The Australian Public Service’s approach to generative AI has matured substantially through 2024-2026. The early policy work that culminated in the APS-wide guidance frameworks of late 2024 has now had eighteen months of practical implementation, and the patterns of what’s working and what isn’t are reasonably clear.

A practical mid-2026 read on where the APS has landed.

The current policy architecture

The current operational picture reflects three intersecting threads of guidance.

The whole-of-government guidance from the DTA and the Department of Finance establishes the overarching expectations for use of AI tools within APS organisations — risk assessment, transparency, accountability, and data protection requirements.

Agency-specific policies, sitting under the whole-of-government framework, have been issued by most major departments and agencies. The agency-level policies vary in detail and emphasis but most have settled into broadly similar structures — permitted uses, prohibited uses, risk assessment processes, and accountability frameworks for staff using AI tools.

The procurement framework, anchored in the Digital Marketplace and the various sourcing guidance documents, has been updated to reflect AI-specific considerations including data sovereignty, model provenance, and supplier accountability for AI behaviour.

What’s working

Several aspects of implementation have settled into functional practice.

Day-to-day staff use of approved tools. Most APS staff now have access to one or more approved generative AI tools — typically Microsoft 365 Copilot, with agency-specific additional approvals for specialised tools. The training and onboarding for staff has matured, with most agencies running structured introduction programs followed by ongoing community-of-practice support.

Risk assessment processes. The triage processes for evaluating new AI use cases have become more workable. The early implementations were heavyweight and often produced bottlenecks; the refined processes in 2025-2026 are more proportionate, with lightweight assessment for low-risk cases and detailed assessment reserved for higher-risk applications.

Data protection and sensitivity classification. The integration of AI use considerations with the existing information classification framework has worked reasonably well. The discipline of asking “what data is being put into the AI tool” has become standard practice, and the friction this creates is generally accepted as proportionate to the risk being managed.

Supplier engagement. The conversations between APS procurement teams and AI tool suppliers have become more substantive. Suppliers know what the APS expects, the contract terms have stabilised, and the negotiation cycles for renewal procurements are shorter than they were in 2023-2024.

What’s still being figured out

Several areas where the implementation continues to evolve.

Bespoke AI development versus tool procurement. The line between when an agency should procure an off-the-shelf AI capability versus when it should commission bespoke development continues to be debated. The cost dynamics, the risk dynamics, and the long-term maintenance dynamics all push in different directions, and the answer varies by use case.

For complex bespoke build work where APS in-house capability is constrained, an Australian AI company or similar specialist provider has typically been engaged through panel arrangements rather than agency-specific procurement, with mixed outcomes on cost-effectiveness depending on the specific engagement.

Outcomes measurement. The APS doesn’t yet have well-established frameworks for measuring the outcomes of AI investment. Time savings, quality improvements, cost reductions — all of these are being claimed but the measurement methodology is uneven across agencies. The next phase of policy development is expected to focus on outcomes evidence rather than process compliance.

Higher-risk applications. The integration of generative AI into citizen-facing service delivery, automated decision-making (or decision-support) in regulatory functions, and similar higher-risk applications continues to be a careful incremental process. The few pilots that have been run have been carefully bounded and the broader rollout pace remains deliberate.

Workforce implications. The honest conversation about what generative AI means for the APS workforce is still nascent. Some functions are clearly being affected; the implications for workforce planning, skills development, and recruitment haven’t been fully worked through. The Australian Public Service Commission’s work in this area through 2025 has been useful but the question is bigger than any single agency can address.

Sectoral patterns

A few observations on how implementation has varied across different parts of the APS.

The departments most actively engaged with generative AI have been the digitally-mature service delivery agencies — Services Australia, the ATO, the Department of Home Affairs (within tightly bounded use cases) — and the policy and analytics functions across agencies. These have generally moved faster, taken more calibrated risks, and produced more demonstrable outcomes.

The line departments responsible for sensitive operational functions — defence, intelligence, foreign affairs, law enforcement — have moved more cautiously, reflecting the higher sensitivity of their data and decision-making environments. The progress has been real but the pace has been slower.

The smaller agencies and statutory bodies have varied widely. Some have leveraged shared services and the whole-of-government frameworks effectively. Others have struggled with the resource implications of meaningful AI capability development on smaller bases.

What to expect through H2 2026

A few patterns I’d expect to continue or develop.

The outcomes evidence base will strengthen. Agencies that have been investing in AI capability since 2023-2024 will increasingly be asked to demonstrate measurable outcomes, and the resulting evidence will inform the next round of strategic investment decisions.

The integration of AI into core business processes will deepen. The current pattern of AI-as-productivity-tool will progressively give way to AI-as-process-component, with corresponding implications for process design, accountability, and risk management.

Capability development will become more sophisticated. The skills the APS needs for the next phase — including prompt engineering, AI risk assessment, AI procurement specialisation, and AI ethics — are different from the skills emphasised in earlier digital transformation work.

External capability partnerships will continue. The complementarity between APS in-house teams and specialist external providers — including AI consultancies, niche developers, and academic partners — has been a pragmatic feature of the implementation so far, and is likely to continue.

The APS’s generative AI implementation in mid-2026 reflects a substantial, deliberate, and broadly successful policy and capability development effort. The work is far from complete and the operational picture will continue to evolve, but the foundations have been laid in a way that should support more ambitious applications through the back half of the decade.