Executive summary
Kitbag Consulting and KB Technology work with Australian public and private sector entities on digital initiatives and AI governance. KB Technology has developed DOLIUM, an Australian owned, IRAP assessed System of Work listed on the Digital Transformation Agency Software Marketplace Panel. This submission draws on that engagement, on submissions already made to the Joint Select Committee on Artificial Intelligence, the Digital Transformation Agency and the Department of Home Affairs and on direct briefings to the Department of Finance and other agencies on the same subject.1
The consultation paper addresses two related but distinct aspects of Australia's AI capability. The first is physical: the infrastructure, energy, water, land, community and workforce requirements associated with large data centres. The second concerns conditions for frontier AI training in Australia, including national security and safety, local capability, skills, research and innovation, and support for Australian creators and rightsholders. Both are necessary. Neither however, directly addresses the organisational operating layer through which AI produces economic and governance outcomes.
This submission does not seek to duplicate those requirements or to address every question raised by the consultation. It identifies a cross-cutting omission: the operating logic of the organisations that will use this infrastructure and these models, meaning the decision rights, policies, processes and institutional knowledge that govern how government and enterprise function.
This submission argues that capturing this layer, in a sovereign Australian governed structure is a precondition for realising the paper's own stated objectives, not an adjacent policy interest.
Six findings follow, each tied to a specific part of the paper.
- First, the broader Australian AI productivity dividend should not be made contingent on attracting frontier training activity. The consultation correctly identifies genuine strategic, economic and national security benefits from frontier training in Australia. However, the substantial majority of routine government and enterprise AI use does not require frontier reasoning. This creates an opportunity to pursue AI productivity through smaller, open-weight or otherwise substitutable models operating against structured organisational context, while separately pursuing the national benefits of frontier training.
- Second, capturing operating logic can reduce the token, compute and energy footprint of AI use under either a hosted or a self-hosted deployment model. This creates a demand-side efficiency opportunity that complements the consultation's proposed supply-side requirements for energy, water and cost management at large data centres.
- Third, a structured operating logic layer can make the paper's proposed safety and security objectives more traceable in practice. Model-level incident reporting can explain what went wrong within a model. It does not necessarily establish what the model was authorised to do inside an organisation, what process it was executing, what decision authority applied or which human was accountable for the resulting action.
- Fourth, without that layer government and enterprise are exposed to two related risks that are not adequately addressed by model-level disclosure alone: the progressive embedding of organisational operating logic into vendor platforms and the potential for foreign third parties to gain structural visibility into how an organisation operates.
- Fifth, sovereignty should be considered in the eventual standards as including effective Australian control over operating logic, rather than being defined solely by data residency or the physical jurisdiction in which a model is hosted or trained. Data residency is necessary in some circumstances. It is not by itself sufficient to establish operational sovereignty.
- Sixth, the cost of getting this sequence wrong can be material. The Commonwealth and Australian jurisdictions have already experienced the difficulty and cost associated with major enterprise platforms that become deeply coupled to organisational requirements. Those examples do not establish that architecture was the sole cause of those outcomes, but they demonstrate why operating requirements should be represented independently before major platform investments are made.
The recommendations at the close of this submission set out how the AI standards can recognise each of these findings.
