximately US$4 billion of cumulative private AI investment across 2013–2024 against the United States' US$471 billion. Australia is, on the available evidence, a consumer of AI capability produced elsewhere rather than a producer of it. The economic returns to supplying AI accrue overwhelmingly outside Australian jurisdiction, while the labour-market consequences of using it fall inside it.
No jurisdiction reviewed has recorded mass AI-driven job losses. Two have recorded measurable divergence in the growth rates of exposed and unexposed occupations.
The Australian result is a divergence rather than a decline. Employment in AI-exposed occupations grew 5.6 per cent over the period against 9.5 per cent in the least-exposed occupations, leaving exposed employment approximately 2 per cent below the level implied by its pre-ChatGPT trend. Occupations examined included registry clerks, telemarketers and administrative roles. Set against that, employment in software development, an occupation frequently named as exposed, rose 25 per cent, and employment among workers aged 20 to 24 grew slightly faster than among older workers.
Internationally the picture is not uniform. The Stanford AI Index Report 2026 records that employment for software developers aged 22 to 25 in the United States has fallen by nearly 20 per cent from its 2024 level, which points in the opposite direction to the Australian software result and suggests the effect may be concentrated at the entry level rather than across an occupation.
The EY-Parthenon modelling should be read with its own caveat attached: it projects 36,000 to 44,000 full-time equivalent positions supported and multifactor productivity gains of 2.0 to 2.4 per cent by 2036, with construction the largest growth sector and agriculture and mining contracting. It is a modelled scenario, not an observation, and it assumes adoption levels that the ABS and Treasury evidence in Sections 1 and 4 indicate Australia has not yet reached.
Two of the five jurisdictions have set a quantified national target for how widely AI should be used by a given date. Australia has not.
| Jurisdiction | Target | Instrument and content |
|---|---|---|
| China | Yes | State Council AI Plus initiative, issued 27 August 2025. Penetration of new-generation intelligent terminals and AI agents to exceed 70 per cent by 2027 and 90 per cent by 2030, with comprehensive entry into an intelligent economy and society by 2035, across six named application areas. |
| Europe (EU27) | Yes | Digital Decade policy programme. Three out of four EU companies (75 per cent) to use cloud computing services, big data or artificial intelligence by 2030. The combined measure stood at 63 per cent in 2025, twelve percentage points short; AI alone stood at 20.0 per cent. |
| Australia | No | National AI Plan, Department of Industry, Science and Resources, 2025. Sets qualitative goals and identifies more than $460 million in existing funding, but contains no quantified national business-adoption target or date. The plan does record a regional divide: 29 per cent of regional organisations are adopting AI against 40 per cent in metropolitan areas, with 26 per cent of regional businesses not aware of AI. |
| Canada | Not identified | No quantified national business-adoption target was identified in the sources reviewed for this report. This is an absence of evidence rather than confirmed absence of a target. |
| United States | Not identified | No quantified federal business-adoption target was identified in the sources reviewed for this report. This is an absence of evidence rather than confirmed absence of a target. |
The Chinese and European instruments differ in kind. China's target is a diffusion target set by the state and directed at penetration of terminals and agents; Europe's is a capability target embedded in a monitored policy programme with annual reporting against trajectory. Both create an accountable number. Australia's National AI Plan creates none, which means there is no published benchmark against which Australian business uptake can be judged as adequate or inadequate.
The following five propositions are the author's analysis of the evidence assembled above. They are inference, and are marked as such. They are not findings of any cited statistical authority. They were re-derived on 6 September 2026 after a verification pass corrected the survey frames used in Sections 1 and 9.
Australia, Canada and the United States all survey employing businesses. On that common frame Australia records 12.0 per cent against Canada's 19.2 and the United States' 19.8, a shortfall of roughly seven percentage points that cannot be explained away by how the surveys are drawn. The European Union's 20.0 per cent sits on a narrower frame that excludes enterprises under 10 employees, so the true Australia–Europe gap is smaller than the headline figures suggest, but it runs in the same direction. An earlier draft of this report attributed the entire spread to survey frame. That was wrong, and correcting it strengthens rather than weakens the case for policy attention.
The Australian evidence shows divergence in growth rates, not contraction: exposed occupations grew 5.6 per cent against 9.5 per cent for unexposed, leaving exposed employment about 2 per cent below its pre-ChatGPT trend. That is a measurable deficit, but the Department itself repeatedly cautioned it is not proof of an AI effect, since several exposed occupations were already weakening. A COAD case built on displacement that has already occurred is not supported by Australian official data. A case built on displacement that the deployment trajectory makes probable is supportable, and is the stronger construction.
Deep deployment sits between 7 and 10 per cent across four separate Australian and Canadian instruments. This weakens any near-term estimate of an AI-generated productivity dividend available to fund a citizen dividend, and it should be stated rather than avoided, because Treasury has already published it. Conceding it strengthens the credibility of the remainder of the case. Note that the robust part of this finding is the deep-use ceiling; Treasury's two-thirds any-use figure is undocumented and should not be relied on.
Approximately US$4 billion of cumulative private AI investment across 2013–2024 against the United States' US$471 billion, and 1.6 gigawatts of data-centre capacity across Australia and New Zealand combined against a global 122.2 gigawatts, indicate that the returns to supplying AI accrue offshore while the labour-market adjustment costs fall onshore. This asymmetry is the strongest available structural argument for a domestic distributional mechanism, and it rests on infrastructure and capital data rather than on contested labour-market projections.
China and the European Union have published quantified diffusion targets against which progress can be measured. Australia's National AI Plan has not. With proposition 1 established, there is now a measured shortfall for such a target to address. Advocating for a published national adoption and depth-of-deployment target, and for an ABS series capable of measuring it on an unchanged definition, is a specific, low-cost policy ask that is independent of the dividend proposal itself and that would improve the evidence base for it.
What each authority measured, and what that means for cross-jurisdiction comparison.
| Jurisdiction | Rate | Instrument | Population in scope | Reference period | Comparability |
|---|---|---|---|---|---|
| Australia | 12.0% | ABS Business Characteristics Survey 2024–25 | All employing businesses in the Australian economy (exclusions apply). Non-employing businesses are out of scope. | 2024–25 financial year; collected Oct 2025 – Feb 2026 | Comparable with Canada and the United States. Understates relative to the Eurostat 10+ employee frame, so the Australia–EU gap is smaller than the headline figures imply. |
| Europe (EU27) | 20.0% | Eurostat, Use of AI in enterprises | Enterprises with 10 or more employees, selected economic sectors only | 2025 | Overstates relative to an all-business frame. Highest internal consistency of the five. |
| United States | 19.8% | Census Bureau, Business Trends and Outlook Survey | Employer firms, nationally representative, biweekly | Reference date 3 May 2026; range 17–20% since Dec 2025 | Broadly comparable with Canada. Higher frequency than any other series here. |
| Canada | 19.2% | Statistics Canada, Canadian Survey on Business Conditions | Business establishments with employees; 9,251 responding of 21,105 sampled. Non-employing businesses are out of scope. | Q2 2026; collected 1 Apr – 6 May 2026 | Directly comparable with Australia and the United States: all three measure employing businesses. The most defensible comparison in this report. |
| China | >30% | MIIT ministerial statement, 5 March 2026 | Manufacturing enterprises with annual main business turnover of at least ¥20 million only | End of calendar 2025 | Not comparable. Sector-restricted and size-restricted. No national all-business survey is published. Do not treat as a national rate. |