Why the APS needs to be picky with AI

As of June 2026, the Digital Transformation Agency (DTA) has launched the first mandatory requirement of its policy for the responsible use of AI in government, highlighting that every use case must be individually governed, assessed for risk and impact, and assigned an accountable owner.

There’s no doubt it’s the right call, but it raises the question of whether agencies are prepared for this? Tight budgets and an expectation to modernise quickly means that it’s tempting for agencies to move on AI as quickly and as much as possible. But this only exacerbates the risk.

With more tools to manage and more room for error, this approach will likely create use cases that are tricky to audit and govern under the new DTA framework, as well as driving inefficiencies, ultimately making AI difficult and expensive to run.

The problem with ‘AI everywhere’

The enthusiasm across the Australian Public Service (APS) is fantastic but it needs a clear plan to turn the ambition into a reality.

Firstly, AI tools aren’t free, and the costs add up fast when deployment is scattered, from licensing to integration to maintenance fees. As this happens across multiple use cases without a clear idea of the return on investment, many agencies are already finding that the budget is getting out of control.

On top of this, most government systems weren’t built for modern AI tools to simply plug in. There are years of accumulated complexity with outdated platforms and siloed data. AI is only as good as its foundations and without addressing the underlying issues, things become more complicated and fragile; and this affects governance, which is now urgent.

The DTA requires agencies to document every AI use case, assess its risks and assign someone accountable for it. It seems straightforward, but when there’s no structure in place and AI is deployed widely and quickly, the questions of accountability, explainability and ownership that nobody has the answer to are sure to creep in fast.

Selective deployment is a strategic advantage

To combat this, agencies must be deliberately selective, placing focus on AI deployment for use cases that are high-volume, repetitive and well-defined: for example, processing applications, arranging correspondence, checking documents or triaging requests. These tasks use up a lot of staff time, but the rules are clear, outcomes easy to measure and they are straightforward to govern. This means AI can give hours back to people for other tasks, while still being able to easily track what the tool is doing, why it’s making certain decisions and who the person responsible is.

This predictability is different from embedding AI across many different, undocumented processes that may be difficult to manage and track, and therefore will likely add further complexity to an already stretched team. Selecting specific tasks it can do reliably and accountably instead significantly reduces existing pressures on workers.

How to fix the foundations first

To allow AI to be embedded for these chosen tasks, the government must modernise its legacy systems so that AI has solid foundations to run effectively. However, there’s no reason why AI can’t help you in this process.

Mapping legacy processes that have never been properly documented, identifying duplicated or redundant steps, generating options for how a service can be redesigned from the user’s perspective: these are all necessary tasks that typically take months of work. But with AI, that time can be compressed significantly. This means agencies get a faster, cheaper path to understanding what they’re working with and what needs to change.

Additionally, agencies that are selective and focus AI on genuine process simplification rather than blindly automating whatever already exists, will find compounding efficiencies such as fewer handoffs, less rework, simpler systems to maintain, and better experiences for both staff and citizens: a much stronger return than if you were to spread AI too thinly.

Accountability must be built in from the start

When laying these foundations, agencies mustn’t forget about governance. The DTA’s mandatory AI impact assessments around fairness, safety, privacy, transparency, security and human-centred values, are most consequential in high-stakes situations — such as welfare entitlements, licensing decisions or case outcomes that affect people’s lives.

This means transparency isn’t optional. Agencies need AI that can show its reasoning, flag cases that require a person to review them, and keep a clear record of what happened and why — one that can stand up to scrutiny. That requires purpose-built workflow and decision management capabilities, not general-purpose AI dropped into an existing process.

The governance framework the DTA has put in place points us in the right direction, but it can only work if the AI is actually governable to begin with.

From strategy to delivery

Recently, the Organisation for Economic Co-operation and Development (OECD) recognised Australia as a global leader in technical maturity: a vote of confidence in our nation’s advancement of human-centred digital transformation. But it also noted our next core challenge is moving from strategy to measurable execution, and the DTA’s governance requirements reinforce that point.

Great intentions and ambition will only get us so far and are only as good as our ability to use AI to actually improve services, reduce workload and deliver demonstrable value, rather than chasing volume of AI for its own sake. This requires focus: identifying the correct use cases, investing in getting the foundations right and developing the proper governance.

In a resource-constrained environment, doing less, but doing it properly is the smarter solution.

Image credit: iStock.com/Just_Super