In some of my last roles in government, you could put a huge part of my day into two buckets: giving information to someone else, or getting information from them. Comments on a draft. Figures for a briefing. An approval before a project could move. An explanation of a policy that made sense to our team but needed translating for everyone else.
That is the part of government work worth looking at closely when we talk about AI. There is a lot of effort between the meeting and the decision, or between receiving a question and giving someone a useful answer. AI creates an opportunity to reduce that effort. Waiting too long to understand it has a cost, both for agencies and for the people building their careers inside them.
Back in June, I wrote a short series about practical AI use in government. We looked at meeting follow-ups, chasing input from other teams and explaining policies in plain English. Those are useful places to start. But once you look behind those individual tasks, you can see a bigger change taking shape.
Look at what happens after the meeting
Imagine a project meeting has just finished. There are rough notes, a few agreed actions and one issue that still needs a decision. Someone needs to work out who is doing what, update the tracker, ask another team for input and prepare something the director can act on. None of this sounds particularly complicated. Put it across several projects and a week full of meetings, though, and it becomes a substantial part of the job.
An approved AI tool can help turn those notes into a draft update. The person reviewing it still needs to notice that a suggestion was never actually agreed, that the proposed deadline is unrealistic, or that the tool has confidently assigned an action to the wrong person. Then the team needs somewhere reliable to record the corrected version and follow it through.
That is where the behind-the-scenes work becomes interesting. A quicker first draft is useful. The larger opportunity is to improve the whole process so decisions are clearer, requests reach the right people and work moves with less chasing. You only find out whether that has happened by looking beyond the moment the AI produces an answer.
It also explains why knowing the work matters so much. Someone who understands the process can see which steps are repetitive and which require judgement. They know the awkward exceptions. They can tell when a neatly written answer would create more work for the next person.
The policies show where the work is heading
The government announcements are useful because they give us a clearer picture of what agencies are being asked to put in place. The national assurance framework was agreed by Commonwealth, state and territory governments in June 2024. It is the earlier foundation for a consistent approach, with room for each jurisdiction to implement it differently.
In NSW, the AI Assessment Framework platform became mandatory for new AI use-case registrations and assessments from 30 September 2026. There is an important practical qualification in the AI Operational Policy: routine drafting, summarising and similar productivity work using approved agency tools generally do not need registration where meaningful human oversight and agency safeguards remain in place. Check the conditions and your agency's own rules before relying on that.
Commonwealth agencies covered by the responsible AI policy have a separate timetable. By 15 December 2026, they must implement mandatory responsible-AI training and responsible-use processes, along with accountable owners and registers for in-scope AI use cases. These requirements do not automatically apply to state employees.
Read those requirements with a practical eye. Someone has to understand the process being changed, decide what good performance looks like and make sure problems can be reported and fixed. Staff need useful training. Managers need to know what they are approving. Those responsibilities become part of the implementation work.
There is an opportunity here for people who understand government
A business analyst can map where an enquiry gets stuck and help define what an AI-assisted process should do. A project officer can organise testing and keep track of the problems it reveals. An experienced customer service officer can identify the cases where a standard answer will let someone down. Corporate services staff can help make guidance usable, while checking that the meaning survives the rewrite.
These contributions draw on skills people already use in government. AI adds new things to learn, including how to test an output and how to handle the information going into a tool. There is a practical starting point for someone who knows their work well and is prepared to learn those additional skills.
For someone applying for roles, that is worth paying attention to. Being able to explain a real improvement, the checks you put around it and what happened afterwards gives a hiring panel something concrete to assess. An AI course can be useful preparation. A genuine example of applying what you learned gives you more to talk about. Be clear about your contribution and never dress up a practice exercise as workplace experience.
For agencies, the opportunity is to put more of people's time into work that needs their attention. A team spending less time reconstructing updates might have more capacity to resolve difficult cases or explain a decision properly. Whether that happens depends on how the team uses the time released. Buying licences alone does not establish a benefit.
There is some encouraging evidence. In the government's 2024 Copilot evaluation, 69% of survey respondents said it helped them complete tasks faster, and 40% reported reallocating time to higher-value work. Those were self-reported results from a non-randomised trial, with important limitations. Checking and editing outputs reduced some of the gains. It is a good reason to test useful tasks and measure the results in your own setting.
What waiting actually costs
The risk of moving too slowly can be hard to see because very little appears to happen. The team keeps using the same process. The same people chase the same missing information. Work that might have become easier continues to take the same amount of effort, and nobody has a useful comparison because the alternative has never been tested.
There is also a learning cost. A team that runs a small, approved trial can discover which tasks are suitable, where the tool makes mistakes and how much checking is needed. It can improve its instructions and decide whether to continue. A team that waits another six months will still have those questions to work through when it starts. Access to a better tool later will not answer them automatically.
At an individual level, delaying that learning can leave you with fewer examples to draw on when a project or role asks for experience improving work with AI. That does not mean everyone needs to become a technical specialist. It does mean there is value in understanding how these tools affect the work you want to do, while there is time to practise in a controlled setting.
There is a further risk for agencies: decisions about new systems may be shaped without enough input from the people who understand the service. If experienced staff have had no chance to test the tools, it is harder for them to challenge a demonstration, spot a weak assumption or explain why a proposed process will fail on a busy Monday morning. Early, structured involvement gives them better evidence to bring to those decisions.
None of this gives us a reliable forecast of how many jobs AI will create or remove. It does give us a reason to take the learning seriously. Agencies need people who can assess a proposed change against the realities of public service work, and that ability takes some practice.
Move faster on the learning and be careful with the consequences
There is a big difference in risk between using an approved tool to draft a reply that you check yourself and connecting a tool to incoming enquiries so it can reply automatically. The second changes who sees the output, what information the system handles and how quickly a mistake can affect someone. It needs its own assessment and approval before you take that step.
A useful place to begin is one recurring task you understand well. Check your agency's approved tools, permitted information and training first. Record how the task works now, including the time spent reviewing and correcting it. Then test the AI-assisted version within the approved boundaries and compare the whole task. Include the checking time and the errors, even when they make the result less impressive.
Share what you learn through the appropriate internal channel. Explain where the tool helped, where it needed correction and which parts still depended on your judgement. If the trial reveals that the process itself is unclear, fix that before trying to automate more of it. And if it produces no worthwhile improvement, that is a useful finding too.
If you do not have workplace access, you can still practise with public government information and fictional scenarios. Try drafting a response to a made-up enquiry from a public information page, then check every claim against the source. Keep workplace, personal and confidential information out of that exercise. A prompt asking the tool to stick to its sources is helpful, but you still have to verify that it did.
Start while you have room to learn
AI gives public servants an opportunity to improve some of the work that quietly consumes their day. The people closest to that work have useful knowledge to contribute, provided they get the chance to build confidence with the tools and understand their limits.
Choose a sensible task, get clear on the rules and start learning from what actually happens. Waiting for everything to settle can leave you trying to catch up at the same time you are expected to deliver. Starting carefully now gives you a better chance to help shape how the work changes.