What are the three things every agency now has to do?
For agencies operating inside the Plan's scope, three obligations carry the most operational weight in 2026: appoint a Chief AI Officer, mandate AI literacy training, and designate accountable officials for every AI use case.
Every agency is required to appoint a senior executive as Chief AI Officer (CAIO) by 31 July 2026. The role is distinct from the AI Accountable Official under the existing AI in Government Policy: the Accountable Official focuses on governance and risk; the CAIO focuses on adoption, capability development, and driving consistent uptake. Smaller agencies may combine the roles in a single leader.
CAIOs operate as a network. A peer working group is developing shared training materials, reusable use-case patterns, and a common forum for the cross-agency questions that AI adoption surfaces. The intent is explicitly that no agency has to solve AI adoption alone.
Through an updated AI in Government Policy led by the APSC, foundational AI literacy training is being mandated for all 220,000+ APS employees. This is supported by GovAI interactive learning, the APS Academy, and live webinars with public servants who have hands-on AI experience.
The DTA Copilot trial across 7,600+ APS staff in the first half of 2024 surfaced a finding that has shaped the Plan's training emphasis: staff who received more than three forms of training were significantly more confident and effective AI users. The Plan treats training as a precondition for adoption, not a follow-up.
The AI Impact Assessment tool is now mandatory for in-scope use cases. Every AI use case must have a designated accountable official responsible for its safe deployment. The DTA is establishing a central register of generative AI assessments on GovAI, allowing agencies to reuse prior evaluations and streamline procurement decisions.
For agencies, this means that AI initiatives that previously sat inside individual teams now need a clear accountability chain ending in a named senior executive. Not every use case will face the same scrutiny, but every use case needs the chain.
What did the Copilot trial actually show?
The DTA evaluation of the whole-of-government Copilot trial, published in October 2024, is the largest documented public-sector evaluation of generative AI to date, and the Plan's implementation strategy clearly draws on it.
86% of trial participants wanted to keep using the tool, with adoption sentiment strongly positive. 69% reported faster task completion, with time savings concentrated in summarisation, search, and meeting minutes. About one hour per day was saved on routine summarisation and information-retrieval tasks, particularly for APS levels 3-6 and EL1. Around 60% of users had to make moderate to significant changes to AI-generated output before it could be used, and only a third of trial participants used Copilot daily, with usage concentrated in a smaller cohort of intensive users.
The picture is genuinely positive, and genuinely qualified. AI is delivering value, but it is not autonomous. The people who got the most from it were the ones who had the most training and engaged with it most consistently. The Plan reflects all three of those findings.
Worth knowing: GovAI Chat, the secure, government-hosted conversational AI platform under the Plan, began APS trials from April 2026. It runs within Australian Government infrastructure, with data remaining under government control. It is intended to provide a sovereign alternative to commercial AI tools for OFFICIAL-level work.
What does this signal for agencies outside the formal scope?
The APS AI Plan is binding on Commonwealth agencies only, but three dynamics are making it operationally relevant well beyond its formal scope.
State governments are often ahead, not behind. NSW introduced mandatory AI assurance requirements in 2022 and strengthened them in 2025. Several states have moved earlier than the Commonwealth on practical AI governance instruments. The Plan does not standardise these; it parallels them. Agencies operating across federal and state arrangements are increasingly navigating both.
Procurement is propagating Plan-aligned requirements. Commonwealth procurement of AI services is increasingly referencing Plan-aligned obligations. Suppliers, including major consulting firms, system integrators, and SaaS platforms, are aligning their internal practices to what they expect Commonwealth buyers to ask for. That ripple effect is reaching into private-sector AI programs without any formal Plan obligation.
Robodebt is the operating constraint nobody names. The Plan does not invoke Robodebt directly, but every paragraph operates inside its shadow. The public expectation that automated government decisions be explainable, contestable, and subject to human oversight is now hard-coded into how Australian government AI initiatives are designed. Agencies, Commonwealth or otherwise, that build AI without that posture are building toward a problem.
What is the deeper signal here?
The APS AI Plan is the first comprehensive whole-of-government attempt anywhere to operationalise responsible AI adoption at scale, and it is being watched closely by peer governments.
If it works, it will become the operating template for public-sector AI globally. If it stumbles, the lessons will travel just as far.
For agencies inside the Plan, the right posture is not just compliance, it is contribution. The Plan succeeds or fails on the use cases agencies build inside it, the lessons they share through the CAIO network, and the trust they earn from the public. That is a programmatic challenge, not just a technical one.
What we would do this quarter
- Audit existing AI use against the AI Impact Assessment toolInitiatives launched before November 2025 may not have been assessed against the framework that now applies. Identify them, prioritise them, and assess them.
- Map the accountability chain for every AI use caseWhere a chain ends short of a named senior executive, fix it. The Plan's expectation is explicit: every AI use case has a designated accountable official.
- Plan training in tiers, not in bulkFoundational literacy is mandatory. But the Copilot evaluation is clear that the value comes from layered training, foundational, role-specific, use-case-specific. Build the layered offer now.