7Architecture (the secure foundation)

Architecture is the force that decides whether AI can scale safely or fails brittle. It is the foundation of data, systems and infrastructure on which AI runs, and the structure that lets behaviour emerge without collapsing into chaos. For government, architecture carries a second, non-negotiable burden: it is where security and sovereignty are made real. This section sets out three things the architecture must provide, a foundation for intelligence, a backbone for trust, and a secure, sovereign place to run, and connects them to the agency’s obligations under the PSPF, the ISM and the Hosting Certification Framework.

7.1Architecting for a new reality

Traditional architecture was static: design the structures, document them, hand them over. Agentic AI breaks this, because the system adapts continuously and useful behaviour emerges at runtime. The architect’s role shifts from drawing fixed boundaries to shaping the conditions in which AI behaves well, and the deliverables shift from static diagrams to living models that can be simulated and adjusted. The agency does not need to embrace the full implications of this overnight, but it should build foundations that will not have to be torn up when its AI ambition grows.

7.2The AI capability stack: a foundation for intelligence

An AI system is best understood as a capability stack, in which each layer rests on the one below and weak foundations collapse what sits above. Renly departs deliberately here from the popular telling, which borrows Maslow’s hierarchy of needs, places raw signals above meaning, and puts emergence at the summit as the goal. For government that ordering is wrong twice over. Meaning is foundational, not a higher need: an AI cannot perceive usefully without a shared vocabulary to interpret what it senses. And the summit is not emergence but oversight: the apex of a public-sector AI system is the human governance that watches and bounds it. The stack below corrects both, and its message for an agency is that the unglamorous foundations, agreed meaning and clean, trustworthy data, matter more than the visible cleverness in the middle.

Four architecture domains, business, data, application and technology, act on every layer: business architecture decides which concepts matter, data architecture structures the shared vocabulary and trust, application architecture embeds meaning into services, and technology architecture provides the platforms to scale. For an agency, the practical message is to invest first in data and meaning: agree what things mean, and make the data trustworthy, before reaching for reasoning and autonomy.

7.3The decision passport: a backbone for trust

If data and meaning are the foundation, the decision passport is the contract that makes a decision trustworthy enough for an agency to stand behind. The idea is simple: every AI decision should travel with the data that lets it be trusted, explained, contested and audited. We model this as nine kinds of data that accompany each decision as a single payload. For government, this is the architecture of accountability: it is what makes an AI decision explainable, contestable and recordable, satisfying both the contestability ethics principle and the Archives Act.

Decision passport · minimum for production

Entity

Who or what the decision concerns

Policy

The rule and version it applied

Model

The model version that produced it

Trust

A confidence or trust score

Rationale

A reference to the reasoning

Also carried: structural, semantic, contextual, behavioural, interactional, lifecycle and provenance data.

The operating rule is exacting and worth adopting verbatim: an AI decision is not ready for production unless it carries, at minimum, the identity of the entity it concerns, the policy it applied, the model version that produced it, a confidence or trust score, and a reference to its rationale. An agency that holds its AI to this rule will find that explainability, contestability and recordkeeping are no longer bolted on after the fact; they are properties of the architecture.

7.4The secure, sovereign foundation

For a government agency, architecture is where the security and sovereignty obligations are met. AI workloads, the data they touch and the third-party tools that act on them must sit inside a hardened, assessed and sovereign environment. The framework recommends a concentric pattern: the AI workload runs inside an Azure landing zone built to the Commonwealth’s standards, wrapped by successive layers of control.

Five properties are non-negotiable: sovereignty, data and keys remain onshore and the classification ceiling is enforced; identity, zero-trust and least-privilege, just-in-time access; isolation, third-party and AI tools run in contained, monitored enclaves; observability, central logging, provenance and immutable audit; and a clean exit, no lock-in, with data and models portable on the agency’s terms. Architecture that holds these five is architecture an agency can defend to its audit committee and to the ACSC.

7.5Architecture outputs