The Claude Nation, and the Institutions That Never Arrived
Australia is number one in the world at using AI. Our policy stack is still writing the risk register.
On 26 June, Anthropic updated its Economic Index with a number that ought to have rearranged the front pages. Australia sits at 6.40 on the Anthropic AI Usage Index. Rank one of one hundred and twenty-one countries. Australians use Claude at more than six times what our working-age population would predict, ahead of Singapore, Switzerland, Luxembourg and New Zealand (Anthropic Economic Index).
The follow-up report, published a fortnight later and christened Cadences, added the texture. Of every hundred Australian Claude conversations, 54.5 stay in what Anthropic calls augmentation, where the human stays in the loop and the model plays collaborator. The global mean is 51.4. In Mongolia, Zimbabwe and Uzbekistan, the pattern inverts and automation wins (IT Brief Australia; Forbes Australia).
We are, on the numbers, the world's most enthusiastic co-workers with a large language model. And almost none of our institutional machinery is designed to notice.
The gap between practice and provisioning
Consider what the Anthropic data actually captures. It sees homework, business operations, self-presentation writing, promotional copy, workplace writing, instructional design, editing, wellness advice. It sees Cowork and Claude Code and chat. It maps tasks to O*NET occupations and it prices the cognitive work in tokens. What it cannot see is the Australian ledger those tasks feed into.
That ledger is largely blank.
The Productivity Commission's Harnessing Data and Digital Technology final report, released 19 December 2025, argued that AI-specific regulation should be a last resort and that existing frameworks can be adapted. Its follow-up inquiry, Making the Most of the AI Opportunity, concluded that Australia's near-term play is downstream of the frontier: adapting general-purpose models to local use cases and implementing them through digitised firms and SaaS (Productivity Commission). Correct as diagnosis. Incomplete as strategy. The Commission is describing a policy posture for adoption we have already, quietly, achieved.
At the Commonwealth layer the settings are catching up rather than leading. MYEFO 2025-26 committed $30 million over four years to the AI Safety Institute and $166 million to GovAI Chat, a public-service chatbot (Good Ancestors newsletter). On 23 March 2026 the government released formal expectations for data-centre and AI-infrastructure developers under the National AI Plan; Defence issued a binding responsible-AI policy the same month; and the Privacy Act's automated-decision-making obligations arrive in December (Trusenta). Anthropic itself has signed a memorandum of understanding with the government and is "exploring adding local capacity through our third-party partners in Australia, using infrastructure already in place", a direct response, they said, to data-residency requests from Australian enterprises and government agencies (Forbes Australia).
The federal AI workforce numbers tell the same story from the demand side. Federal roles explicitly requiring AI or ML skills fell from 202 in 2022 to 179 in 2025, but the share of AI-linked vacancies climbed from roughly one in one hundred and twenty in 2020 to one in fifty-five by Q1 2026 (awardedtenders.au). The absolute headcount is shrinking; the ratio inside a shrinking APS is not. Neither number is remotely commensurate with a country whose civilians already run the world's most intense per-capita AI usage.
There is a name for this configuration. It is not innovation policy. It is retrofitting.
Augmentation as an institutional signal
The augmentation share is the more interesting number. Anthropic's Cadences framing treats it as a preference. Australians like to stay in the loop. That reading is available, but a stronger one is that augmentation is what you get when the underlying tasks are shot through with tacit knowledge, regulatory nuance, professional judgment and reputational stakes. Those are precisely the features of Australian white-collar work: heavily credentialled, thin-market, deeply relational. You cannot delegate a partnership-form KPMG engagement to a model. You can, and Australian professionals evidently do, use a model to draft, refactor, argue with, and rehearse against.
The Innovation Commons frame here is not "AI is helping Australians be more productive." That framing collapses too quickly into GDP arithmetic. The frame is that a distributed cognitive commons has assembled itself, without a plan, on top of foreign infrastructure, using foreign models, priced in tokens denominated in another currency, governed by terms of service written in San Francisco. The commons exists. The question is whether Australia treats it as a substrate to build institutions around, or as a curiosity to write briefing notes about.
Three institutional gaps follow directly.
First, we do not measure it. The ABS Business Characteristics Survey asks about "use of artificial intelligence" as a yes/no. The Productivity Commission models AI adoption through generic diffusion curves. Anthropic's O*NET-mapped, token-weighted usage data is a more granular economic instrument than anything in the Australian statistical system, and it belongs to a private company. If usage this dense is not a measured input to productivity accounting, then our productivity accounting is describing a phantom economy.
Second, we do not price it. Every one of those augmentation conversations is a small transfer of value out of the Australian economy, to Anthropic (via AWS or Google Cloud), to the model's original training corpus, and to whoever holds the enterprise contract. There is no domestic settlement layer, no Australian pricing benchmark, no local model-of-record for regulated professions. The MOU with Anthropic and the Data Centre Expectations are the start of a settlement layer. They are not yet one.
Third, we do not credential around it. Almost 10 per cent of Australian Claude topics are homework (Forbes Australia). Universities and TAFEs are, at policy level, still relitigating whether this is cheating. The evidence base has moved. Assessment design has not. If the median graduate has spent three degree years in augmentation with Claude, then the credential that certifies them needs to certify the practice, not police it. It largely does not.
The Kirznerian moment we are close to missing
Kirzner's account of the entrepreneur turns on alertness, the capacity to notice a discovered opportunity before it becomes a priced-in condition. Australia has been alert as consumers. We have not, so far, been alert as an economy. The Anthropic data is a discovery: the country is running the world's most intense collaborative-AI experiment, in the open, at scale. The commons is real. The rents are, currently, foreign.
The institutional response worth designing sits along three axes. A domestic measurement layer that treats AI-mediated work as a first-class economic input, not a survey checkbox. A domestic settlement layer that gives regulated professions a model-of-record and a data-residency posture strong enough to keep sensitive workflows onshore. And an education layer that credentials augmentation as a legitimate practice, rather than treating it as an integrity problem to be defended against with detection tools that do not work.
None of that requires a national LLM. It requires the same posture Australia adopted with the corporations law, the superannuation guarantee, and the RBA payments architecture: a willingness to build settlement, measurement and standards around a practice that has already happened, rather than waiting for the practice to arrive.
What the number is really telling us
Anthropic's headline reads as a compliment. Read a second time, it is closer to a warning. Australia is number one at using someone else's tools to do our own thinking, and we have not built the institutions that would let us capture the surplus, price the risk, or credential the practice. The commons exists. The question is who is administering it.
For a country whose historical strengths run through the design of durable public institutions (RBA, HECS, Medicare, the compulsory super system), the current gap is not a technology problem. It is a nerve problem. The signal is loud. Whether the response is proportionate is a separate question, and one Canberra has not yet answered.
The number to watch is not next quarter's usage index. It is whether, twelve months from now, any Australian institution can point to a measurement, a settlement or a credential it built because 6.4 was real.
— The Editor.
P.S. tag someone who should know about this on LinkedIn