
Written by Vito Rinaldi, Managing Director at Blue Crystal Solutions
The language around artificial intelligence sovereignty has become muddied. Terms like "sovereign" and "private" are used interchangeably across boardrooms, procurement documents, and vendor pitches when discussing AI, but they describe fundamentally different things, and the distinction matters.
Sovereign refers to models that are Australian-built, trained, and owned domestically. Private, by contrast, is about where a model runs and who holds the keys. The model itself may originate from anywhere in the world, but it is hosted on Australian infrastructure inside an environment the customer controls. Blue Crystal Solutions deploys the latter and is deliberate about not conflating the two. Organisations should be wary of anyone who does.
The position is straightforward: if an organisation handles sensitive information, its technology should be ring-fenced so that data goes in and never comes out.
Settings Are a Promise; Architecture Is a Guarantee
A common refrain in enterprise adoption is that public tools are safe because the right toggles have been switched on within an enterprise subscription. This assumption deserves scrutiny. While enterprise-tier settings may offer meaningful protections, every one of them ultimately belongs to the vendor, who defines what those settings mean and reserves the right to change the terms.
Once a prompt leaves an organisation's environment, the only thing left to audit is the vendor's word. For industries handling privileged legal advice or patient health records, that is a fragile foundation on which to build trust.
The more important question is not what settings are enabled but who controls the deployment. Settings are a promise. Architecture is a guarantee.
A Matter of National and Organisational Control
This position finds an unlikely echo in federal policy. In his July keynote address, AI in Australia's Interests, Prime Minister Anthony Albanese warned against "subcontracting our national sovereignty and security to the control of foreign monopolies." It was a statement about the country's strategic posture, but the logic extends directly to the private sector.
Swap "national" for "client," and the sentence still holds. Too many law firms and health providers are subcontracting their confidentiality obligations one prompt at a time and paying for the privilege. Businesses spending $50 per user per month on OpenAI licences could direct that money toward infrastructure they own, ending the cycle of renting access to their own information.
Control over these tools is the central issue. Whether the conversation is about a nation's data sovereignty or a firm's duty of confidentiality, the principle is the same.
A Lesson from Education
The idea that private deployment is a niche concern for risk-averse IT departments does not hold up under scrutiny. One of the most compelling counter-examples comes from the NSW Department of Education.
NSWEduChat is the department's own generative assistant, hosted within its cloud environment in Sydney. According to the department, all data remains in Australia, and only the department can access it. This means one of the largest school systems in the country has made a deliberate decision that its students' interactions with the platform belong on infrastructure the department owns.
Teachers gain access to modern generative capabilities, while the department retains full authority over where the tool lives and who sees what goes into it. Ownership and hosting are where the real value sits, and they are also the two things most organisations hand to a vendor without a second thought.
If a public school system can hold that line for student data, the case for law firms, healthcare providers, and other custodians of sensitive information to do the same becomes difficult to argue against.
Not Anti-Vendor, Pro-Deliberation
This is not an adversarial stance toward any particular technology provider. OpenAI's models are effective for research, and there is a broad range of non-sensitive work that can be conducted through public platforms without issue.
Private deployment is a portfolio decision. A marketing team brainstorming campaign ideas on ChatGPT presents a very different risk profile to a solicitor summarising privileged client instructions or a clinician documenting patient notes. Organisations need to draw the line deliberately, determining which workflows belong inside the fence and which can operate outside it.
What cannot be accepted is sensitive work landing on a public tool by default, simply because no one wanted to make a hard decision or take on another project. Allowing private information into a public platform is itself a decision, and one that will carry consequences eventually.
The Simple Rule
The principle can be distilled to three words: in, not out.
Information enters a ring-fenced, privately hosted environment. It does not leave. The organisation retains control over the infrastructure, the data, and the terms under which the technology operates.
As adoption of these tools accelerates across every sector of the Australian economy, the organisations that treat deployment architecture as a strategic decision rather than a procurement detail will be the ones best positioned to harness the technology without compromising the trust their clients, patients, and stakeholders place in them.
Frequently asked questions
What is the difference between sovereign AI and private AI?
Sovereign refers to capability that remains under Australian control and jurisdiction across its data, compute, models, operations, and governance, often including locally built or trained models, but not limited to that. Private refers to deployment in an environment controlled by a single organisation, where that organisation controls data access, infrastructure, and keys. The underlying model may originate from any provider or country.
Why should organisations ring-fence their AI?
Organisations handling sensitive information should ring-fence their technology so that data goes in and never comes out. Security settings on public tools are a vendor promise that can change at any time, whereas privately hosted architecture provides a structural guarantee of control over data.
Can organisations still use public tools like ChatGPT?
Yes. Private deployment is a portfolio decision. Non-sensitive work such as marketing brainstorming can be conducted through public platforms, while privileged legal advice, intellectual property information, and other sensitive workflows should remain inside a ring-fenced environment.
Top comments (0)