Skip to content
Public sector

AI for Australian government

Agentic RAG and knowledge platforms for agencies that have to defend the answer, not just produce it. Permission-aware retrieval, auditable citations, and inference that stays in an Australian region.

Talk to our public sector teamRead the agentic RAG guide
01 / The problem

Why agency AI pilots stall after the demo

The demo is rarely the hard part. Agencies get stuck at the point where the system has to survive a privacy assessment, a records obligation and a question from an executive about where the answer came from.

Assurance arrives too late

Assurance treated as a gate at the end rather than a design input. By the time the framework is applied, the architecture has already made choices that cannot be unwound cheaply.

Sovereign storage, offshore inference

Documents held in an Australian region while every query and retrieved passage is processed overseas. The residency question gets answered. The processing question never gets asked.

Permissions bolted on

Retrieval that pulls everything and filters afterwards has already exposed restricted material to the model context. Entitlements have to be resolved before retrieval, not after.

02 / What we build

Knowledge platforms built to be reviewed

Every capability below exists because an agency had to answer for the system in front of someone. They are build requirements, not features. We are currently building the Knowledge Sharing Platform for the Victorian Collaborative Centre for Mental Health and Wellbeing, and we run the managed service behind the National Cancer Screening Register.

Permission-aware retrieval

Role-based access control enforced on the index itself. A user's entitlements are resolved before anything is retrieved, so restricted content never reaches the model context.

Traceable answers

Every response carries its sources. Which documents were consulted, in what order, and why the system judged them sufficient. Designed to be defensible under review.

Australian inference

Retrieval and model inference both run in an Australian region on AWS Bedrock (ap-southeast-2, Sydney). We map the full data path during discovery, including anything that leaves the boundary.

Audit logging as a deliverable

Retrieval logged against the requesting identity, retained to meet records obligations. Treated as a first-class output rather than something recovered from application logs later.

Ingestion at estate scale

CMS content, records systems, network drives and a decade of PDFs. Ingestion and content structure are most of the work in a real deployment. We run migrations as a practice.

Human oversight by design

Accountable officer identification, escalation paths and review points built in from the start, in line with Commonwealth guidance on responsible AI use in government.

Filtering after generation means restricted content already entered the model context. Permission-aware retrieval excludes it before the prompt is assembled, and logs every retrieval against the requesting identity.
03 / Approach

Assurance first, then the build

We start with the obligations, because they determine the architecture. Agencies that build first and assure later tend to rebuild.

01. Assurance discovery

Map the applicable obligations: Commonwealth policy and technical standard, state framework where relevant, privacy position, records requirements. Output is the constraint set the design has to satisfy.

02. Corpus and permissions

Inventory the knowledge estate and its access model. Establish what can be retrieved by whom. This is where most of the real effort sits, and skipping it is why pilots fail at scale.

03. Grounded pilot

A working system over a real corpus with real permissions, not a sandbox. Evaluated on answer quality, citation accuracy and whether it holds up when someone senior tries to break it.

ProofAll case studies
National Cancer Screening Register case study

National Cancer Screening Register / UX Research / Design Systems

Enterprise UX transformation & managed AEM support

A complete redesign with improved UX, a new design system, and ongoing managed support.

Read the case study →
04 / Guides

The AI in government reading list

Six guides written from inside the Australian obligations: DTA policy and technical standard, the Agentic AI addendum, ASD guidance, OAIC and OVIC. Start with the first and follow the links.

05 / Questions

What agencies ask us

Retrieval-augmented generation (RAG) grounds a language model in your own documents so answers cite real sources instead of being generated from training data. Agentic RAG adds a planning layer: the system decides what to retrieve, checks whether the retrieved material actually answers the question, and retrieves again if it does not. For government the difference is accountability. A single-shot RAG system returns a plausible answer. An agentic one can show which documents it consulted, in what order, and why, which is the difference between a tool an agency can defend in a review and one it cannot.

Both questions matter and most suppliers only answer the first. Data residency covers where documents rest. Inference locality covers where the model reads them. It is entirely possible to store a corpus in an Australian region and then send every query and every retrieved passage offshore for processing. Noice runs AI workloads on AWS Bedrock in ap-southeast-2 (Sydney), so retrieval and inference both stay in an Australian region. We will map the full data path for any deployment during discovery, including anything that leaves the boundary.

The relevant instruments are the Digital Transformation Agency's Policy for the responsible use of AI in government, its technical standard for government's use of AI (including the agentic AI addendum published in 2026), the National framework for the assurance of artificial intelligence in government, and the Guidance on AI procurement in government published through BuyICT. We design against these rather than retrofitting to them: transparency artefacts, accountable officer identification, logging and human oversight are build requirements from the outset, not a compliance pass at the end.

Victorian agencies work to the state guidance for the safe and responsible use of generative AI in the VPS, alongside privacy obligations that OVIC has set out for AI systems handling personal information. Procurement runs through the VPS AI Assurance Framework, referenced in the Victorian Government buying guidance. We are currently building the Knowledge Sharing Platform for the Victorian Collaborative Centre for Mental Health and Wellbeing, so this is the framework we work in day to day rather than one we have read about.

Permission-aware retrieval, enforced at the retrieval layer rather than filtered out of the response afterwards. The distinction matters: a system that retrieves everything and then tries to redact has already put restricted content into the model context. A system that resolves the user's entitlements before retrieval never sees it. We enforce role-based access control on the index, log every retrieval against the requesting identity, and treat the audit trail as a deliverable rather than a by-product.

Progress Agentic RAG as the retrieval and orchestration platform, where the requirement suits it. Noice is a registered Progress reseller through BlueChip Infotech. We are also a Magnolia Platinum Partner, which matters when the knowledge platform has to sit alongside an existing content estate. We are not tied to a single stack: platform selection happens in discovery against your assurance requirements, not before it.

No, and agencies should be clear about the boundary. GovAI Chat is a whole-of-government generative AI tool for APS staff, and it is a sensible default for general drafting and summarisation. It does not reach into an individual agency's permissioned document corpus, and it is not available to state government. Where an agency needs grounded answers over its own knowledge base, with its own access controls and its own audit obligations, that is a different system with a different risk profile.

Usually, yes. Most agencies hold knowledge across a CMS, a records system, network drives and a decade of PDFs. Ingestion and content structure are the majority of the work in a realistic deployment, not the model. Where the estate is Drupal or GovCMS, see our note on choosing a GovCMS alternative. Where content needs restructuring before it can be retrieved reliably, that is a migration problem and we run those as a practice.