The Agentic State: What AI Agents Mean for Government Service Delivery

For the past year I’ve been grappling with the question that sounds simple but turns out to be quite hard: what does it actually mean when an AI agent delivers a government service?

Not a chatbot. Not a better search. A genuine agent — something that understands intent, takes action, coordinates across systems, and completes a task on behalf of a citizen without requiring them to navigate bureaucracy.

The technology is advancing quickly, which makes thinking carefully about it all the more worthwhile.

What’s actually changing

The shift from digital forms to agentic delivery isn’t just a UI improvement. It’s a fundamental change in what government does versus what government enables.

When a citizen fills out a form, they are doing government work. They are translating their situation into administrative language, navigating eligibility rules, and providing evidence the system requires. The cost of this is enormous and largely invisible — it sits in the unpaid labour of citizens.

An agentic system inverts this. Given appropriate authorisation, an agent can understand the citizen’s situation in natural language, determine eligibility, gather evidence from authoritative sources, and deliver an outcome. The citizen articulates a need. The system handles the rest.

The hard problems aren’t technical

So where does the difficulty actually lie? The capability to build these systems exists. What we don’t yet have are adequate answers to:

Who is accountable when an agent makes a wrong decision? In traditional service delivery, there’s a person who made a choice. Algorithmic accountability is genuinely hard, and getting these frameworks in place ahead of wide deployment matters.

How do we preserve the right to human review? Some decisions — around welfare, visa status, child protection — carry consequences that make automated delivery inappropriate without meaningful override mechanisms.

What happens to the public servants whose work changes? Agentic delivery doesn’t eliminate public service roles, but it fundamentally changes them. That transition deserves careful thought and support.

Technical challenges we do need to consider

There are challenges we need to consider around privacy and the citizen right to their own data. Do we see a future in which we need a form of an agentic action ledger. An immutable, queryable record of every action taken on a citizen’s behalf (or as a customer for other systems).

As agencies (and businesses) move towards agentic delivery such as pre-filling, automated compliance checks, AI-assisted decisions, the citizen loses the natural visibility they had when they initiated each transaction. A standard is needed in this mix to restore that transparency - as the action or the event is no longer citizen-initiated, the receipt has to come from the other side.

I’m going to try to talk about this some more in an article on this agentic action ledger.

Something interesting

The Agentic State, a vision paper initiated and supported by the Berlin Global Government Technology Centre and the World Bank, is explicitly trying to get ahead of these questions. The document is a vision paper, not a deployment plan — it’s meant to shape the conversation about what capability we want to build toward, and what conditions need to be in place.

The core argument is that agentic government is coming regardless. The opportunity is to shape it deliberately, with accountability and equity built in from the start.

In the meantime, I’m genuinely interested in what others working in this space think — particularly on the accountability frameworks. Would love any thoughts below.