Australia, we need to talk about AI
AI capability is rising fast and its cost is falling faster, yet most Australian businesses use it shallowly, if at all. This paper sets out what the evidence says about that gap, what Australians are already building, what government and industry have shown they can do together, and where the Digital Economy Council of Australia’s AI Working Group says policy should focus.
Where we write from. Mark Monfort co-chairs the AI Working Group of the Digital Economy Council of Australia (DECA), which prepared DECA’s submission to the Joint Select Committee on Artificial Intelligence. He is Chief AI & Innovation Officer of the law firm Madison Marcus and Managing Director of Foundry Labs, its innovation arm, and he has taken part in government and industry tests of emerging technology, including the Reserve Bank’s CBDC pilot and Project Acacia.
Capability is up, cost is down, adoption is shallow. Epoch AI finds that the cost of reaching a fixed level of AI performance has fallen about 47% a quarter since 2023, or 13 times a year. About 12% of Australian businesses used AI in 2024–25, and on the closest like-for-like measure Canada was level a year ago and is now at 19.2%.
Australia can build, and has the people. Unsloth, Artificial Analysis, Build Club, Gamgee and LegalQuants all have Australian founders. Private AI investment here was US$2.52 billion in 2025, ninth in the world, against US$285.9 billion in the United States.
Government and industry have already shown they can test emerging technology together. The Reserve Bank’s CBDC pilot and Project Acacia are the model for how to do it with AI.
The Working Group’s ask: regulate the deployed system, build practical and verifiable standards on those that exist, and back commercialisation through procurement, test facilities and sandboxes.
01 · The conversation we are not having
Public discussion of AI swings between promises that it will fix everything and headlines that it will break everything. Meanwhile three things are happening that deserve more attention than either.
Capability is rising. In September, OpenAI announced that an internal system had produced a proof of finite-time blow-up for the 3D Navier–Stokes equations in the unforced case, one of the Clay Institute’s Millennium Prize Problems. It is a claim under review, not a settled result: mathematicians are still checking a 166-page manuscript, and reporting quotes Oxford’s James Maynard saying it is hard to extract human understanding from it. A few days later a group of leading mathematicians published a declaration, “A Severe Misalignment of AI in Mathematics”, warning that answers matter most when people can understand them. Whatever the verdict on the proof, the profession is now arguing about how to absorb AI results, not whether they are possible.
The same shift is visible in software. In 2026 independent developers have released free, AI-written reimplementations of professional tools, among them Filmcraft, an open-source video editor modelled on Adobe Premiere Pro and written in Rust with AI assistance. It is early and reviewers report gaps, so it shows direction, not replacement. Investors read the direction seriously: software shares fell sharply in early 2026 after AI agent products launched, with Adobe and Salesforce each down more than 25% by February by one count, though analysts disagree about whether the selloff was overdone. Closer to home, Sydney engineer Paul Conyngham, working with UNSW’s Ramaciotti Centre and RNA Institute and with AI tools including AlphaFold, produced a personalised mRNA cancer vaccine for his dog Rosie. It is a single case and she had other treatment. He went on to found Gamgee Technologies, now in Y Combinator, and said the red tape was harder than the vaccine.
Cost is falling. Epoch AI’s report “The plunging price of thought”, published on 22 September 2026, measures the actual cost of reaching a fixed score on five benchmarks covering maths, hard science and games of skill. It finds that cost has fallen about 47% a quarter since 2023, or 13 times a year. The pace varies by task, from about 7 to 10 times a year on game puzzles to 16 to 19 times on maths, and prices fall fastest soon after a level of performance first appears. The chart below uses the cost data of Artificial Analysis, the benchmarking company co-founded by Australian Micah Hill-Smith. Work that was uneconomic to attempt is now cheap to try. Total spending is rising because usage is, and frontier models keep premium prices, but the direction is clear.
Today’s price lists show the same pattern, and a new feature: a price for every kind of task. On 13 October, OpenAI’s GPT-6 Astra and Anthropic’s Claude Fable 5.1 are both listed at $50 per million output tokens, GPT-6 Sol and Claude Sonnet 5.5 at $10, and GPT-6 Luna and Claude Haiku 5.5 at $0.50. That is a hundredfold spread between the top and the small tier, and the small tier itself has become ten times cheaper than its predecessor, Claude Haiku 4.5 at $5. A task that needs the best model pays for it, and a task that does not can be done for a fraction of a cent. Choosing the right tier for each task is now a core part of building AI cheaply.
There is now a price for every task: the top score costs about 19 times a score of 50
Each dot is one model at one reasoning setting. The highest score, 58, costs $5.98 a task. A score of 50 costs $0.32, and a score of 43 can cost $0.21.
Source: Artificial Analysis LLM leaderboard (artificialanalysis.ai/leaderboards/models), Intelligence Index and cost per task, read on 13 October 2026. 109 model settings are plotted; 5 that report a cost of $0.00 are not shown. Cost per task is Artificial Analysis’s own measure and includes the tokens a model uses to finish a task, so it is not the same as the price per token. Index versions change, and scores from different versions are not comparable. Total spend is rising because usage is.
Adoption is shallow. Against that backdrop, most Australian businesses are not using AI in any depth. The Australian Bureau of Statistics found about 12% of businesses used AI in 2024–25, up from 1% in 2021–22.
02 · How Australia compares, and why the numbers differ
Other surveys give much higher figures: the National AI Centre’s SME tracker has 43% to 44% of small and medium businesses reporting some use. Both can be right because they measure different things. The ABS counts every employing business, including micro-businesses, over a year. The National AI Centre counts self-reported use at any level, including one-off use, among SMEs. Eurostat counts firms with ten or more employees. The US Census Bureau asks about the past two weeks.
Every adoption number on one chart: the headline gap is mostly definitions
Colour shows the measure. Only bars of the same colour are like for like.
Sources: ABS Characteristics of Australian Business 2024–25; Statistics Canada, Canadian Survey on Business Conditions (Q2 2025, Q2 2026); Eurostat ICT usage in enterprises 2025; OECD (Jan 2026), as reported secondhand; US Census Bureau BTOS (May 2026); National AI Centre SME adoption tracker, as reported by ai.gov.au. Not charted: UK ONS (headline reported as 29% to 35% with an unclear size base) and the auDA, RBA, MYOB and Deloitte surveys, which are known to us only through secondary sources.
The closest like-for-like comparison is Canada, whose national statistics agency also counts employer businesses of every size and asks about the previous 12 months. A year ago the two countries were level, at 12% in Australia (2024–25) and 12.2% in Canada (Q2 2025). Canada is now at 19.2%. The next ABS reading is not yet out, so we cannot say how Australia compares today.
The closest like-for-like comparison: Canada, and we were level a year ago
Both count employing businesses of every size and ask whether AI was used in the past 12 months.
Statistics Canada samples establishments with employees (21,105 sampled, 9,251 responded in Q2 2026) and asks whether the business used AI to produce goods or deliver services in the last 12 months. The ABS covers about 7,000 businesses over the 2024–25 year, fieldwork October 2025 to February 2026. Neither page gives a formal definition of AI, so wording differences remain. The next ABS reading (2025–26) is not out, so it is not yet possible to say how Australia compares with Canada's 2026 figure.
Where the gap is clearer is in larger firms: 35% of large Australian businesses use AI against an OECD average of 52%, while small firms are at about 11% against 17.4%. Adoption is also higher where businesses innovate: the ABS found 20% of innovation-active businesses use AI against 6% of the rest. The Nordic countries that lead Eurostat’s table show what is possible: Denmark at 42%, Finland at 37.8%, Sweden at 35%.
Australian large firms are behind their OECD peers; the small-firm gap is narrower
Share of firms using AI, by size.
Sources: ABS 2024–25 (small and micro about 11%, large 35%; medium 22% not shown); OECD January 2026 release as reported secondhand (small 17.4%, large 52.0%). Size cut-offs and the population covered differ between the two, so treat it as indicative.
03 · The prize, and the part Australia must build
The Productivity Commission’s December 2025 report reviewed studies putting AI’s effect on multifactor productivity over the next decade at between 0.5% and 13%, and judged that gains above 2.3% are likely. That implies roughly 4.3% higher labour productivity, or about $116 billion of additional GDP, which the Commission itself describes as a rough estimate with considerable uncertainty. Treasury’s 2026 Intergenerational Report likewise treats AI as likely to support its long-term productivity assumption.
The growth prize is large and uncertain
AI's effect on multifactor productivity over the next decade (Productivity Commission, December 2025).
Source: Productivity Commission, Harnessing data and digital technology, final report (released 19 December 2025). Treasury's 2026 Intergenerational Report also says AI is likely to support its long-term productivity assumption; Its scenario range is not charted because it was available only through commentary.
That prize is only available if the technology is deployed, and it matters who builds it. Australia ranked ninth in the world for private AI investment in 2025 at US$2.52 billion, against US$285.9 billion in the United States, and had 38 newly funded AI companies against 1,953 in the US, according to the Stanford AI Index. Business R&D on AI is growing, up 142% in two years to $668 million, but is still about 2.7% of business R&D. In venture funding, Cut Through reports that 61% of the $5.4 billion raised by Australian startups in 2025 went to startups with an AI offering, with AI the top funded sector at $1.0 billion, though early-stage deal flow in the second quarter of 2026 was the weakest since 2020.
Australia is ninth for private AI investment, at US$2.52B
Global private investment in AI by country, 2025, US$ billion (US shown in the note).
Source: Stanford AI Index 2026, Figure 4.2.11 (Quid). Australia also had 38 newly funded AI companies in 2025, against 1,953 in the US (Figure 4.2.8). Cut Through reports $5.4B of total startup funding in 2025 across 390 deals, with 61% going to startups with an AI offering and AI the top funded sector at $1.0B; those are Australian-dollar, all-sector venture figures and are a different measure.
Business R&D on AI grew 142% in two years
AI business R&D expenditure, A$ million.
Source: ABS, 'AI now fastest growing area for business R&D', 22 August 2025. The 2022–23 value is not in the release text, so the chart shows the two end points.
Our view is that Australia needs its own modelling capability, and the people to build it. The talent is already here. Unsloth, which makes it much faster and cheaper to fine-tune open models, was founded by Sydney brothers Daniel and Michael Han, who moved to San Francisco through Y Combinator. Artificial Analysis, whose independent ratings of AI models and providers are used across the industry, is co-founded by Micah Hill-Smith and has a Melbourne office. Build Club, started by Annie Liao in Sydney, has grown a community of AI builders that it says now spans more than 50 cities. LegalQuants, co-founded by Raymond Sun and Jamie Tso, trains lawyers to build their own tools, in a legal AI market where the two leading vendors are reported to be valued at about $21 billion between them (reports differ on the latest rounds). Gamgee is commercialising a personalised cancer vaccine approach starting with dogs, with clinical trials in Australia.
Where Australian builders already sit in the AI stack
A map, not a data chart: examples are named so the room can see who they are.
Examples are from public sources; the placement of each is the author's judgement. Funding and scale claims are left out where sources conflict.
Hear from them
Short interviews with five of the Australians named above. Each video loads from YouTube only when you press play.
The government has begun to back buyers as well as builders. The Buy Australian AI Partnership, run by Stone & Chalk with the National AI Centre as principal sponsor and ANZ, Commonwealth Bank, NAB, Westpac and Cuscal as founding enterprise partners, aims to help Australian AI companies reach large enterprise buyers, starting with banking, insurance and superannuation. The government’s own framing is that Australia has built more than 1,500 AI companies and has too few Australian buyers.
04 · What government and industry have shown they can do together
Australia has already shown it can coordinate with industry, and with regulators in the room, to test emerging technology. In 2022–23 the Reserve Bank and the Digital Finance CRC ran a central bank digital currency pilot in which the RBA issued a limited-scale pilot CBDC, a real legal claim on the RBA. Its use cases included construction payments, an area NotCentralised, where Mark then worked, took part in.
Project Acacia then extended the work into wholesale markets. Its final report, published in May 2026, records that participants used pilot wholesale CBDC across private and public third-party platforms, and that regulators observed the work. NotCentralised piloted a tokenised structured finance security backed by loans, issued on a bond exchange, with payment flows settled through a private money token backed by wholesale CBDC. In a collateralised-loan use case, repayment, loan discharge and release of collateral were combined into a single atomic transaction, so collateral could be released instantly and reused, removing delays that are common today. The report is balanced and records the limits: it notes limited central bank appetite for permissionless platforms where real legal claims on the central bank are involved.
This is not about blockchain for its own sake. It shows a method: government, regulators and industry agreeing a scope, testing something real, and publishing what worked and what did not. That is the method AI needs. It also gave the AI debate two ideas it sorely needs.
- Digital attestations. Blockchain work pushed a long way on proving who did what, on what, and with whose authority. AI has the same problem: when a system acts, who
deployed it, under what authority, with what model, on what inputs? The Working Group’s submission calls for the same shift, from narrative self-declaration to attested
evidence of what a system did and who approved it.
“It is between narrative self-declaration and attested evidence of what a system did, on what inputs, and who approved it.”
DECA submission, 14 September 2026 - Cheap, programmable, instant settlement. Transactions that settle in an instant, with the rules built in, can be safer than today’s multi-step processes, and they suit software agents that need to pay for things within limits a person has set.
05 · Where the Working Group says policy should focus
DECA’s AI Working Group, which Mark co-chairs, represents mostly deployers and users of AI, not frontier model developers. Its submission to the Joint Select Committee puts three themes at the centre.
“Most of our members are deployers and users of AI rather than developers of frontier models.”
DECA submission, 14 September 2026
Deployment-focused regulation (Question 3). Regulate the deployed system and the decision pathway, with responsibility attached to the people and organisations that control it. When a system behaves unexpectedly, ask what actually happened: the model, the software around it, the permissions it was given, or a failure of oversight. Each answer points to a different responsible party and a different safeguard.
“AI harm arises at deployment, at the level of the system rather than the model.”
DECA submission, 14 September 2026
Standards (Question 4). Build on the Voluntary AI Safety Standard and AS ISO/IEC 42001, and make governance practical, measurable, independently verifiable and proportionate to risk.
“obligations that cannot be independently verified produce documentation rather than safety.”
DECA submission, 14 September 2026
In our own testing at Madison Marcus through Foundry Labs, the lesson has been that different kinds of check catch different kinds of error, so a standard should say what each check does and does not cover. The detailed research is on foundrylabs.com.au.
Innovation opportunities. Australia can develop, commercialise and govern valuable AI applications, and should keep selective sovereign capability where it has an advantage, including the training tools and evaluation that Australians already lead. The practical levers the submission names are procurement, test facilities, sandboxes and export.
The National AI Plan released in December 2025 sets goals of capturing the opportunity, spreading the benefits and keeping Australians safe, and relies on existing law. The Working Group’s emphasis is on coordination: a published map of which instrument and portfolio addresses each AI issue, before further legislation.
06 · What we do at Madison Marcus and Foundry Labs
Madison Marcus is a national law firm that has to meet professional duties of confidentiality, accuracy and supervision. Foundry Labs, its innovation arm, builds and tests AI for the firm and its clients. Our practice is simple. We test before we deploy, on synthetic and public material. Where real client information is involved, the data and the model calls stay in Australia. We keep the ability to change models when a better one appears, and we check each model and route rather than assume. Every output must be traceable to a source a person can check. We publish what we find, including what does not work, and the detailed research is on foundrylabs.com.au.
07 · The Working Group’s ten recommendations, in brief
| Recommendation | What it asks |
|---|---|
| 1. A coordination map | Publish, before any AI legislation, a map of which instrument and portfolio addresses each AI issue, and how Commonwealth and State measures interact. |
| 2. Red tape reduction as AI policy | Measure regulatory complexity, discretion and the number of approvals between a firm and a new activity, to give a checkable baseline. |
| 3. Regulate the deployed system | Direct obligations at the deployed system and the decision pathway, on deployers in proportion to the consequence of the decision. |
| 4. Standards built on existing ones | Build any mandatory Australian Standards on the Voluntary AI Safety Standard and AS ISO/IEC 42001, with the scope, threshold, commencement and binding mechanism of each obligation stated separately. |
| 5. Agent identity and attribution | Address agentic AI expressly, starting with who deployed an agent, under what authority it acts and what model drives it. |
| 6. Verifiable disclosure | Design disclosure and reporting for independent verification, and pilot cryptographically attested reporting of AI decision lineage before mandating a schema. |
| 7. Selective sovereign capability | Concentrate sovereign AI investment where Australia has comparative advantage, and connect research to commercialisation through procurement, test facilities, sandboxes and export. |
| 8. Copyright | Resolve the uncertainty quickly, with transparency on training-data governance and obligations on those who control training rather than downstream businesses. |
| 9. Procurement and contract terms | Publish guidance and model clauses covering supplier due diligence, post-deployment changes, audit and information rights, incident notification, liability, and exit and portability. |
| 10. Assurance capability | Address the shortage of people able to govern, audit and certify AI, and treat assurance capability as a sovereignty question. |
“to convene members to support a verifiable disclosure pilot, and to contribute to work on agent identity standards”
DECA submission, 14 September 2026
Method
- Figures come from the sources named under each chart. Adoption figures use different definitions and are not directly comparable except where stated, and several secondary-source figures are marked as such.
- Claims under review. The Navier–Stokes result is a claim being checked. Legal AI vendor valuations are reported, not confirmed. Build Club’s “50 cities” is the organisation’s own figure.
- Quotes from the DECA submission were checked against its text.
Sources: ABS (adoption 2024–25; AI business R&D); Statistics Canada; Eurostat; OECD; US Census Bureau; National AI Centre; Productivity Commission; Epoch AI, The plunging price of thought (22 September 2026); Stanford AI Index 2025 and 2026; Cut Through Venture; RBA and Digital Finance CRC (2023 pilot report; Project Acacia final report, May 2026); Stone & Chalk and the Department of Industry; OpenAI, Quanta, NPR and TechCrunch (Navier–Stokes and the mathematicians’ declaration); GitHub (Filmcraft); company and press reports for the people named.
General information only. Nothing here is legal advice.