Introduction
The adoption of AI within the Australian Public Service (APS) presents a unique set of challenges. The imperative to leverage AI for efficiency and improved citizen services must be balanced against stringent requirements for data sovereignty, privacy, and security. This framework provides a reference architecture for deploying "Sovereign-First" AI systems.
Defining Sovereign AI
In the context of Australian government deployments, Sovereign AI is defined by three core principles:
- Data Residency: All data—including prompts, completions, embeddings, and fine-tuning datasets—must remain within Australian borders.
- Operational Control: The government agency retains absolute control over the infrastructure, access policies, and audit logs. No external entity (including the AI vendor) has implicit access.
- Algorithmic Transparency: The ability to audit the decision-making process and enforce deterministic constraints on non-deterministic models.
The Reference Architecture
The proposed architecture leverages a Virtual Private Cloud (VPC) or on-premise deployment model, isolating the AI infrastructure from the public internet.
Key Layers
1. The Application Layer: Internal government applications, citizen-facing portals, or internal chatbots.
2. The Control Layer (Songlines Control): The critical enforcement point. This layer handles identity verification (via Entra ID / Active Directory), policy enforcement (e.g., preventing PROTECTED data from entering the model), and comprehensive audit logging.
3. The Model Layer: Locally hosted open-weight models (e.g., Llama 3, Mistral) or sovereignly deployed commercial models (e.g., Azure OpenAI Australia East) configured with zero-data-retention policies.
Implementation Guidelines
"Sovereignty is not merely a geographic constraint; it is an architectural guarantee of control and visibility."
Agencies should begin by classifying their data and AI use cases according to the Protective Security Policy Framework (PSPF). The Control Layer must be configured to enforce these classifications dynamically, ensuring that sensitive data is redacted or blocked before reaching the model layer, regardless of user intent.
Conclusion
By adopting a Sovereign-First architecture, Australian government agencies can harness the transformative power of AI while maintaining the highest standards of security, privacy, and public trust.