Closing the gap between what we say we want and what we do: AI in the public service

At every event, public sector leaders say the same thing. "We want a public service that knows how to use AI to deliver better services for citizens, not a public service that gets replaced by AI." It's obviously the right ambition. But at those same events, we never see how AI tools could be used inside specific departments to deliver a better service. There's no demo, no walk through, no case study, no aha moment. Instead we tend to have helicopter chats, talking about the risk and opportunity of AI in general terms.

Last week I attended The Mandarin Live: Future Ready Public Service conference in Brisbane. The Premier opened the day, and somewhere between his speech and an audience question a few sessions later, something about how we're approaching AI in government came into focus for me. When you hold what's happening on the ground up against the ideal we keep describing, of people using the latest tools to deliver a better service, there's a real gap. The good news is that it's a closeable one.

The QChat moment: a useful lesson before you build your own chatbot

A lot of organisations reach the same conclusion early on: "let's build our own internal tool." The trouble is it's never as good as using Claude or Co-Pilot. People know what that feels like, and this ain't it. Worse still, it's just a bot relying on prompts to provide answers. Some people try it once and think, "if that's AI, I don't get the hype, this is terrible." It's a fair reaction, but a dangerous one, because of course that is not a good example of an AI tool helping a team deliver a better service. Unfortunately everything gets lumped into one "AI" box because of that experience.

The Queensland Premier shared that QChat, Queensland's internally built AI bot, has seen limited adoption. Fewer than one in ten public servants in the state have used it. It's been generally available since February 2024, and on the Premier's own numbers it hasn't quite landed, though nothing has replaced it either. To his credit, he said so plainly. That takes some courage in front of a room of people who are being told to leverage AI while feeling, rightly or wrongly, that they couldn't try a new tool if they wanted to.

We also heard from the team that built it, around five people who genuinely tried to make it work. Anyone who's built a product feels that. It doesn't matter how much effort went in or how many people contributed. What matters is whether the tool delivers better outcomes that the public can see and feel over time. Success isn't how many people prompt it. Are hours being saved? Are approval cycles shorter? Do citizens get answers faster? Are there fewer project variations? For most internally built tools so far, the honest answer is "not yet," and in many cases we haven't even agreed on what success would look like or how we'd measure it. That's not a failure of effort. It's a sign we're still learning how to do this well. But I'd argue we're making that learning curve much harder than it needs to be.

The "get in touch with my team" problem

Later in the program the National AI lead was on stage. Someone in the audience asked the question every public servant in the country is asking right now: how do I, as a public servant who wants to do my job better, actually start using an AI tool? Not a general bot, not a co-pilot, but AI-native software designed to get a specific job done.

The answer was a familiar one: reach out, get in touch with my team.

Everyone in the room understood what that tends to mean in practice. Get in touch with a small, under-resourced central team and join the queue behind every other request. We've created long lines for clarity, and we've done it because it feels safer. The intent is good. But if the vision every Premier in the country describes is to become remotely real, "we use AI tools to deliver a better service, we increase productivity, we empower our teams to do more with less," then you simply cannot have "get in the line" as the answer. The people closest to the work, the ones who would actually know whether a tool saves them time, need a way to:

• Jump online
• Say what they want to trial and why
• Have confirmation as to whether existing tech may already help
• Get support to write up a short business case
• Answer the data and security questions any IT team would ask
• Explain whether a low-risk pilot can be facilitated by the vendor
• State how they will measure success
• Be given the red or green light, fast

This is exactly the kind of workflow Sumday can support, through our decision and procurement platform. If we mean what we say about wanting an AI-capable public service, we have to change how the decisions get made, without sacrificing governance or security of course.

The public sector can't run fast and break things. Time to experiment and break nothing.

Government cannot adopt the "move fast and break things" mindset, and it obviously shouldn't. When the thing that breaks is a citizen's payment, their safety, their data or their trust, the cost is real and it lands on people who didn't choose the risk. Caution here is appropriate.

But there's a difference between moving slowly to manage risk and moving slowly because you haven't figured out how to manage risk.

The opposite of recklessness isn't paralysis, it's structured experimentation. You can give a small team a frontier tool, with proper enterprise terms, a clear scope, human review on anything that matters and a defined way to measure whether it helped. The risk of a careful, well-governed pilot is low. The quieter risk we don't talk about enough is the cost of standing still, while people either continue without AI entirely (not what we say we want) or reach for personal accounts to get their work done.

What "leveraging AI to deliver better services" can look like in practice

If we want a public service that uses AI to do better work for citizens, a few things help.

Give people access to the right tools. Where building your own makes sense, build it. But for most jobs, the faster path is to procure proven frontier tools with proper enterprise terms and let public servants pilot them. Set up a quicker process for a yes or a no. What's the business case? Can you run a small pilot and measure success before committing? Does it meet the bar on ISO, SOC 2 Type II, frontier models, and human review of consequential actions? If yes, let people try it and tell us whether it helps them do their job better, and how. Then you have a real basis for the next step.

Train people on what good looks like. A login is not enablement. People need to see what these tools can do in their own context. The procurement officer needs to see how AI changes a market analysis. The policy officer needs to see how it changes a briefing. The frontline worker needs to see how it changes a case note. This is the work that actually moves adoption. It's also striking how rarely we see any AI tool demonstrated in action at these conferences. There's an opportunity there. Imagine a demo day for the public service, where agencies and providers show what's genuinely possible, side by side.

Put governance where the work is. Risk-based and role-based, with clear guidance on what can and can't go into a prompt. A central team approving every individual use case one at a time is hard to scale, and it can unintentionally signal that good judgement sits somewhere other than with the people doing the work. Strong central standards paired with trusted local decisions tends to move faster and hold up better.

Why this matters

The Premier was right about the vision. A public service that knows how to use AI to deliver a better service is exactly what Queenslanders, and Australians more broadly, should expect from their government.

You don't get there by building your own QChat and hoping adoption follows, and you don't get there by moving so cautiously that nothing happens at all. You get there by trusting people with the best available tools, training them well, designing experiments that are safe by construction and putting governance effort into the decision making process.

That was the conversation I wanted to have in the room and it's the conversation we're having every day at Sumday with the organisations we work with. The leaders getting this right are the ones who've stopped trying to build the chatbot and started focusing on the workflows, the decisions, the high-friction areas where AI can actually support people to deliver a better service.