Architecture description
Environments
Architecture does not exist in isolation. It is shaped by and can shape its external environments.
Environments can be grouped into the following categories:
- People and the physical world, including the public, staff, partner agencies, agency roles and natural environments.
Legislative, policy and regulatory, including legislation, delegated authorities, administrative law requirements, government policies, standards, assurance frameworks, reporting expectations and regulatory obligations.
- Business processes, including operational workflows, service delivery processes, decision-making processes, and organisational operating models that define how work is performed within and across agencies.
- Hardware and software systems, including Internet of Things sensors, devices, platforms, external agents and other digital systems.
For agencies exercising statutory powers, legal and regulatory obligations are often a primary driver of architecture decisions. Agencies operating under legislation with coercive powers, compliance functions or statutory decision-making responsibilities require these obligations to be considered and implemented across the functional stack, communication pathways, trustworthiness and assurance domains, and human agency, oversight and control. This includes maintaining traceability from legislation, policy obligations, delegated authorities and operational procedures through to architectural controls, decision pathways, audit mechanisms, assurance activities and human oversight arrangements.
Functional stack
Experience layer
This layer includes the interfaces through which people, systems, devices and agents interact with an AI-enabled system. It covers human-facing experiences, machine-consumable services, sensing functions and actuation functions. Sensing functions recognise physical properties in the environment, for example, temperature or proximity sensors. Actuation functions translate signals into environmental change, such as smart air conditioners or robotics. The experience layer focuses on the interaction points and capabilities exposed by the system, while Communication defines the protocols, integration patterns and trust mechanisms used to connect them.
This layer supports inclusive, accessible, usable, transparent and responsive experiences, while supporting appropriate trust, oversight and user understanding of AI‑enabled interactions. Where AI is introduced into a service channel, the architecture should also consider routing users to non-AI alternatives and sustaining equitable service access for users who decline or cannot engage with AI.
Typical considerations for the layer include:
- channels – human-to-system, system-to-system and physical interaction points
- inclusion and accessibility – multi-modal input/output, feedback channels, age-based experiences, cultural adaptation, reading-level adaptation, multi-lingual, compatibility with assistive technology
- AI transparency and disclosure – communicating when and how AI is being used, progressive onboarding
- user traceability and consent – visibility and consent of AI interactions and data use
Application and orchestration layer
This layer covers domain-specific workflows, decision logic and AI orchestration to enhance business services and processes with AI capabilities. It provides guidance on coordinating interactions between AI agents, AI models, deterministic systems, human decision-makers, Internet of Things (IOT) systems and third-party services. Deterministic systems are systems where the given input always produces the same output, without randomness. Coordinated interactions in this layer are auditable, resilient, cost-effective and aligned with policy and regulatory obligations.
Typical considerations for the layer include:
- AI orchestration – coordinating AI, systems and workflows
- AI agentic application – coordinated agents
- API management – system integration for orchestration and system management
- context engineering – providing context, knowledge and instructions
- business rules and decisioning – combining AI and deterministic logic
- workflow integration – embedding AI into business processes.
Intelligence and analytics layer
This layer covers the analytical and AI capabilities that generate insight, predictions, classifications, recommendations or derived outputs from data. This layer includes conventional analytics such as reporting, statistical analysis, forecasting, optimisation and decision support, as well as AI techniques such as machine learning, generative AI, reasoning and planning that agentic applications may consume. The layer also covers the acquisition, development, evaluation, deployment, monitoring and maintenance of models and analytical artefacts over time. This layer helps agencies combine analytical and AI approaches appropriately.
Typical considerations for the layer include:
- Operational frameworks for AI model development, delivery and operations
- AI model acquisition and development – build, procure and customise models
- AI model evaluation and validation – performance, safety and quality assurance
- AI model registry and versioning – lineage and version control
- AI model deployment and promotion – controlled release management
- AI model inference and serving – running and exposing models
- AI model monitoring – performance, drift and compliance monitoring
- AI model retirement and decommissioning – lifecycle closure and replacement.
Data and metadata layer
This layer includes operations on data that are collected, stored, transferred, processed and used by AI systems. It enables coordinated handling of large, structured, semi‑structured and unstructured data. It covers the storage, processing, governance and access requirements for data, where the data varies by type, scale, source, sensitivity and legal or policy constraints.
AIRA supports use cases that handle Indigenous data. Guidance specifies where the architecture applies directly to Indigenous data, where it requires adaptation and where additional metadata, governance or control conditions are needed.
Imperfect data does not automatically block AI experimentation or the development of AI systems. In many government contexts, data will be incomplete, inconsistent, historically accumulated or uneven in quality.
AIRA supports the safe use of imperfect data through measures such as lineage and provenance tracking, quality measurement, confidence signalling, human review points, constrained use of outputs, fallback pathways and iterative improvement of datasets and models over time.
Typical considerations for the layer include:
- data acquisition and store – ingest, integrate and store data
- data preparation and transformation – clean, enrich and structure data
- data provenance and lineage – track origin, movement and usage
- data access and sharing – secure access and data exchange
- metadata management – classification, cataloguing and discovery.
Platform and infrastructure layer
This layer provides the computing environment to host and operate AI systems. It encompasses compute and emerging accelerators including GPU, TPU and NPU. This layer also includes storage, networking, containerisation platforms, run-time environments, developer workstations, data centres, edge environments and associated land, water and energy considerations. It provides workload isolation, tenancy separation and secure execution environments to enforce trust boundaries and protect sensitive workloads.
While most AI systems in government rely on third‑party infrastructure providers, AIRA provides a consistent basis for assessing infrastructure options against AI workload requirements, risk profiles, sovereignty constraints and whole‑of‑government expectations.
Typical considerations for the layer include:
- compute – CPUs, GPUs and AI accelerators for processing
- storage – secure and scalable data and model storage
- networking – connectivity, performance and security
- hosting – cloud, on-premises and hybrid platforms
- developer environments and tooling – development toolkits
- edge and sovereign deployment – local, disconnected and sovereignty-sensitive workloads.
Communication
Communication provides integration patterns, protocols and trust mechanisms to enable secure, reliable and auditable communication across the functional stack and across systems, agencies and external partners. It provides architectural guidance for integrating with legacy systems, managing trust boundaries, selecting appropriate communication protocols and enabling reuse through shared services, models and data catalogues.
Cross-agency communication patterns specify if they are applicable to Indigenous data, or whether adaptation is required to preserve provenance, purpose, cultural sensitivity, access conditions, authority-to-control requirements and constraints on reuse, onward disclosure or secondary use across organisational boundaries.
Typical considerations for the layer include:
- identity and trust – authentication, authorisation and federation
- internal network and API gateway – secure connectivity and integration
- legacy integration – connecting existing and modern systems
- agentic communication protocols – agent-to-agent and agent-to-system interactions
- IoT – connectivity with IOT devices, sensors and actuators
- shared catalogues and reusable assets – standard and reusable data, models, APIs and services.