Beyond Digital Government: The Rise of Agentic Government

Access, not agency

Digital government solved access more successfully than it solved agency. Over two decades, the world’s administrations moved forms online, consolidated portals, stood up digital identity, and made it possible to pay, file and prove things from a phone. What they largely did not change is who acts — who initiates a process, who decides, who is accountable, and who can contest the outcome. The citizen still carries the burden of knowing a right exists, finding the service, and chasing the result across agencies. We optimised the interface of the state far more than its agency.

That gap is about to matter much more, because the technology is crossing a line. AI systems are moving from answering to acting: planning steps, calling tools, coordinating across systems, and carrying a process toward an outcome. The question this raises is not “should government use AI?” — it already does — but “what model of government comes after the digital one, once software can act?” The answer proposed in Andrey Tikhonov’s Agentic State framework is agentic government — and, importantly, that is a claim about institutions, not a synonym for deploying agentic AI in the public sector.

A ladder, not a leap

It helps to see the present as the latest rung on a ladder the field has been climbing for years.

Each rung is real and well documented. “Government as a platform” was articulated by Tim O’Reilly in 2010–2011 and built in practice by the UK’s Government Digital Service, whose shared modules — GOV.UK Notify and GOV.UK Pay (the latter reported to have processed around £8bn) — are used across departments. Digital public infrastructure took it to scale: Estonia’s X-Road, running since 2001, connects more than 450 organisations and powers over 3,000 services without a central data store; India’s Aadhaar identity layer covers roughly 1.4 billion people and UPI became the world’s largest real-time payments system; the GovStack initiative (Estonia, Germany, the ITU and the Digital Impact Alliance) now publishes reusable “building blocks” as open specifications. The OECD has made this a measured norm: “government as a platform” and “proactiveness” are two of the six dimensions in its Digital Government Policy Framework, and its 2025 index records the average score rising as governments advance on them.

The newest rung is the AI one. The OECD’s 2024 stock-take Governing with Artificial Intelligence — the first broad survey of AI across government functions — describes systems beginning to act, not just inform, and warns that monitoring and evaluation are weak enough that risks can go undetected. (This analytical, predictive stage is what the framework’s own five-model comparison calls cognitive government — AI as the state’s analytical core, before it becomes an actor.) That is the threshold. None of these stages is “agentic” in the framework’s sense; each is the ground the next step stands on.

What “agentic” actually adds

If platform government built the infrastructure and proactive government began acting on the citizen’s behalf in narrow cases, agentic government changes two things at once.

First, the unit of action shifts from the service to the life situation — birth of a child, job loss, relocation, a medical event — carried end to end rather than reassembled by the citizen from a dozen separate services.

Second, and more consequentially, the model adds two actors, not one. A government agent acts for the state: it checks eligibility, prepares and executes decisions, takes on the first layer of bureaucracy. But an agentic state with only a government agent is a one-sided machine. So the framework pairs it with a citizen agent — a representative that acts for the person, understands their rights and goals, engages the state at the state’s own speed, demands explanations, and contests decisions. The pairing is the load-bearing idea: it is what keeps automated administration from becoming unilateral.

And because both agents act, the framework treats rights not as an after-the-fact appeal but as part of the architecture. Every significant action should be explainable, logged for audit, scoped by explicit permissions, and reversible through contestation. This is where agentic government parts company with “more AI in government”: the point is not capability, it is bounded capability.

The shadow: administrative power at machine speed

Every serious proposal should name its own failure mode. Here it is concrete, not hypothetical. When public decisions are made or shaped by automated systems at scale, errors propagate at machine speed, and the people affected are slow, partial and outmatched.

The Netherlands lived this. In the childcare-benefits scandal (toeslagenaffaire), a risk-scoring model used by the tax authority helped wrongly brand roughly 26,000 families as fraudsters between 2005 and 2019, with risk flags tied to traits such as nationality; families were ordered to repay tens of thousands of euros, and — as a consequence of the wider scandal, not as a direct act of the algorithm — more than 1,600 children from affected families were taken into care; in January 2021 the government resigned. Australia’s Robodebt scheme automated debt calculation by “income averaging” and raised some 794,000 unlawful debts; a 2023 Royal Commission called it “crude and cruel… neither fair nor legal” and recommended a body to monitor and audit automated decision-making.

Describe these precisely: in each case a machine-generated signal carried too much weight inside a human process, and the individual had neither equal speed nor a real way to contest in time. That is exactly the asymmetry agentic government is meant to address — and exactly the disaster it would amplify if built without the rights frame.

The safeguard

What separates agentic government from automated administrative power is a set of institutional conditions, not a better model:

• Citizen representation — the second agent, owned by and loyal to the person, independent of the authority it challenges.
• Traceability — every meaningful action leaves an auditable record; nothing decisive happens off the books.
• Human responsibility — a named, accountable human remains answerable for outcomes; “the system did it” is not a defence. (Europe’s emerging law points the same way: the EU AI Act requires effective human oversight of high-risk systems and warns against automation bias, while GDPR Article 22 limits decisions based solely on automated processing — context, not endorsement, but a clear direction of travel.)
• Review and contestation — decisions can be explained, challenged and reversed.
• Alternative models — no single system holds a monopoly on seeing and deciding; contestability implies the possibility of another model and another route.

These are not embellishments. Strip them out and you do not get a leaner agentic state; you get the digital Leviathan — the same technology, the opposite regime.

Why this is the right argument now

The temptation, for anyone impatient with bureaucracy, is to read all of this as a productivity story: agents that make government faster and cheaper. That story is real but secondary. The deeper shift is constitutional. When the state can act through software at machine speed, the design question is no longer how to digitise a form; it is how to keep a fast, capable, seeing state answerable to the people it serves. Digital government answered the access question well. Agentic government is the attempt to answer the agency question before the acting state becomes a settled fact.

That is what lies beyond digital government. Not a smarter portal, but a different distribution of agency — with the explicit, deliberate condition that the citizen gets an agent too.

Related

This article develops ideas from Andrey Tikhonov’s book “The Agentic State.” For the definition and principles, see the canonical page: The Agentic State. Related: Agentic vs digital government, Agentic vs cognitive government, about the author.