AI Reference Architecture: Architecture overview

Architecture overview

AIRA provides two architecture views. The foundational view provides a common conceptual model for specialist and non-specialist audiences across disciplines and decision-making levels. The detailed view decomposes the architecture into capability groups and components for practitioners. Readers may begin with either view, depending on their role, purpose and phase of analysis.

Foundational view

Figure 1 Foundational view of the AI Reference Architecture is composed of 4 distinct and complementary architectural constructs:

Functional stack
Communication
Trustworthiness and assurance
Human agency, oversight and control

The Functional stack represents the technical execution environment of AI systems. It includes the Experience layer, Application and orchestration, Intelligence and analytics, Data and metadata, Platform and infrastructure. These layers describe where AI capabilities are implemented and operated.

Communication is a full stack connective tissue. It provides the patterns, protocols and controls required for secure, governed interaction between components within a system, across systems, agencies and external partners.

Trustworthiness and assurance (T&A) domains represent crosscutting qualities that are realised across the functional and communication stacks. These domains operationalise Observability, Traceability, Explainability and interpretability, Validity and reliability, Safety, Fairness, Security and resilience and Privacy, by embedding them into system design, rather than addressing them through post hoc controls.

Human agency, oversight and control (HAOC) form an overarching domain that is applicable across the functional and communication stack as well as T&A domains. This is concerned with engineering meaningful human agency, intervention and lifecycle oversight of AI systems.

Figure 1 Foundational view of the AI Reference Architecture

Diagram illustrating the foundational view of the AI reference architecture: Trustworthiness and assurance, Functional stack, Communication, Human agency, oversight and control. It is shaped by and shapes its Environments. Trustworthiness and assurance includes observability, safety, fairness, security and resilience, traceability and reproducibility, privacy, sustainability, explainability and interpretability, validity and reliability; Functional stack consists of layers from platform and infrastructure,

Detailed view

Beneath the layers and domains, AIRA provides capabilities and components to support the detailed architecture work of agencies. Capabilities are clusters of components that together deliver related functionality. Components are the most granular tier: well-defined, reusable and adaptable elements that each serve a specific function within AIRA. The capability level maps out a comprehensive set of logical components, such that any AI system in government could be assembled from these. Not every component, nor every capability group, will be needed for any given AI system. The capabilities adopted in practice will depend on the agency’s transformation model. At lower levels of transformation, agencies may rely on a smaller core subset of capabilities, while more mature or more integrated operating models may draw on a broader and more complete set across AIRA.

Figure 2 Detailed view of AIRA decomposed into reusable capability groups and components

Shows an illustrative decomposition of AIRA into capability groups. Both capability groups and components are expected to evolve as AI technologies, standards and government implementation patterns mature.

Diagram of AIRA demonstrating how the architecture is organised across three abstraction levels: Level 0 foundational constructs, Level 1 capability groups, and Level 2 components. It shows how the detailed capabilities and components support the foundational architecture. For a complete explanation, refer to the long text description below.

The diagram presents the AI Reference Architecture (AIRA) decomposed across three levels of abstraction: Level 0 foundational constructs, Level 1 capability groups, and Level 2 components.
Level 0
The architecture is organised into four Level 0 constructs:
Functional stack
The functional stack contains five layers:
- Experience
- Application and Orchestration
- Intelligence and Analytics
- Data and Metadata
- Platform and Infrastructure
Communication
Communication is represented as a separate vertical construct that spans interactions between systems, services and infrastructure.
Trustworthiness and Assurance
Trustworthiness and Assurance is represented as a cross-cutting construct containing governance and assurance domains that apply across the architecture. It includes Observability, Safety, Fairness, Security and Resilience, Traceability and Reproducibility, Privacy, Sustainability, Explainability and Interpretability and Validity and Reliability.
Human Agency, Oversight and Control
Human Agency, Oversight and Control spans the architecture and contains capabilities that support human intervention, review, oversight and operational control.
Level 1 capabilities
At Level 1, each Level 0 construct is decomposed into capability groups.
Experience layer
The Experience layer contains the capability groups Channels, Inclusion and Accessibility, AI Transparency and Disclosure, and User Traceability and Consent.
Application and Orchestration
The Application and Orchestration layer contains AI Orchestration, Agentic Application, API Management, and Context Engineering.
Intelligence and Analytics
The Intelligence and Analytics layer contains Model Acquisition and Development, Model Evaluation and Validation, Model Registry and Versioning, Model Deployment and Promotion, Model Inference and Serving, Model Monitoring, and Model Retirement and Decommissioning.
Data and Metadata
The Data and Metadata layer contains Data Acquisition and Store, Data Preparation and Transformation, Data Provenance and Lineage, Data Access and Sharing, and Metadata Management.
Platform and Infrastructure
The Platform and Infrastructure layer contains Compute, Storage, Networking, Hosting, Developer Environments and Tooling, and Edge and Sovereign Deployment.
Communication
The Communication construct contains Identity and Trust, Internal Network and API Infrastructure, Legacy Integration, Agentic Communication Protocols, Internet of Things, and Shared Catalogues and Reusable Assets.
Human Agency, Oversight and Control
The Human Agency, Oversight and Control construct contains the capability groups Decision Support and Intervention.
Level 2 components
At Level 2, selected capability groups are further decomposed into individual components.
Decision Support
Within Decision Support, example components include Alerting, Anomaly Highlighting, and Escalation Routing.
Intervention
Within Intervention, example components include Human Override, Circuit Breakers, Irreversibility Guardrails, and Transfer of Control.

Applying AIRA proportionately

AIRA supports a broad range of AI use cases, from productivity through to systems that support consequential decisions and actions. The same foundational architecture applies across these. Figure 1 remains relevant regardless of the technology, deployment model or use case. What differs is the detailed architecture: which capabilities are realised, and to what depth and rigour.

For example, a productivity assistant may need relatively simple assurance and oversight patterns such as source attribution, and a human reviewing AI-drafted content before use. In more consequential uses, traceability may need to support decision reconstruction, review, appeal and audit. As such, the detailed architecture will need to support legal defensibility, procedural fairness, contestability, evidentiary integrity.

Choosing the right level of detail

Early phases of an initiative typically focus on foundational architecture, such as the functional stack, communication, trustworthiness and assurance, and human agency, oversight and control. As understanding of the problem, requirements and constraints matures, architectural analysis can progressively move toward more detailed capabilities and components. This allows agencies to begin with broad architectural considerations, avoid premature architecture decisions, and introduce depth as information becomes available.

Building evaluation into the architecture

Evaluation is considered across the architecture. Establishing how a system will be evaluated, including whether meaningful evaluation is possible, is part of architecting it.

Technical standard for government’s use of artificial intelligence: Evaluate | digital.gov.au sets expectations on evaluating AI systems before deployment and continuously throughout their lifecycle. It also includes expectations on comparative assessments against existing human and hybrid baselines.

Evaluation applies at multiple architectural levels. AIRA’s scope includes evaluation patterns for individual layers, across layers, across systems and systems-of-systems. It covers evaluating trustworthiness and assurance, and human agency, oversight and control, across the functional stack and communication boundaries.

Understanding architectural trade-offs

Architectural decisions often involve trade-offs across the architecture. Optimising one element may influence another. For example, increasing human oversight may reduce operational efficiency, while strengthening security controls may affect interoperability.

Different architecture approaches optimise for different outcomes. More integrated platform solutions may accelerate delivery, reduce integration effort and simplify operations. More modular or best-of-breed approaches may provide greater flexibility and specialised capability, but can introduce additional integration, governance and operational complexity. Neither approach is inherently preferable. Architecture helps agencies understand these trade-offs and determine which approach best aligns with their objectives and constraints.

The foundational architecture in Figure 1 provides a shared frame of reference for understanding and evaluating these trade-offs across disciplines. It allows executives, architects, engineers, assurance practitioners and business stakeholders to discuss architectural decisions and their impacts at a strategic level.

Architecture description

AI Reference Architecture