Podcast 🎧 & blog: Governments cannot ignore the agentic AI shift

Written by Federico Plantera, Researcher and Journalist

The Agentic State vision paper, launched at the Tallinn Digital Summit in 2025, describes agentic AI as the most significant transformation in public administration since the rise of modern bureaucracy. Governments have heard versions of that claim before – about e-government, proactive services, digital transformation. But this time, the tools in question can handle a part of government work that previous automation could not: the part that involves judgment. 

In this episode of the Digital Government Podcast, we discuss what this means in practice with Luukas Ilves, co-founder of The Agentic State initiative, former Chief Information Officer of the Estonian government; and Piret Hirv, Head of the Data Management and AI Competence Centre at eGA. Ahead of the Tallinn Digital Summit 2026, what’s in store for one of the chief trends in digital government through the coming year? 

Agentic AI is set around outcomes, not processes 

The countries that have led digital government – Estonia, Denmark, Singapore – built their advantage on deterministic automation: if a process can be fully specified, it can be digitised. What agentic AI changes, as Ilves explains, is the scope of what can be handed over: “All of the sort of automation, digitisation that we’ve really done – what makes your Estonia or your Denmark or your Singapore really good so far – has been about getting deterministic automation right. And suddenly now we can take large chunks of the end-to-end workflows of everything government does, including the judgment bit, and hand that over mostly to the machines.” 

This has a practical consequence he calls attention to: “The biggest bottleneck from the beginning of time for good administration, which is skill and talent of people, becomes much less of a bottleneck.” And a more ambitious one further out – giving an agent a public mandate and letting it optimise across the full set of constraints that make government hard. “Government is never just maxxing on one outcome. It’s maximising on a whole bunch of different things at once: the policy outcomes, the ethics, the rights, the cost.”  

Most governments today define good performance as following a process correctly. Agents allow instead to formalise the outcomes that processes exist to serve, and let the machine figure out how to get there within constraints of legality and cost. As Ilves puts it: “Dear agent, make sure not just that the process gets done, but it gets done in a manner that takes care of the individual user’s rights and the legality of the process. But you figure out, within the constraints of legality and ethics and cost, how to do it.” 

How to manage a flood of agentic queries? 

Governments are not the only ones adopting these tools. On 3 June 2026, Cloudflare’s Radar showed automated bots accounting for approximately 57 per cent of bot-versus-human HTTP requests to HTML pages on Cloudflare’s network. The figure covered all automated traffic, not just AI agents. 

However, citizens and businesses are deploying agents that interact with government services – filing applications, submitting complaints, querying entitlements – whether or not those services were designed for machine-to-machine interaction. Research calls this agentic flooding: “Governments whose services may be digitised, but are not ready for interactions where it’s machines talking to machines, are just overwhelmed – especially where you have a process that’s somehow gated by friction, an application process, a complaints process,” Ilves adds. 

The pressure this creates goes beyond technical friction. Let’s follow a personal example, offered by Ilves in this episode: his bank in Estonia already operates as an MCP (Model Context Protocol) server, meaning his AI agents can access his account information. For the next step – having agents switch his energy provider, arrange a doctor’s appointment, file his taxes – he would tolerate, accept, the level of risks we foresee for the technology today. “If you have a sizeable chunk of your citizens saying, ‘Well, I want to take the risk,’ and your procedures don’t let you do it, then the unhappy path is a legitimacy problem.” The gap is between a population whose risk tolerance for AI use is moving fast and institutions where this is structurally low. 

This creates two parallel tracks that run simultaneously. “One is the short-term push factor – all governments deal with it just because the agents around them are going to be interacting with their services. And then, how do we at the same time maintain this more positive, longer-term, comprehensive transformation?” The question Ilves asks is no longer whether governments adopt agentic AI. It is rather whether they get to shape the terms on which it arrives, or absorb it reactively. 

Scoping authority and accountability, in and outside of government 

Both Ilves and Hirv stress that, next to talking about technology, the hardest problems are actually institutional, as we have witnessed while working with municipalities of Riga and Rotterdam. “Technically, we increasingly know how to connect systems. But institutionally, it is often much less clear who actually owns the processes and who actually is accountable. This is something that deeply impacts the infrastructure that’s needed for building functional and useful digital systems,” Hirv explains. 

Agents sharpen this problem because they operate across organisational boundaries by design. A workflow that used to sit inside one agency’s silo – HR, finance, compliance, and more – can now be handled horizontally, with an agent managing onboarding, transparency criteria, and legal compliance in a single pass. But that means responsibility moves into a matrix, and most governments are not good at matrix management.  

Ilves points to Finland, where central platforms for financial and administrative back-office functions have been built with explicit contracts delegating responsibility across agencies. That makes it easier to deploy agents horizontally, because someone is already accountable for how cross-cutting processes work. Ukraine’s model, in which chief digital transformation officers serve as deputy ministers who also answer to the minister of digital transformation, is another case to keep in sight. But Ilves is clear that these are exceptions: “Working in a matrix like that is really tough for most governments, because you get into the autonomy of different ministries, different bodies in government, the political responsibility.” 

This institutional question connects to a technical one. In the agentic state, identity is no longer just about people and organisations – it includes the agents themselves. Estonia is piloting a framework for giving identity to agents, both internal ones and those arriving from outside. And Hirv offers here a useful operational governance checklist for agentic AI. “Can we identify the agents? Can we restrict their data? Can we restrict their actions? Can we reconstruct what actually happened? Can we revoke authority? Can we stop the agent? Can we restore the service without the agent?” 

Broccoli before steak (but the steak is coming) 

Hirv frames the political economy problem as lying in governments increasingly adopting AI strategies, finding political support, but without addressing the readiness of enablers. “Transformation itself requires less visible expenditure too: data, integration, cybersecurity, workforce development, vendor management, maintenance. Political support without recurrent financing produces only pilots and not transformation.” Baby steps, she says – “You can’t fix the world at once.” 

Ilves, agreeing, reaches for a metaphor he says he borrowed from a minister: “You have to eat this very boring broccoli, and then you get to have the exciting steak and the dessert. Can we create the link between the broccoli and the steak well enough that the politicians say, ‘Okay, I trust you enough, and I give you the mandate and the budget to figure out the broccoli, if in a reasonable time we also get some steak?’“ 

The stakes, he argues, are higher than they were for digital government. “The cost of un-agentic government is going to be much more than a couple percent of GDP. It’s going to be state capacity failures.” And the speed at which governments learn from each other will determine how many of them avoid that outcome.  

Ilves frames it almost as a research question: if Estonia pilots an agent identity framework, or Singapore develops a model for civil servants building their own AI tools, how long does it take for those practices to move from one government to ten, and then to a hundred? “In the world of gov tech and digital government, that speed has been measured in years. And the cost of that speed has been the immaturity that we’ve seen for a long time.” 

This podcast and podcast blog were produced as part of the project “AI-ready Municipalities”, with financial assistance from the European Union via the Technical Support Instrument. The information and views set out herein are those of the author(s) and do not necessarily reflect the official opinion of the European Union or its institutions.  

Interested in more?