Services Australia is growing its staff AI trial to 2,000 people, one thumbs up at a time

There is a particular kind of frustration that anyone who has worked a service counter will recognise. The customer in front of you has a situation that touches three programs, the rules for each sit in different places, and the right answer exists somewhere in a library you are supposed to search with the right keyword. Services Australia is trying to fix that problem first, before it reaches for anything more ambitious.

The agency's Operational Blueprint is its approved how-to library, used by some 21,000 staff every day to work out service delivery processes. An AI trial now lets staff ask it a question in plain language and get relevant guidance pulled together, with links back to the source material. About 780 staff were in the trial as of October. By the end of the year it grows to roughly 2,000.

What stands out is how the agency is measuring it. Staff give each answer a thumbs up or thumbs down, and so far that has produced around 26,000 thumbs up and about 2,500 thumbs down, plus written feedback on what works and what does not. Lisa Keeling, a senior Services Australia executive, wrote that the feedback is used to improve search, responses, design, training and support, and to decide whether more guardrails are needed. Further expansion depends on readiness, assurance, operational capacity and consultation, including with the Community and Public Sector Union.

The wider program is called everydayAI. It pairs approved tools, Microsoft Copilot and GovAI Chat, with training, governance and assurance. Unapproved generative AI tools are not allowed at work, and customer information is not accessible to the AI tools. Staff completed mandatory AI in government fundamentals training earlier this year, and further courses are tailored to different roles.

Keeling put the principle simply: "The pace is guided by evidence and readiness, not enthusiasm for the technology." She also made the point that a new desktop tool does not, on its own, lead to better ways of working. Staff need the skills and confidence to use it, and to recognise when it is not suitable.

That second skill gets far less attention than the first. Most AI training teaches people how to get an answer. Very little teaches them when to distrust one, which is exactly what a staff member applying entitlement rules needs, because the accountability for the decision stays with them. Building that judgement on purpose is the point of our Critical Thinking workshops, and it is the part of a rollout that tends to decide whether staff trust the tool a year later.

Sources:

• The Mandarin, 3 November 2026: A people-first lens for AI in government services