Jump to content
Sectors
Services
Work
Future Led
Journal
About
Let's talk
More than automation: What agentic AI means for person-centred government
While agentic AI may be grabbing the headlines, automation itself isn’t the end goal. For government, it’s the promise of more connected, trusted and person-centred services that unlock equity for all.
19 Jun 2026
8 min read
AI & innovation

Consider a man in his late fifties who is made redundant. He is pointed to employment services and shown how to claim a payment. What no part of the system registers is that he is also the main carer for a parent with dementia, that the mortgage is becoming unmanageable, and that his own health has started to slip under the strain. Each of those things is somebody’s responsibility. The whole of it is nobody’s.

Or consider a woman who registers the death of her father, then spends the following weeks repeating that single fact to a dozen agencies, each demanding its own proof and its own form, while she is also trying to grieve.

These are not unusual cases. They are the ordinary shape of how people meet government when their lives are hardest. And in almost every one of them, the state already holds the pieces of a fuller response.

Health sits in one place, employment support in another, housing and financial assistance somewhere else again, and community connection somewhere else still. What rarely happens is that anyone brings those pieces together. The person is left to do that themselves, at precisely the moment they have the least capacity for it.

Someone always ends up joining the dots between agencies. Today it is the citizen, doing it from the outside, with none of the information or authority that would make it easy.

When we talk about the possibilities of agentic AI in government, it’s here we have a problem worth solving and a new means of solving it, if we can design for trust and safety.

A person-centred vision worth organising around
This is the ‘North Star’. A public service that recognises the whole of a person’s situation and proactively brings the right support to them, rather than waiting for them to work out which agency owns which service and chase each one down in turn.

Person-centred, not agency-centred.

The technology to attempt this at scale now exists. The harder question is whether government is willing to reorganise itself to use it well.

This is where agentic AI enters, and where it is most often misunderstood. The autonomous agent is the part of the story that grabs headlines, but automating an existing process faster is not the goal, and a faster, broken process is still broken.

The goal is to deliver seamless, proactive service and genuine equity of access.

What agents really do, for now, is less flattering than much of the marketing suggests: they shine a hard light on limitations that are already there. The data silos. The organisational boundaries. The absence of any remit or incentive for one agency to coordinate with another around a single person’s needs.

The friction people feel when dealing with government is partly a real technology problem: services and data sit in systems that were never built to talk to each other, and connecting them is hard, unglamorous work. But that’s increasingly more feasible to address with cross-agency collaboration.

From there, with the uplift in technology making the promise of automation possible, it’s the lessons learned from automation gone wrong that demonstrate what’s needed for safe implementation.

A warning that never quite leaves the room
Robodebt is never far from any conversation about automation in Australian government, and it should not be. It was not artificial intelligence, but it remains the clearest warning of what happens when automated decision-making is aimed at people without meaningful human judgement in between.

That history points to two AI safety principles worth getting familiar with, since they are part of what separates a trustworthy connected service from a reckless one.

The first is bounded autonomy: being deliberate about exactly what an agent is permitted to do, and where a human must step back in.

The second is zero trust, a security principle borrowed from cybersecurity that means “never trust, always verify”. Every request to reach a piece of data is checked, rather than assumed safe because of where it came from.

Keeping a person accountable for consequential decisions, and being transparent about how those decisions are reached, are not constraints on the vision. They are the conditions that make it trustworthy enough to attempt.

It helps to picture what that looks like in practice. Someone applies for housing support on a Sunday night. Rather than the request disappearing into a black box, they can watch it move from agent to agent, much as a parcel can be tracked through Australia Post.

Documents gathered here, eligibility checked against the rules of three different agencies there. Each agent with defined bounds. Each step visible as it happens. If a check fails or something is missing, the person is told at the time and given the chance to fix it, instead of learning weeks later through a rejection letter.

Throughout, they decide what is shared and with whom. When the package is complete and verified, a human caseworker reviews it and makes the decision. The person is spared weeks of repeating themselves at different counters, and the accountable decision still rests with someone who feels its weight.

What good already looks like
There are working examples of this approach overseas.

Ukraine’s national AI assistant, Diia, lets citizens request services through ordinary conversation and completes them end to end, issuing an official income certificate within minutes of a plain-language request.

What makes it instructive is not the speed but the discipline beneath it. A human retains oversight. It verifies every request rather than trusting it, holds to strict data protection standards, and does not give the model direct access to personal data.

In the United States, the City of Boston is building a shared, secure agent layer that lets outside AI systems reach public resources in a governed way, rather than either blocking them outright or letting them roam unchecked.

Both real-world use cases point to the same working definition of a good, connected service: one that earns trust through how it is built, not merely through what it can do.

The encouraging part is that Australia has already started laying these foundations.

The Digital Identity program has been maturing for years, and it is essentially the verify-don’t-trust handshake that made something like Diia possible. A way for systems to confirm what they need to confirm without sensitive information being copied and stored across every department that touches it.

In health, the sector’s adoption of the Fast Healthcare Interoperability Resources (FHIR) standards shows the same thing from another angle: a hard-won agreement on how to share data securely between systems that were never designed to work together.

The building blocks of trustworthy, connected service are not hypothetical. They exist. The open question is whether we join them up, and at what level of ambition.

A leadership challenge, not a tooling one
What stands in the way is more than technology.

The governance frameworks are largely in place: Chief AI Officers across agencies, an AI Safety Institute, and fresh technical guidance from the DTA on deploying agents responsibly.

That work comes from a sensible place given how fast the field is moving and how much is at stake when things go wrong. But a theme keeps surfacing in conversations across government, and it should give the sector pause.

For all the governance now built, the teams responsible for actually delivering services often cannot find their way through it well enough to get projects off the ground. Governance designed to create confidence is instead producing hesitation.

That is a leadership problem, not a tooling one. So, the work ahead belongs to leaders more than to engineers.

It means treating governance as something that gives teams the confidence to move, not a gate that holds them in place. And it means putting human-centred design at the centre of the method, because design is what keeps the whole effort anchored to the real problems people have rather than the tools available to solve them.

Get that wrong and connected services become a faster way to deliver the same fragmented experience. Get it right and the system finally starts to organise around the person.

None of this waits for a more convenient moment. The expectations are already arriving, shaped by every other part of people’s digital lives, and the agents that will soon act on citizens’ behalf are not far behind.

The groundwork for a public service that genuinely lifts the burden off people, rather than adding a more sophisticated layer on top of it, has to be laid now.

The conversation about agents is really a conversation about this, and anyone shaping an agentic strategy inside government should be designing for that future, not just for faster versions of today’s transactions.

What’s in your AI toolkit?
Tell us what you want in the ‘Let’s get real about AI’ 2026 public sector survey by 5pm on June 30. Take the survey.
Results will inform a national report to be published later this year.

The Mandarin

Sam Daley
Partner
Share
Journal
Sign up to our newsletter.

This is Liquid.
Newsletter
LinkedIn
Instagram
We acknowledge the Traditional Owners of country throughout Australia and recognise their continuing connection to land, waters and culture. We pay our respects to their Elders past and present.