4The North Star: the Governed Intelligence Model

To govern AI across an enterprise, an agency needs more than a list of controls. It needs an organising idea that connects ambition to delivery, delivery to foundations, foundations to safe design, and design to accountable assurance, and that keeps all of them in balance as the agency learns. This section sets out that idea, the Governed Intelligence Model. Because the model draws on a fast-moving and uneven field, we begin by saying plainly what the prevailing agentic-AI thinking gets right and where it falls short, so that what we keep is kept deliberately and what we change is changed for a reason.

4.1A critical reading of the agentic playbook

A body of practitioner work has grown up around agentic AI, most visibly Jesper Lowgren’s writing on agentic design. Much of it is useful, some of it overstated. Renly has read it closely. A great deal of it is genuinely valuable, and we build on it. But for a public-sector audience that will be held to account by an audit committee and a court, some of its habits are liabilities, and we have replaced them with sources that will bear weight. Honest engineering means doing both: keeping what works and being candid about what does not.

Four insights from the field are sound, and we keep them. First, agent-first design: for an autonomous system you must design from what the machine perceives and decides, not from a human’s journey through a screen. Second, runtime governance: rules must be enforced as the system acts, not reconstructed from logs afterward. Third, the autonomy threshold: the step from an AI that suggests to an AI that operates is a genuine change in kind, not degree, and governance must step up to meet it. Fourth, the decision passport: every AI decision should carry the data that lets it be trusted, explained and contested.

Four habits, though, do not survive contact with government accountability, and we have replaced each. Where the popular frameworks reach for evocative but arbitrary scaffolds, the five elements of classical cosmology, a hierarchy of needs, three laws of robotics, we ground our model instead in the NIST AI Risk Management Framework, ISO/IEC 42001, and the Commonwealth mandate. Where they impose a memorable numerology, everything arriving in fives, sevens or nines, we let structure follow function, even when the resulting count is untidy. Where they offer taxonomies that blur together, most visibly a set of nine overlapping agent archetypes with inconsistent names, we separate the few things that genuinely differ. And where they place emergence, the system surprising its makers, at the summit as an aspiration, we place oversight there instead, and treat emergence as a risk to be governed. The result is the same breadth of coverage, made rigorous, original, and defensible to the people an agency must answer to.

4.2The Governed Intelligence Model

The Governed Intelligence Model names five domains that every organisation must get right to adopt AI safely and at scale: Strategy, Operating Model, Architecture, Design and Governance. The five are not a checklist to be completed once but a system to be kept in balance, and they correspond, in plain terms, to the four functions of the NIST AI Risk Management Framework: govern (Strategy and Operating Model), map and measure (Architecture and Design), and manage (Governance, closing the loop). Applied to AI, the model gives leaders a way to see the whole, and to notice when one domain is racing ahead of the others.

• StrategyAmbition and risk appetite
• Operating ModelAccountability and value flow
• ArchitectureSecure, sovereign foundation
• DesignSafe, contestable systems
• GovernanceAssurance and feedback

4.3Two loops: enablement and assurance

The domains relate to each other through two loops that run in opposite directions. The enablement loop carries the organisation forward. A clear strategy mobilises the operating model, giving people and processes direction. The operating model grounds the architecture in real executional needs. The architecture lets design scale beyond one-off pilots. And design defines what must be governed, because every new behaviour and permission is something governance must address. This is the accelerator: it turns ambition into delivered value.

The assurance loop runs the other way, and it is the brake and the steering. Governance observes how the system behaves, holds it to account, and feeds what it learns back to strategy, refreshing priorities and surfacing risk. It applies control inward too, so that delivery does not outrun safe practice, ambition does not demand what the foundation cannot bear, and yesterday’s comfortable patterns do not quietly become tomorrow’s constraints. Neither loop is sufficient alone. A model that only enables becomes reckless; one that only assures becomes paralysed. The discipline is to run both, and to keep them in balance.

4.4Why a self-balancing model, and why now

Autonomy introduces a kind of systemic instability that organisations have little experience managing. Small misalignments between strategy, systems and governance can produce outsized, unintended consequences, because AI acts independently, learns iteratively and scales quickly. A linear framework, a plan executed step by step, cannot keep up, because by the time the last step is reached the system has already changed. A two-loop model is built for this. It expects feedback, it expects correction, and it makes the second- and third-order effects of a change visible before they compound. For a public-sector leader, that is the difference between governing AI and being surprised by it.

4.5The five domains, anchored to the Australian mandate

The model’s real power for government is that each domain has a clear home in the Commonwealth mandate. This framework treats the domains as the structure and the instruments as the content, so that working through the five domains is, at the same time, discharging the agency’s obligations. The table below makes the mapping explicit, and Sections 5 to 9 work through each domain in turn.

4.6How to read the rest of this framework

Part II takes the five domains one at a time. Each section explains what the domain must achieve, introduces the supporting models that make it concrete, anchors it to the mandate, and ends with the outputs an agency should be able to point to. Part III then assembles the domains into an assurance lifecycle and a maturity roadmap, gives leaders the executive agenda for driving it, and closes with practical guidance for applying the framework in your agency. The supporting models, the AI Autonomy Ladder, the capability stack, the decision passport, the two-plane account of agent roles, the autonomy gates and the runtime controls, are our own synthesis, grounded in the standards named above and translated into the language and obligations of Australian Government.