Nearly all of this missed what is actually being built. What Estonia is proposing is both less dramatic and more interesting than the headlines. We are not (now) aiming to grant agents new rights or legal personhood. What we are looking to do is extend the identity and delegated authority ecosystems Estonia already has to agents acting on behalf of users. Estonia’s plan is remarkable not because it is big but because it is modest, building on the unique identity and delegated authority ecosystem Estonia already has.

(DISCLAIMER: This article contains my personal reflections on public discussions, it does not reflect government policy or internal discussion)

The country where identity is an address

To understand the agent ID, you first have to understand what “digital identity” means in Estonia — because it means something subtly different than in most of the rest of the world.

When identity people abroad talk about digital ID, they usually mean a credential: a token you present to prove something about yourself, ideally revealing nothing more. For instance, the EU Digital Identity wallet, verifiable credentials, most of the modern standards stack are all about issuing credentials.

Estonia’s system certainly has credentials — the ID card, Mobiil-ID, Smart-ID, each backed by keys and PINs. But the deeper layer, the one that makes the whole machine run, is the identifier: the isikukood, the unique personal code every Estonian resident has.

What the isikukood provides is an address in a shared namespace. Every state registry, every bank, every pharmacy, every school portal keys its records to the same eleven digits. When systems talk to each other over X-Road (the data-exchange layer that carries several billion queries a year, the overwhelming majority of them already machine-to-machine) the isikukood is how they agree on who they’re talking about. Furthermore, Estonians now give more than 100 million legally binding digital signatures a year and authenticate themselves using their e-IDs nearly 500 million times a year.

This has a consequence Americans in particular find hard to believe: Estonia has no meaningful identity-theft economy of the “someone got my number” variety. The isikukood is not a secret. It appears on invoices and class lists, and knowing someone’s code gets you precisely nothing (for that you need someone’s identity token and passcodes). The number is an address, not a lock.

On top of this namespace sits the layer that matters most for the agent story: delegation. Estonians act on each other’s behalf constantly, formally, and digitally, and have for years:

• An accountant files a company’s taxes through the Tax and Customs Board’s portal under her own identity, with scoped access rights the company granted her — not by borrowing the owner’s login.

• A parent sees a child’s prescriptions and school records; an adult manages an elderly parent’s affairs through the health portal’s representation functions.

• A corporate bank account has multiple users with individually configured rights and limits — this user can view, that one can prepare payments, only the board member can approve above a threshold.

• For public services, the government has managed a single platform to let businesses manage who may represent them across public services from a single interface.

Every one of these mandates is the same thing: a machine-readable relation between two identifiers — who grants, who receives, what scope, until when, in many cases digitally signed by the individual delegating the auhority.

The mandate table has two columns of isikukoodid (or registry codes). Hold onto that detail. It is the entire logic of the agent ID.

Images: Menu for giving access permissions in the Estonian Tax and Customs Board; A small selection of the 191+ granular permissions I can give for tax and customs reporting alone.

Now consider what separates your accountant from a bookkeeping agent. Not the work — software can already assemble the declaration, reconcile the statements, draft the filing. The difference is that the accountant holds a place in this delegation infrastructure and the software does not. When she acts for you, every system involved knows three things: that she is not you, what you have allowed her to do, and when her authority ends. When you task an agent to do the same work for you, the systems know none of this — they see only you. So the question almost asks itself: if we can give delegated roles to people, why not to agents?

What is actually being proposed

The design under discussion looks more-or-less as follows:

A person (a natural person, and perhaps also a legal entity) requests a linked machine identity: an identifier issued in the same format as the personal code, but from a visibly distinct series (the working assumption: codes beginning with 9), held in a separate register, and bound to its owner.

This linked machine identity is not the agent itself. Rather it is a digital twin* of the individual that is in turn linked to one (or many) agents and their delegations. (*Side note - what terminology do you think Estonia should use? Twin, clone, linked identity, something else?)

As Taavi Kotka notes, a person might run dozens of agents in dozens of versions on any given day; what they need is one separate digital identity, tied to them, that they can entrust to their machines. In other words: a persistent, person-linked identity that your software — whatever software, however often you swap it — can authenticate as and act through.

To it attach mandates, expressed in the same rights-and-scopes logic the Estonian ecosystem already uses: ‘may read bank statements at this bank’; ‘may read health records’; ‘may not transact’.

Because the identifier conforms to the isikukood standard, the thousands of systems that key everything to that standard don’t need re-architecting — a bank’s rights-management screen that today lists your accountant can tomorrow list your twin. Queries about a 9-series code simply route to the agent register instead of the population register. Alongside the twin, the Estonian eID ecosystem — public and private — provides the supporting machinery: authentication for machines, a register, and logging designed so that every action traces through the agent to the liable human.

This is not about giving agents legal personhood; it is not an agent doing anything a human hasn’t authorized; it is not a mandatory registration scheme for all agents; and it is not a permission gate agents must pass before anyone may serve them.

How Estonia’s Proposal Fits Into a Larger Landscape

If you follow the technical side of the agent world, you know that agent identity is currently a hot topic. Anthropic’s Model Context Protocol added server identity to its spec and moved to the Linux Foundation. Google’s Agent2Agent protocol ships “Agent Cards.” The W3C’s decentralized identifier work now explicitly contemplates AI agents; the IETF has competing drafts on agent authentication and on binding an agent’s identity to its owner’s. NIST launched an AI agent standards initiative and published a concept paper on agent identity and authorization in February.

The plumbing is arriving fast, and much of it is good. But notice what kind of thing it all is: grammar. These protocols specify how an agent authenticates, how it presents a mandate, how it calls a tool. What the grammar deliberately leaves to someone else is who vouches for the principal and how a delegation acquires legal effect.

Effectively, there is a division of labor. The world’s protocol layer — OAuth flows, token exchange, verifiable credentials, MCP — is where machine-level agent authentication will happen, and an Estonian identity code series is invisible to that layer. And the twin does nothing for the operational craft of serving machines — hardening interfaces for machine-tempo traffic, rate-limiting, keeping sessions short-lived and replay-proof, remembering that an authorized agent can still be tricked by a poisoned document into doing something foolish within its mandate.

But these technical conventions don’t bridge machine identity to social and legal identity. “I authorize X to read my prescriptions” requires an X — an addressable party to which rights attach. In OAuth-land, X is a client ID scoped to one service, and every service mints its own.

That works, service by service. What it cannot give you is what Estonia’s architecture is uniquely positioned to provide:

• one namespace across the whole society, and therefore one control plane for revocations and governance,

• one audit trail that means the same thing in every system,

• one revocation act — one kill switch — that severs an agent from the tax office, the health record, and the bank simultaneously.

Per-service identity gives per-service legibility. A shared namespace gives societal legibility. And it is why the proposal runs the identifier through the isikukood standard rather than a fresh UUID scheme: the mandate infrastructure Estonia already has — the accountant’s rights at the tax board, the bank’s per-user limits, the central mandate service — is keyed to that namespace. Give the agent an address in it, and twenty years of delegation plumbing lights up for machines with minimal rebuilding.

The most systematic survey so far — a policy memo on designing agent IDs, published this spring by the Singapore AI Safety Hub together with the Singaporean and Korean national AI safety institutes — concludes:

That same memo points to a function not covered by global technical standards:

That is the wager Estonia is making, stated plainly: the standards community will solve the semantics, the handshakes, the cryptographic primitives of agent interaction — and will leave a hole where all of it connects to legal, organizational, and societal structures. The agent ID is Estonia’s bid to fill that hole first, with the delegation infrastructure it has spent twenty years building anyway.

Boundaries and Open Questions

Estonia’s proposal isn’t going to solve for all elements of delegating authority to agents on day one. The current proposal has clear boundaries:

It addresses the structure of delegation, not the substance of mandates: which services actually open to agents, how far, and on what schedule is not being decreed centrally, even for public services. This will be decided gradually, service by service, by public authorities and private providers alike, each applying its own risk appetite. One can imagine a gradual progression up the risk ladder, starting from read-only, with providers raising the cap as capabilities and guardrails prove out.

Nor does it yet give you mandates that are native to machines — richer schemas for instance-level identity, behavioral limits, and semantics that mean the same thing across services. RIA’s Aruait (“reason reserve”) project is prototyping capability manifests, kill switches, and audit hooks.

A further question — whether software can ever hold a signing key the way European signature law assumes a person does — is real but proves less than it seems, because “agent signatures” decomposes into three different problems:

1. Most acts an agent will perform need attribution, not signature: authenticated identity plus valid mandate plus logs, a chain to which national law can give legal effect on its own authority.

2. Where a cryptographic act is genuinely required, the legally operative signature can be the human’s: the principal signs the mandate once, with the tools they already have, and the agent executes inside that signed envelope — and for integrity and origin proofs, Estonia has issued non-human cryptographic credentials for years (the e-seal organizations use to stamp documents automatically).

3. Only the top of the pyramid — an agent executing qualified, handwriting-equivalent signatures — genuinely waits on European rules. And there Estonia should queue actively rather than passively: remote signing with server-held keys also violated the naive reading of sole control once, and it became respectable through ETSI standards with Estonian practice — Smart-ID — as the running code that made the argument. The same play is available again.

Why do this at all — and why now

Everything above explains how Estonia can do this cheaply. The case for why starts from a fact that most governments haven’t yet internalized: the agents are already here, and they are already acting as us.

I wrote earlier this year about what happens when agents meet authenticated government services, and the conclusion bears repeating: eID is not the moat governments think it is. Browser automation drives a logged-in session as easily as a human at a keyboard. Identity verification protects against impersonation, but does nothing to limit the speed and scale at which an authenticated “user” interacts with a system. Absent a well-governed option, citizens in countries with secure eIDs will start handing their credentials to agents. Unregistered agents don’t stay out of the system; they enter it as their owners, invisibly. When they do, the bank and the tax office see a person. They will try to use fancy analytics to detect the agents, and they will fail, because users themselves will work to design around those limitations.

The demand side is not hypothetical. This spring, in a demonstration that quietly made the rounds, Claude Code filed and paid a real tax obligation end-to-end through the Estonian tax authority and a bank — no human touching a UI. LHV, one of Estonia’s largest banks, now ships an MCP server: customers can plug Claude or Cursor directly into their account data, on the bank’s own terms, because — in the bank’s words — customers should decide how and where they use their data, and the bank’s job is to make that safe. Estonian law already lets a provider and a customer agree, bilaterally, to agent-mediated access, if both sides accept the terms.

What a bilateral deal cannot create is the shared rail: the one identity a person can point at every service, the one view of everything their software is authorized to do, the one revocation that severs it all at once. Each provider can build its own fence. Only the ecosystem can draw the map.

There is a sovereignty dimension, too. An agent has two parts: capability and authority. The capability comes from the model, and Estonia will never be fully sovereign in that dimension (frontier models and many digital tools and services are foreign and will stay foreign). Authority is different. The layer that determines Who may act, for whom, within what bounds, revocable by whom can be built in Estonia. The alternative is to wait until platform OAuth grants and acceptable-use policies drafted in Silicon Valley become the effective law.

Estonia has faced this exact question before:: whether identity would be a public function or a platform feature, It answered by building the national version (in a PPP with many private companies) so well that it become the default. The agent ID is the same answer, one level up: keeping the grant of authority — including its outer boundaries, the things no agent may ever be mandated to do, which belong in statute rather than in a usage policy — a public, democratically governed function.

There’s a governance corollary, too: a state can only regulate what it can distinguish. If agent traffic remains indistinguishable from human traffic, Estonian governance of the agentic economy loses its empirical grip before it starts. None of this is sovereignty against the world — the whole design fails unless it interoperates. It is sovereignty in the older sense: that on this territory, the law of who-acts-for-whom is still made in public.

But defense is the smaller half of the story. The larger half is what agentic interaction is actually for — and those benefits, while more speculative, could be enormous. Representation has always been a luxury good: the wealthy (and big businesses) have accountants and lawyers; everyone else has evenings. Delegation infrastructure gives every person and every one-person company what, until now, only those who could afford professional representation have had:

• For consumers. The world’s payments industry is already racing to serve shopping agents: Google’s Agent Payments Protocol, launched with sixty-plus partners including Mastercard and PayPal, exists to answer the merchant’s basic question: did the human actually authorize this? But all of it is scoped to the checkout flow. Estonia’s rail generalizes the same idea to the rest of a consumer’s life: LHV customers already plug Claude into their account data, agents analyze statements and the same mandate logic extends to the insurer, the utility, the pharmacy, the subscription you keep meaning to cancel.

• The business-to-business side follows. One company’s agent requests a quote, the other’s issues the invoice, the first schedules the payment — logged and auditable end to end — while compliance agents run the sanctions checks and flag anomalies in real time. Commerce at machine tempo, with the accountability chain intact.

• For citizen services. Governments elsewhere are building their own assistants as the new front door: Ukraine’s Diia.AI delivers services directly in chat — the first national agent to do so — and GOV.UK Chat is rolling out in the UK government’s app. Estonia can enable a complementary, structurally different path: bring your own agent. Instead of the state building the one assistant everyone must use, the twin gives any agent — yours, chosen by you, the same one that already reads your calendar and your bank statements — a legible, mandated way in. The individual agent can weave together your public data (medical records, social benefits) with private service providers (insurance, gym) and deeply private information (health tracking data) to enable deeply personalized healthcare, education, finances (etc) in a way no government agent could (All while tracking on behalf of the citizen that the government’s use of citizen data remains ethical and legal).

• For compliance and reporting. The monthly VAT and payroll declarations, the annual report, employee registration, statistics returns, customs declarations, grant applications and their reporting, the official notices scattered across half a dozen agency portals — assembled, filed, and reconciled by an agent-accountant working against the company’s actual bank data. A big-company back office at the price of software, for every OÜ in the country.

And the benefits compound: every service that opens to mandated agents makes every other service’s opening more valuable, because the twin, its mandates, and its audit trail carry over. eID and X-Road created their value the same way — not per integration, but per network.

These use cases in turn make Estonia a natural incubator for the agent economy. Skype, Wise, and Bolt all grew from Estonian engineers and founders solving everyday consumer problems. The agent ID sets up the same dynamic for the agentic economy: Estonian founders get years of practice against real agent rails — real mandates, real banks, real liability — before such rails exist anywhere else.

The first wave is easy to name: agent accountants for hundreds of thousands of companies, agent-native compliance and audit for regulated firms, agentic personal health coaches. Existing players will get in on the act too: LHV is shipping one of the world’s first banking MCP servers and Bolt now allows users to search for rides via ChatGPT. The companies that figure out agentic trust on Estonian rails first will be selling that knowledge to everyone else when the world catches up.

But the opportunity space is global: Estonia’s platform need not be limited to Estonians. Since 2014, Estonia’s e-residency program has offered anyone in the world a state-backed digital identity, allowing them to run an Estonian company entirely online, with more than 142 000 e-residents at current count. E-residency proved that a small state can turn its administrative infrastructure into a platform others build on — that “access to a trustworthy, fully digital jurisdiction” is a product the world wants.

The agent ID is the same move for the agentic era, and the value Estonia can offer is concrete: the first jurisdiction where an agent-native company — anyone’s, from anywhere — can operate with legible, scoped, revocable, legally anchored authority across an entire society, banks included, under supervision rather than in the gray. For builders, that’s a runway that exists nowhere else. For everyone else, Estonia becomes the reference implementation the standards conversation currently lacks: 1.3 million people, one namespace, real banks, real government services, real liability rules — the place where the abstractions get debugged in production, and where the lessons, including the failures, become public goods.

What has to go right

I’ll end where a credible explainer should: with the ways this could fail. Estonia’s ecosystem doesn’t just create the opportunity — it creates risks that are specific to Estonia, and the initiative’s success depends on treating them as design requirements.

We are well-versed with the risks arising from our identity architecture. A system where one identity opens taxes, health records, banking, and voting is a system where credential compromise is catastrophic rather than merely annoying. The price of this risk is eternal vigilance. In 2017, when researchers found the ROCA vulnerability in the Estonian ID card’s chips, the state suspended and replaced certificates for the better part of a million cards, publicly and quickly (with no known exploitation). The lesson institutionalized then still holds: this kind of infrastructure buys you enormous convenience and eliminates whole categories of fraud, at the price of standing vigilance over a concentrated attack surface. The agent initiative inherits both sides of that bargain.

There are other risks, too:

Consent fatigue, one level up. The OAuth world’s dirty secret is that nobody reads permission scopes. People who click “allow all” on a calendar app might click “allow all” to their agent delegation.

But at least agent identity gives us a handle on safety: narrow initial mandates, short expiries, aggressive summaries of what an agent actually did. Estonia’s data-tracker helps here - we have the infrastructure to tell citizens who queried their data (eg “Agent X, acting for you, read your health record at 14:32”).

Finally, the project might simply not take off. Building platforms ahead of demand is a bet and the whole thing could linger as a well-designed answer to a question nobody was ready to ask. It has happened to digital infrastructure elsewhere. But building early is also the most Estonian move in this story. The eID and X-Road both shipped years before the demand that justified them; e-residency launched as a half-curiosity and became a platform for a hundred thousand businesses.

The bet is never that you know what the platform will carry, only that nothing can be carried until the platform exists. If the agent ID follows that arc, success will look reassuringly mundane. Nobody will write headlines about robot personhood. A founder’s agent will close the books at month-end, a pensioner’s agent will renew the prescription before anyone noticed it lapsing, and a company on the other side of the world will do supervised business through Estonian rails — because giving your software a mandate will have become as ordinary, as visible, and as revocable as giving one to your accountant. That is the modest, radical thing Estonia is actually building.

ANNEX: Chronology

The intellectual thread for what this linking of agents to national ID could look like runs roughly like this:

In January 2025, Alan Chan, Markus Anderljung, Gillian Hadfield and colleagues published “Infrastructure for AI Agents”, arguing that as agents act in the world, we will need technical systems outside the agents themselves to manage their interactions. They named attribution and identification as the first function of that infrastructure: connecting actions to specific agents, and agents to the people responsible for them.

I picked the thread up in the Agentic State whitepaper (May 2025), which carried the question into government — citing that previous paper directly — and sketched what we called Identity 2.0: alongside verifiable credentials for people and businesses, “cryptographically signed IDs for software agents, delegation tokens spelling out who may act for whom, on which tasks, for how long,” so that every agent transaction starts with the question “who am I, on whose behalf, and am I authorised?”

The updated Agentic State vision paper (October 2025) turned the sketch into architecture: accountability “anchored in a public registry of agents,” and national identity systems “extended with machine identities for AI agents and sophisticated delegation frameworks.” That architecture has since formed the basis of Ukraine’s planned register of government AI agents and of RIA’s Aruait project here in Estonia.

In late 2025, the technologist Vattan PS gave the idea an Estonian formulation — agent residency: every agent needs “a verifiable identity, authorized permissions, and an auditable record of its actions,” anchored to a responsible legal entity, with Estonia’s identity infrastructure as the natural first trust anchor.

At the start of this year, a group of tech entrepreneurs — chiefly Sten Tamkivi and Taavi Kotka (also the initiatoe of e-Residency) — formulated a concrete proposal built around Estonia’s eID infrastructure.

In June, the Eesti.ai Advisory Board chaired by the Estonian Prime Minister endorsed the proposal.

Since then, Estonia’s e-residency team has been leading the work to put together the proposal, with the contribution of a large number of experts and leaders across Estonia’s public sector in consultation with private sector service providers and experts.

Estonia laid out an early vision for AI-enabled government back in 2019, when it became one of the first countries to adopt a national AI strategy. The Bürokratt concept, introduced in 2020, envisioned an agent-based public sector, even if the technology at the time was not yet mature enough to fully deliver it. Recent advances in agentic AI are now making it possible to redesign how the state operates.

This post maps concrete developments in Estonia against the 12-layer Agentic State framework (co-authored by Luukas and Ott). It is intended as a working inventory and will be updated over time as new initiatives, decisions and implementations emerge.

Three strategic milestones frame these developments:

Estonia recently updated its national Digital Society Strategy 2035 to reflect the growing importance of AI adoption at a societal level. The renewed strategy sets concrete targets around AI adoption and trust in public and private sectors. While the AI and Data strategy 2030 and the yearly action plans translate those ambitions into concrete government activities.

The Eesti.ai initiative, launched in early 2026 by the Estonian Prime Minister aims to identify high-value AI deployment projects and brings to bear whole-of-government political support and resources to support their implementation.

Note: links to specific initiatives are included where available, in both Estonian and English.

Layer 1: Public Service Design & UX

Public value: Citizens interact with public services through a single intelligent front door, in their own language, across any channel — rather than navigating dozens of separate agency portals.

• Bürokratt is Estonia’s public-sector virtual assistant, now live across 18 government agencies. It is not a single chatbot but a framework for multi-agent collaboration: a citizen-facing front door that routes interactions to agents and workflows managed by specific government bodies. The architecture is designed so that a citizen asking about, say, a building permit and a tax question in the same conversation gets routed transparently to the right institutional agent each time. Bürokratt is maintained by RIA, Estonia’s digital agency. More broadly, the aim is to enable people to interact with the state in a channel-agnostic way and, over time, receive more personalised and proactive services.

• The eesti.ee citizen portal is the central gateway for public services. Under the DÜAK 2035 plan, the services portal eesti.ee, the government app and Bürkoratt are being merged into personalised, AI-driven interfaces.

• Estonia is also confronting the growing phenomenon of “inbound agents” - interactions with the state will increasingly be initiated not only by people, but by software agents acting on their behalf. As in many countries, public authorities are facing a deluge of AI generated filings, funding applications and queries. But Estonia is also looking to meet this inbound with constructive proposals:

  • The Agent Residency initiative proposes to give every AI agent a digital ID and allow the delegation of rights and responsibilities from individuals and companies to their agents in interactions with public authorities.

  • A signal of where this is heading: in March 2026, developer Stefano Amorelli demonstrated Claude Code executing tax payments through Estonia’s Tax and Customs Board (MTA) and LHV bank, with no human touching a UI.

• The Estonian Language Technology Programme (EKT, 2018–2027) underpins everything in this layer. It funds the development of real-time automatic speech recognition (ASR), neural speech synthesis, sign-language recognition, and avatar creation for Estonian — a language spoken by barely a million people. Much of this work is done by the TartuNLP research group and TalTech Language Technology Laboratory and published as open-source components available through the national Kratid AI component library. Without this programme, Estonian-language AI services would depend entirely on commercial providers with limited Estonian capabilities.

  • A particularly important focus is the development of Estonian-language corpora, which now contain more than 15 billion words. These are intended to strengthen context, cultural understanding, and language performance in both commercial and open models, while also supporting local AI development.

  • A complementary eesti.ai-funded project at the Estonian Language Institute aims to create the first public Estonian-language fine-tuning dataset — at least 10,000 expert human preference judgments across law, administration, education, medicine, and culture — to measurably improve open-source models’ Estonian performance.

Layer 2: Government Workflows

Public value: Faster decisions, lower costs, and civil servants freed from mechanical tasks to focus on judgment and policy.

• The reform of the Administrative Procedure Act (HMS) is one of the key legal enablers of workflow automation in government. Approved by the government on 3 April 2026 and now under discussion in the Riigikogu, it establishes a horizontal legal basis for automated administrative decisions, allowing the widespread introduction of automated administrative procedures without requiring changes to individual legislative acts. Safeguards include automatic notification of decisions, the right to contact a human official, transparency about decision logic, and human review on request. (see post)

• The “10 käsku” (Ten Commandments) is a planned list of government services where the use of AI will be mandatory (e.g. meeting memos, preparing talking points, preparing procurement documents).

• A function repository was launched in March 2026 to address a chronic problem: agencies building the same solutions and workflows separately. Building on Estonia’s existing mandate to make all code publicly available, the function register addresses sharing and reuse at a higher level of abstraction, and will grow into a marketplace of agentic workflows and use cases.

• The function repository is planned to include a separate section for algorithms, functioning as an algorithmic registry that provides an overview of AI systems used across the public sector. This also supports Estonia’s broader direction of making algorithmic transparency more systematic across government, with an ambition to move toward mandatory disclosure requirements.

• The function repository is planned to be paired in 2026 with a prompt register that standardises how officials interact with AI systems, building institutional knowledge about what works.

• A government-wide classifier is being developed by RIA (Estonia’s digital government agency) to automate the categorisation and routing of internal communications across agencies.

• Estonian civil servants report a high rate of using AI in their work, but the adoption of enterprise AI tools is low. Under the Eesti.ai initiative, licenses for approved AI tools will be funded centrally for any civil servant, with plans to conduct pilots of agency-wide adoption and workflow redesign in several agencies.

Layer 3: Policy & Rule-Making

Public value: Laws that are designed from the start for automated execution, reducing compliance costs for citizens and businesses and enabling the state to act proactively rather than reactively.

• Estonia’s lawmaking guidelines are also being updated to require that new regulatory initiatives consider, from the outset, whether AI could address the problem more effectively or reduce the future compliance burden (see post).

• An eesti.ai-funded project (€920K) is building dedicated AI tools for the legislative process: consolidating consultation comments, comparing EU and Estonian law, and detecting contradictions, overlaps, and reference errors in draft legislation. Piloting begins in the Justice and Digital Ministry, then scales to all ministries. Targets include a 25% reduction in the time spent analysing and processing legislative consultations per draft act, and fewer rejections and correction rounds.

Layer 4: Regulatory Compliance & Supervision

Public value: Moving from sample-based, periodic inspections to more continuousdata and AI-driven oversight, and reducing the administrative burden on citizens and businesses — catching problems earlier without increasing the burden on compliant actors.

• Several regulators are already piloting AI for their specific mandates: TTJA (the Consumer Protection and Technical Regulatory Authority) is developing advertisement compliance monitoring; the Gender Equality and Equal Treatment Commissioner is scanning job advertisements for compliance; and the Data Protection Inspectorate (AKI) is automating checks on personal data processing conditions.

• The Real-time economy initiative focuses on the elimination of “unproductive work” performed by businesses, citizens and the state, which can be done through real-time exchange of high-quality data.

Layer 5: Crisis Response

Public value: Using AI to predict, prepare for, and respond to crises faster — from flood risk to medical logistics.

Most activity here is at the pilot or planning stage. Some projects of note:

• A government crisis resilience tool aims to improve coordination and decision-making during major incidents.

• The eKiirabi (e-Ambulance) project is an initiative supported by Eesti.ai to develop AI-powered automated transcription for ambulance incident documentation, enabling paramedics to spend their time on patient care instead of paperwork.

Layer 6: Public Procurement

Public value: Making it faster and cheaper to buy the right things, while making the entire procurement process transparent and data-driven.

• The Ministry of Justice and Digital Affairs has created an assistant to automate the creation of simpler procurements.

• An eesti.ai project aims to automate various labor-intensive steps of the procurement process, including supplier matching, bid invitations, negotiation support for simpler procurements, and IT architecture compliance assessment. The targets include reductions in document preparation time and increases in the number of companies bidding.

• A new procurement plan module consolidates previously scattered annual procurement plans into one searchable, machine-readable system. Scale: 9,000 public procurements per year, €6 billion in annual procurement volume. Velsberg’s framing: “The next phase of digital government is not only about new applications or new AI tools. It is also about making the underlying operational data of the state more visible, machine-readable and usable across institutions.”

• The national procurement register has also added multilingual procurement capabilities and open data access, making Estonian public procurement more accessible to international suppliers, including through EU tenders portal

Layer 7: Agent Governance — Accountability, Safety, Redress

Public value: Ensuring that as government deploys AI agents, citizens can trust them, understand what they do, and challenge their decisions.

This is one of Estonia’s deepest and most distinctive layers.

• TARK provides central trust assessments for cloud-based AI tools used across government. It currently covers ChatGPT, Gemini, Claude, and Copilot, with approximately 20 additional services planned. Each assessment includes risk scoring, usage principles, and mitigation measures. Regulatory and compliance uncertainty — not technology — is the primary obstacle to AI adoption: TARK addresses that by giving agencies a clear, centrally vetted risk profile before they adopt a tool.

• A new AI technical and regulatory sandbox, launched in March, will be rolled out in phases through end 2026, targeting at least 30 AI systems assessed in a national AI risk register. Capabilities include legal and compliance guidance, real-world controlled testing under supervision, secure data processing, local HPC access, algorithm transparency alignment, and audit trails.

• The Trustworthy AI toolbox entered early beta in March 2026. It addresses the observation that public-sector AI typically fails not at the model layer but earlier — at problem framing, data quality, procurement, testing, and accountability. Estonia now has 220+ public-sector AI applications in production; the toolbox aims to move from isolated pilots to repeatable, coherent adoption.

• RIA is piloting an Agent registry to create a unified system for managing AI agents across government. The design includes cryptographic machine identities (extending the national eID to machines), capability manifests defining what each agent can and cannot do, authorisation scopes, and kill switches.

• From an agentic state perspective, EquiTech initiative was launched to detect and prevent bias, discrimination, and opaque decision-making of AI systems. The project built the practical methods, guidelines and training materials needed to ensure that AI systems remain fair, transparent, and trustworthy. Equitech

Layer 8: Data & Privacy

Public value: Citizens control their own data, can see who accesses it, and can grant or revoke consent — while government and business get the data they need to build effective AI.

• The Andmejälgija (Data Tracker) gives citizens visibility into government access and processing of their data. It becomes mandatory for all government agencies by 2028 - a world first in operationalising data transparency at scale. Citizens can view their data trail at eesti.ee.

• The Andmenõusolekuteenus (Data Consent Service) lets citizens grant private-sector entities access to their government-held data. Over 2 million consents have been processed. Estonia is now the first country with a legal basis for consent-based personal data sharing that is already in practice. In healthcare, data consent enabled pre-checkup analysis of medical data that has freed ~100,590 doctor appointment slots and saved ~€900,000 annually.

• The andmed.eesti.ee data portal has grown from ~700 datasets in 2022 to over 8,000 today — a 10x increase driven by treating open data as core infrastructure rather than a compliance exercise. From August 2026 the data portal aims to provide a holistic overview of all government held data and different means how data can be re-used, including X-Road data services, open data, data scientist environment, as synthetic data, using restricted query interfaces and more.

• A Data Management Tools Catalogue is now publicly available, helping organisations match their data governance needs to available tools — covering data catalogues, quality platforms, master data management, modelling, and integration.

• Privacy-Enhancing Technologies (PETs) have a dedicated government funding line, supporting the development of language data spaces and health data spaces. A four-stage PII pipeline — detection, tokenisation, anonymised cloud query, token rehydration — allows AI training on sensitive data without exposing personal information to external providers.

Layer 12: People, Culture & Leadership

Public value: A population and civil service that can actually use, govern, and shape AI — rather than being displaced by it.

• Survey data gives a snapshot: 37% of public-sector workers and 45% of leaders use AI daily. 31% use it for task substitution. 67% believe AI raises productivity, and 35% think their job is at least partially substitutable — but only 2% believe full substitution is possible. Estonia’s updated Digital Society Strategy targets 80% of public and local government employees using AI daily, with a unified training system in place by 2028. Digiriigiakadeemia (Digital State Academy) is Estonia’s central training platform for civil servants.

• Over the next year, Estonia aims for 100 000 Estonians (10% of the adult population) to participate in an AI workshop and 1 000 000 completions of AI-related online education modules. The goal is not just general awareness, but deeper skills for professionals across every sector (including the public sector and governments)

Extra note on sovereign compute

For sensitive and higher-trust use cases, Estonia is moving toward a more sovereign AI compute model in which workloads can rely on local HPC capacity as well as shared European infrastructure such as the LUMI AI Factory. The Ministry of Justice and Digital Affairs assessed Estonia’s future AI compute needs, while also preparing for larger-scale investment choices, including possible participation in an AI gigafactory model. These steps build on significant investments made in infrastructure in 2025, and reflect a broader recognition that trusted AI in government also requires trusted compute.

Defining the Agentic State

Valeria: You’ve said that agentic AI is a bigger shift for government than the previous wave of digital transformation. If you had to explain to a policymaker in one minute — what is the Agentic State?

Luukas: Agentic AI is taking the incredible capabilities that AI has gained in the last couple of years and connecting them to tools — creating software that actually does useful things in the world, that can plan, execute, and perform tasks similar to what a human co-worker does.

The magical power of agentic AI is that it’s a force multiplier for what has always been the biggest bottleneck in most organizations, and especially in public administration: having talented people. Agentic AI suddenly makes talent across many different tasks much more broadly available, and that allows you to do real transformation in how government actually works.

Where this differs from digital transformation is that in practice, a lot of what’s happened has been taking existing 19th-century bureaucratic procedures and digitizing them. The hope with agentic AI is that we can shift the core functions of how government works — copying what’s already happening in the enterprise world, where companies old and new are finding completely new ways to work when human capacity is no longer the major constraint.

Which Government Functions Change First?

Valeria: In your vision paper, you argued that agentic AI can transform twelve core functions of government. Which two or three will change first, and what will it look like?

Luukas: Let’s be inspired by what we already see happening in the commercial world.

The first shift will be in how users — individuals and companies — interact with government. We don’t have to personally go and consume services anymore. We can have an agent engage in transactions on our behalf.

Take a complex life event, like the birth of a child. Government only thinks about the public services part of that portfolio. But when you have a child, yes, there are public services — registering the name, signing up for insurance, applying for maternal support — but you also need to go shopping for diapers, schedule video calls with the grandparents, and a hundred other things. Your personal AI agent can take care of all of that, where public services become just one element of managed complexity. You no longer need to think about what’s a government service versus a private-sector service. You set the goal — have your life managed well in the context of this life event — and let the agents take care of it.

We can imagine similar changes in how regulated companies interact with government: tax payments, regulatory compliance, procurements, and tenders.

The second big shift is in the internal workflows and processes of government. Governments are very focused on doing things procedurally correctly — and for good reason, because process is how we ensure rule of law, transparency, and appropriate decision-making. But what gets lost is actually focusing on outcomes. With agents, we can assign more complex goals while also programming in the proper procedural rules.

Let me make that concrete. Think about tasking an agent with ferreting out corruption. That agent isn’t limited to inspecting individual cases. It can crawl through data, identify patterns, query people, and invent new tests and filters — new ways of detecting deviations from how we expect government to work.

Or take procurement. The simple application is to have AI prepare our procurements. But that misses the bigger picture: procurement is a broken process today. If we give the buying process itself to agents — letting agents from the buyer negotiate with potential sellers — we get a process that is a hundred times faster, much more efficient because it can actually go find suppliers rather than waiting for who comes to you. You get better public value, and you can encode transparency and procedural requirements more consistently than in current human-run procurement.

As we go through the functions of government — tens of core functions, hundreds of thousands of tasks — every one of those can be redesigned, bearing in mind the public purpose but changing the how quite fundamentally.

This is going to be a longer-term process. Andrej Karpathy has said that we’re not in a year of AI agents, but a decade of AI agents, precisely because of the transition challenges. What we’re trying to do with the vision of the agentic state is make sure that governments are not hopelessly behind the rest of the world.

Estonia: From Digital Pioneer to Agentic Leader?

Valeria: Ukraine has been learning a lot from Estonia for all these years — we love Estonia, we love Estonian products, we use Estonian experience, and we are extremely grateful for everything we’ve learned that helped Ukraine become a digital state in Europe. But Estonia is perceived as the best case for digital government in the world, and also sometimes as a kind of finished product. What’s Estonia’s AI strategy, and how does it see the transition from digital state to agentic?

Luukas: Estonia had the good luck in the late 1990s and early 2000s to be going through a fundamental rethinking of its institutions. The end of the Soviet Union meant rebuilding basic governance with a very strong ambition of making those institutions accountable and transparent — at a time of fundamental technological transformation.

What happened was that Estonia had to solve problems like how a justice system works, how to build up healthcare, how to make the administrative apparatus function — and you grab for the means available at the time. Suddenly, that was the internet. The computer and the internet weren’t new, but in the 1990s, everyone had a PC and a browser. So by default, you build services that people can access through a browser. Then you think: if people are accessing services online, I need digital identity, and I need to share data between institutions.

A lot of what’s today a standard digital government toolkit — Estonia was among the first to invent it, and crucially, to deploy it broadly. When you’re building new institutions, you can achieve diffusion. Estonia learned a lot of the lessons of digital government first, and it’s humbling that those lessons have been useful around the world.

We’re now in a similar moment of technological shift. The capabilities of AI are on par with what the World Wide Web and the PC brought in the 1990s, if not more fundamental. And for Ukraine, because you’re in this situation of having to build and rebuild a lot of institutions, you have a lot of dynamism — when change is happening anyway, it’s easier to build in new processes, new technologies.

Now, where Estonia is in all of this: on a vision level, Estonia is very much there. Much of what we talk about in the Agentic State isn’t new. Estonia in 2018–2019 had its “Kratt” vision — its Bürokratt concept, which started to think about what happens when you have agents. The Estonian vision uses this very nice metaphor taken from Estonian mythology: the Kratt is a spirit that executes tasks for you. But if you give it poor instructions or leave it without instructions, it does the wrong thing and creates mischief. That’s a very good metaphor for AI agents.

The problem was that in 2018, the tech wasn’t there yet. We started trying to build a lot of this in 2021, before GPT-3.5, before the transformer models, before AI agents really became broadly available. So Estonia is now paradoxically doing a bit of catch-up because we started too soon.

My hope is that Estonia remains among the early movers. But the challenge is that the country is, as you say, in many ways a finished product — very well-functioning institutions, efficient digital-era ways of working, high satisfaction with services. In the 1990s, we looked at Western Europe and said, “They are so slow, so set in their ways.” Now I look at Estonia with that same anxiety: has Estonian bureaucracy become set in its ways the way Northern European countries are?

The challenge for Estonia, lacking the sense of urgency that Ukraine has, is not doing cool innovation projects — those will happen. It’s having the transformation of how government as a whole works.

The Leapfrogging Question

Valeria: There aren’t that many truly digital states in the world, and we’re already talking about the agentic state. What are the scenarios for countries that aren’t digitized at all?

Luukas: There are probably ten or twenty countries with a consistently high level of digital maturity — your Denmarks, your Singapores, some European and Asian countries. But I don’t think it’s fair to say the rest of the world is undigitized. There’s a huge amount of sophisticated technology use, but it’s not well governed, not consistent, and lacks the central platforms — like Diia, or systems for data exchange and identity — that give you a systemic effect.

Look at the US or the UK. At the application level, in individual agencies and departments, they’re often doing very impressive, cutting-edge things. But between administrative silos, things break down because they can’t exchange data and don’t have common core platforms or governance.

That lack of ability to work as a single enterprise, the being stuck in silos — that’s potentially going to replicate itself in the agentic world. If you have the agentic foreign ministry and the agentic Department of Motor Vehicles and the agentic healthcare system, but they can’t share data or use common platforms, there’ll be a lot of inefficiency.

Whether leapfrogging is possible is an open question — not just in government, but in organizations more generally. Estonians like to say that in the 1990s we leapfrogged Western Europe, but the reality is you still have to do the preceding hard work. Companies and organizations that haven’t done digital transformation are having much more difficulty with AI. There are some very tangible reasons, but I think the big one is data governance: if you don’t have some understanding of what your data assets are and governance processes in place, you really can’t deploy agentic systems at scale.

There’s also a mindset and change-muscle dimension. A lot of the rethinking of how organizations and institutions work that digital transformation has already forced you to start doing — that same muscle you need to use much more intensively in the AI context.

Convincing the Finance Minister

Valeria: How do you convince a finance minister to invest in AI — not as a nice-to-have experiment, but as something that could have a huge impact?

Luukas: You need to start with the categories and boxes a finance minister thinks in: activities, lines of work, outcomes they’re budgeting. Work through the different tasks and functions of government and show how agentic capabilities make a real shift.

Part of that is identifying common functions across government — user interaction, application processing, regulatory supervision — showing these commonalities and where savings can arise.

There are two approaches. One is the CFO planning-function view: find the functions where AI makes a difference and show how it executes better, more efficiently. The other question the finance minister will ask is: how does it save me money?

The tricky part is that the evidence is still emerging. At the firm level, productivity gains from AI are starting to appear, but it’s a lagging indicator. This is repeating the dynamic from the 1990s, when economists said you could see the effect of information technology on productivity everywhere except in the productivity statistics.

A Playbook for Getting Started

Valeria: Some European countries are thinking about moving toward the agentic state without enough institutional capacity. What are the first three to five steps?

Luukas: There’s a central element and a bottom-up element.

The central element is what Ukraine is doing incredibly well: having a core AI transformation team and thinking about the core platforms of an agentic state. There are citizen-facing platforms — you have Diia and Diia.AI as the agentic interface for consuming services — and you have the AI center of excellence.

Platforms are critical because they reduce the marginal cost of each new service. Why digital identity, data exchange platforms, and agentic AI platforms matter is that if you’ve already solved identity and data, those are costs and friction factors you don’t have for each new application.

How do you build this? You need to bring in AI-native talent — Ukraine has done this with an AI CTO and an AI CEO with their own team. But the tricky part is you also need to convert the teams already doing digital transformation. Pretty much every country has teams dealing with digital skills, technology adoption in small businesses, digital transformation of public services, infrastructure. Those people and teams need to become AI-native.

It’s almost a religious conversion: from having a small AI team to having an AI-first digital ministry as a whole. And I think Ukraine has a really mature plan for how to do that, using the platforms and institutions you’ve built over the last five years with the digital ministry.

Part of that has to come from leadership. Companies moving quickly in AI adoption have a CEO who spends significant time using AI, using coding agents. They don’t become the company’s CTO, but they have an intuitive understanding of the technology.

Valeria: So your call to action is that every minister should become an AI ambassador in their ministry and build an AI-first culture?

Luukas: Let’s start with the central unit that deals with technology — whether that’s a digital ministry or a government digital service. That’s the top-down piece: the human element, the policy element, the technological platforms, and the governance and regulatory element.

But there’s also a bottom-up element. You need teams across government — in health, education, defense, the boring back-office functions like your finance ministry and government auditors — to enthusiastically start using this. What’s interesting in this moment of coding agents, which wasn’t true even six months ago, is that non-technical teams can really do a lot of product development and iteration — they can’t do finished products, but they can do a lot — without intensive, expensive tech support.

Countries that are moving quickly are thinking about how to build a culture where a critical mass of product owners, strategic thinkers, and service owners get their hands dirty with coding tools, understand what the technology is capable of, and become consumers of the central platforms.

The risk if governments don’t do this is not that everything continues happily. The rest of the agentic society will start running circles around them. Countries that don’t even have digital identity — people are already tasking agents with filling in their online forms. We’ll soon see people authenticate with eID and then hand off service completion to a browser agent, which will lead to all sorts of interesting problems.

If government IT doesn’t support using agents, if approved tools are a year behind, the part of your civil service that wants to do cool things will use shadow IT and build their own agentic workflows — with massive security problems. And the broader disillusionment will worsen: when your smartphone and consumer services are in a fully agentified world running circles around the state, people won’t think the state is ethical and trustworthy. They’ll think it’s behind, incapable, and they’ll lose even more faith in public institutions.

Valeria: I completely agree about the early adopters. That’s what we’ve seen with digital transformation — specifically with Diia and digital passports and digital driver’s licenses. Those came first, and when people started using them, they began to push and demand more and more services. We’re seeing the same dynamic now with AI.

The Architecture Question: Build or Open Up?

Luukas: There’s a third element beyond top-down and bottom-up: how the agentic state interacts with the agentic society. One of the most interesting design questions is how much of the core agentic capability needs to be built in-house by government versus opening up the data, registers, APIs, and rule sets to the private sector.

Let me make it concrete with the citizen agent. Scenario one: the citizen agent is software developed and managed by the government. In Ukraine, you’ve got Diia — you have distribution, authentication, and an agent managed by government.

But there’s an alternative pathway where people are sitting in ChatGPT, Claude, or Gemini, and those tools are gaining more agentic capabilities. They could have an OpenAI agent or an Anthropic agent that accesses public services via an MCP server. Google and OpenAI would love this because it puts them in the central position of owning the consumer relationship. There are clearly risks, but from a service-delivery standpoint, it might work better over time than building your own agents.

I’m not telling governments to stop building their own agents. If you just throw your hands up and give this function to the private sector, there’s a loss of sovereignty and responsibility. But these architectural design questions are genuinely up in the air right now — in government and in the private sector alike. Amazon tried to kick all the browser agents off their website, and that same dynamic governments will have to think through for service delivery.

Personal Productivity and the Coding Agent Revolution

Valeria: What tools are you using, and how are you learning?

Luukas: The short answer is: get your hands dirty.

I’ve had the good fortune in the last year and a half of not managing large teams, which has given me time to experiment — that’s a real challenge for anyone in an executive function.

I remember when ChatGPT first became a thing, the standard advice was to try doing everything with it: don’t use a browser, don’t use a word processor, don’t use a spreadsheet. You’d figure out what it was good for and what it wasn’t. If you tried that in 2023, you reached a certain conclusion. If you tried again in the second half of 2024, once thinking models came out, you arrived at a new, higher equilibrium. And then with coding agents, you arrive at a third, even higher equilibrium.

Right now, my productivity suite always starts with Claude Code and the data I’m working with — locally or connected via MCP servers. The usual tools — email, Word, spreadsheets — become post-production in some sense.

The tricky bit is that it’s very easy on an individual level to learn this and to be productive, but scaling that up to organizations is more complex.

Valeria: Product placement in this video!

Luukas: [Laughs] No, no — and this is actually a real challenge. I’m sharing my personal experience. I’ve used OpenAI’s tools most intensively for the past couple of years. It’s not that one model is better — Anthropic got some things on the UX and product design really right. And this is a challenge for Google, for Microsoft, for OpenAI to keep up with. You see this in the statistics: in the last year, Anthropic has really displaced a lot of OpenAI’s enterprise market share.

But today it’s Claude Code, and tomorrow that could shift. The things that are good about it aren’t super hard to copy. I’d expect Codex, Copilot, and others to have those capabilities within weeks. So: whatever the coding agent is, get your hands dirty with it.

And for non-technical listeners — one of the magical things about coding agents is that they help you accomplish technical tasks that would previously have seemed out of reach. You need to use an API, set up a workflow, push a snippet of code to GitHub, set up a server — the agent can actually do those tasks for you. A lot of previously mystical capabilities that required a technical team are now things you can do yourself. That doesn’t eliminate the need for technical wisdom — there are many failure paths, especially in an enterprise context with sensitive data — but it turns technical capability into more of an orchestrating, governing function rather than purely an execution function. There’s a parallel there to how we would use AI in healthcare or in education.

AI Sovereignty: A Higher-Order Question

Valeria: What is AI sovereignty? And how does working with big tech relate to it?

Luukas: I actually dislike the sovereignty word precisely because it’s so vague and abstract — it’s a Rorschach test for everyone who uses it. But it does point to a really big underlying question.

There are negative paths, some known from the digital transformation era and some new, that will impinge on sovereignty — whether individual, organizational, or national. We’re in a situation where, at the frontier, there are only three to five major actors providing AI models. Behind that, the infrastructure is dominated by a few hyperscalers and chip manufacturers — almost a monopoly on chip manufacturing, and even with new players, it’ll be an oligopoly.

So what dependencies are you creating — commercially, geopolitically, cognitively? If all the ways we consume information become mediated through this technology, do we even know what’s true anymore? The way we interpret data, process information, summarize things — if the algorithm is biased, implicitly or not, you end up with a different perception of truth. Ukraine is very aware of this because you’ve already confronted anti-Ukrainian narratives baked into LLMs — versions of what is Ukrainian territory, the history of the Ukrainian language and people, that are more aligned with Russian information warfare than with the truth.

Given this massive tapestry of sovereignty challenges, what do you actually focus on? If you try to build the whole stack yourself, true autarky is impossible. Maybe two nation-states in the world could attempt to cover most of it, and even they’re struggling. Everyone else will depend on technology from outside their sphere of control.

There’s a spectrum of approaches. The conservative approach: go slow, only use AI when you’ve validated it, try to build as much as you can yourself. The more liberal approach, which I favor: focus on the functional question of whether you ultimately have control over the processes.

To achieve that, you need to be able to switch quickly and avoid lock-in at any layer of the stack to a single provider. The best way to do that is to actually understand the tech and tools deeply. The more aware and capable a user you are, the less dependent you’ll be.

I could be independent of individual AI providers by not using any of them — but I’d rather use them all, have my architecture set up so I can swap models, have guardrails to detect junk information, and maintain the ability to react quickly if any risk materializes.

Ukraine actually knows this intuitively. What made Ukraine digitally sovereign in 2022 wasn’t a specific architecture. It certainly wasn’t having things on-prem, because suddenly that was the problem. When the bombing started, there was a massive shift from on-prem government systems to the cloud — I played a small part in the Estonian government then, helping save some of the data — and suddenly, the fact that cloud is someone else’s computer outside your territory became a feature, not a bug.

The real thing that made Ukraine sovereign was its ability to move quickly: quickly shifting architecture, constantly adopting new technologies, staying in the race and frequently moving faster than its opponent. That’s been true for the last three or four years of the war: Ukraine’s ability to constantly change tactics, constantly adopt new technologies — more so than any individual technology or supply chain — is what’s given you sovereignty.

That’s a higher-order level of sovereignty. When I think about AI sovereignty, my counsel is: aim for that higher order. Use all the tools, understand them all, be a very powerful user, and maintain the ability to move quickly when circumstances shift.

The main reason to pursue this strategy, and not the “don’t use tools you can’t control” strategy, is that the opportunity cost of not using the tools is enormous. Digital sovereignty should not be viewed as a standalone concept. There’s a broader question of sovereignty — starting from the individual’s freedom and cognitive sovereignty, going to the level of the nation-state — and digital sovereignty is one element of that. I would rather be digitally un-sovereign, but use digital technologies to contribute to my sovereignty as a whole, than have some perfect version of digital sovereignty but therefore be less effective, less agile, and — in the case of countries at war, that means more territory in Russian hands.

Why Ukraine?

Valeria: You know that you’re here in Kyiv just for a few days, and you have a packed agenda. I would like to spend the time we have left to tell the listeners the story of how you’re working with Ukraine. You’re my great friend and a great friend of the Ministry of Digital Transformation and of Ukraine. When the full-scale Russian invasion started in 2022, you were among the first who supported us — with data storage, with building a strong bridge between Ukraine and Estonia. We have done great projects together and learned so much from you. We are incredibly grateful for this partnership, for our friendship. You’ve been working as an advisor to Minister Fedorov, and now also to the acting Minister Alex Bornyakov. You have such great experience working with governments worldwide. Why Ukraine?

Luukas: I had the privilege of working closely with the teams from the digital ministry from 2022 onwards. When I left public service, Mykhailo and you invited me to work with the ministry, and I leapt at the opportunity.

There are two reasons. One is that every Estonian has a deep wish to do something in their line of work that supports Ukraine. Whether you’re an entrepreneur manufacturing saunas, a tech entrepreneur, a lawyer, or someone in government — everyone’s looking for a way to help.

But the flip side is what motivates me in my work: it’s not specific to any one country. We have this technological change, we have the institution of government in democratic societies that we want to keep up. We need models for the world of what good can look like. There are 190-odd countries, half of them hopefully democracies, and below that thousands of jurisdictions and entities that could be adopting this. They’re not all going to reinvent the wheel.

So the question is: where can you do new things first? Where is the innovation and experimentation happening? Ukraine is to no small extent the answer. Partly because the existential pressure of the war focuses you, and partly because of the innovation engine you personally helped build from 2019 onwards. Those two things have merged together in the last couple of years.

There’s the push factor — wanting to help Ukraine — but there’s a very strong pull factor: if I want to do something exciting in my field anywhere in the world, Ukraine has to be part of it. It’s kind of unavoidable.

And the conceptual development, the ideas around the agentic state — yes, some of that was born from the vision Estonia has had for almost a decade and my work as government CIO. But I think that without the interaction with my colleagues in the digital ministry, other government bodies here in Ukraine, the ecosystem, and for a moment thinking “what if we had a blank sheet and reimagined the possible” — it wouldn’t have happened. And I think that’s true for a lot of people in a lot of lines of work.

My call to action to anyone: pretty much whatever your global mission is, there’s probably a really strong Ukraine element where you can do something novel. And the hope is that the post-war, winning-the-peace scenario comes sooner rather than later — and that the massive challenges Ukraine will face become an impetus for disruptive innovation across many sectors, the way the existential challenge of war has driven amazing defense innovation these past few years.

Valeria: What advice would you give to any investor, government, business, or expert who wants to start working and partnering with Ukraine?

Luukas: If I had to give one very generic piece of advice: actually come here and find your way in. In any kind of frontier innovation, human connections matter. Start building them by coming to Ukraine.One of the challenges is that many investors say they want to invest, they see the business case, but they’ll be ready at some future date — maybe linked to the end of the war or other uncertainties. Some of that’s unavoidable. But you can ask: how do we shift the thresholds so that someone who is close to putting in more resources can make that decision?

Different industries, types of capital, and risk expectations will have different specific answers. But the commonality is: be here, develop the connections, ideally work actively with Ukrainian companies in Ukraine. And if you can’t yet, you should absolutely be laying the groundwork now — building connections and preparing so that whenever your risk calculation shifts, you’re ready.

Because if we look at what needs to happen in Ukraine — the rebuilding, dealing with post-war challenges, continuing to build up defense capabilities — those who wake up to this opportunity once it’s completely obvious and low-risk are going to be too late.

This is an edited transcript of an episode of “Forward with Valeria,” a podcast by Valeria Ionan. Watch the original conversation on YouTube.

While some work remains on formal consultation and adoption of the document, this is a good occasion to explain the underlying thinking of Ukraine’s draft strategy - and why Ukraine’s approach is noteworthy in a global context. (Reader beware: this text isn’t a summary of the strategy (use an LLM for that!), but rather an attempt to explain and expand on some of the underlying logic and thinking. And the obligatory disclaimer that what follows is my own opinion).

Most national AI strategies are organized as input lists: invest in compute, train talent, update regulation, fund research, repeat across sectors. These are necessary components, and Ukraine’s strategy contains all of them. But the strategy (as a hypothesis connecting means to ends over the long term) is organized around a different foundational bet: that AI capacity is built through the act of deployment itself.

If we visualize the strategy as a value chain (AI inputs (data/infra/people) → Foundational AI, Responsible/Sovereign requirements, accessibility → Deployment and Transformation - all common components of a national AI strategy), this means the center of gravity is on deployment of AI - and subsequently realizing broader societal value. Everything upstream of that is to be evaluated in terms of whether it contributes to those results.

The Deployment Bet

The strategy adopts an “adoption-first, demand-led approach” — language chosen deliberately. This is a conscious departure from two established models: a research-first approach (invest in fundamental AI research and pilot a lot, hope it diffuses) and a regulate-first approach (establish comprehensive rules before scaling deployment), which has been the EU’s default trajectory.

For a mid-size country with limited resources, neither model works. You cannot out-research the US or China — Ukraine spends 0.3% of GDP on R&D compared to the EU’s 2.3% (and that gap isn’t closing soon). And you cannot afford to wait for a complete regulatory framework before deploying, particularly with the rapid development of the underlying technology.

By material measures, Ukraine’s starting position isn’t great (something the strategy is brutally honest about). While individual AI adoption is high, enterprise AI adoption in Ukraine stands at 4.6%. Due to the war, Ukraine has missed out on the massive data center buildout other countries have seen. Energy infrastructure has been systematically attacked, transforming the country from an electricity exporter into one facing regular deficits. Over 360,000 students are studying remotely from abroad.

Most strategies would bury numbers like these. We chose to lead with them — because the logic of intervention has to be built from where you actually are, not from where you wish you were. If you start from 4.6% enterprise adoption and your target is 75%, you need a fundamentally different approach than a country starting from 30%. You need to create the market, not just facilitate it.

The risk of non-use weighs on the strategy:

“an excessive focus on risk avoidance may significantly limit the scale of deployment and the potential benefits of AI application. … For Ukraine, the significant strategic risk is the inability to ensure wide-scale deployment of artificial intelligence technology.”

In other words, the danger isn’t moving too fast — it’s moving too slowly. This isn’t recklessness. The strategy takes risk seriously. But it treats risk management as something you learn through deployment, not something you perfect before deployment begins.

The strategy is therefore organized around finding domains where Ukraine can deploy now — not where conditions are ideal, but where there is enough existing capability, urgent need, or political mandate to make meaningful deployment possible — and then deploying at scale, deliberately, with the goal of using each deployment to build the surrounding capacity that makes the next one easier. This is what “demand-led” actually means in the strategy: investment in talent, data, compute, regulation, and ecosystem services flows toward the domains where there is real demand for AI being put to use, not toward abstract preparation for deployment in the future.

The implication is concentration. A strategy organized around using deployment to build capacity cannot try to do everything everywhere at once. It has to make choices about where the compounding returns are highest, where existing strengths can be exploited, and where political and operational conditions create room to move fast. The rest of this post is about the choices we made.

Reconstruction - may it begin soon in ernest! - amplifies this logic. Over half a trillion euros/dollars in war damages creates a situation that most countries never face: the need to rebuild entire systems — energy, transport, healthcare, municipal infrastructure. The strategy frames this not as a burden but as an opportunity to build AI-native systems by default, bypassing the legacy integration problems that make AI deployment in established Western bureaucracies so slow and expensive. If you’re building a new energy grid, you can design it with AI-powered management from day one. If you’re rebuilding a hospital system, you can embed AI into clinical workflows at the design stage rather than retrofitting it later. … And the need to deploy massive amounts of capital at speed while complying with gold-standard requirements on transparency, consultation, openness make AI indispensable as a planning and execution aid.

As an aside - while this approach is particularly suited to Ukraine’s circumstances, I would argue this should be the lead

Sequencing from Wartime Learning

The strategy is organized around three waves. Wave 1 focuses on defence, where urgency and concentration of resources allow rapid progress. Wave 2 expands to public administration and the broader economy, translating early capabilities into institutional transformation. Wave 3 consolidates competitiveness through scaling domestic production, innovation, and export.

The important thing here is the logic connecting the waves. Defence isn’t a separate track — it’s the foundation. If you had to pick one sector where Ukraine has world-leading capabilities and an extremely clear demand signal that cuts through all the noise of AI hype, you would obviously pick defence.

Since 2022, Ukraine has deployed AI in unmanned systems, geospatial intelligence, cyber defence, and logistics under real operational conditions. This has generated unique datasets, built teams, established rapid iteration practices, and created operational approaches to data-driven decision-making that no other country possesses.

The strategy ups the ante in several dimensions:

• Speed-to-deployment as a measurable target. Probably the most distinctive operational commitment: the strategy sets a KPI for the time from “confirmation of operational need” to actual field deployment of new defence AI capabilities — under 12 weeks by 2028 and under 8 weeks by 2030. It’s a direct codification of what wartime iteration cycles have taught.

• Institutional architecture. A Defence AI Factory to coordinate development across the government, the Armed Forces, and the private MilTech sector. A Defence Data Coordination Center to manage data flows across defence-sector agencies. CAIO-type leadership roles (Chief AI Officers) in all key defence-sector institutions that hold operational data. Centralized annotation and synthetic-data units for restricted information that individual units can’t process on their own.

• Shared infrastructure. A Military Data Lake as the unified data foundation for defence AI development. A Sovereign AI Forge — a secure national repository for source code and AI models tailored to the MilTech ecosystem, giving Ukrainian defence developers a common substrate they can build on without exposing it externally. An Agentic AI System providing multi-agent analytics across government defence data holders, and an Agentic Defence System providing multi-agent decision support for the command structure of the Defence Forces — both operational by 2028.

• Dual-use and export. Explicit commitment to a dual-use commercialization pipeline, moving battle-tested defence AI into civilian markets and allied procurement. This is both an economic strategy (revenue, ecosystem growth) and a security strategy (interoperability with allies).

• Norms and doctrine. A commitment to “codify the experience of conducting autonomous warfare of high intensity within the framework of international humanitarian norms.” This is Ukraine positioning itself as a contributor to rather than just a consumer of — emerging international standards on autonomous weapons. Ukrainian is today the country with the deepest operational experience of drones and robots in combat - and a deep (and among currently active combatants perhaps unique?) commitment to following the Law of Armed Conflict and international humanitarian law even in challenging circumstances, so this is a non-trivial claim to standard-setting authority.

The strategy explicitly frames defence AI capacity as a national asset that should flow into civilian transformation — not as something that stays behind a classification wall. This is conscious departure from how most countries handle defence AI, as something that lives in a separate world with separate funding, separate procurement, and limited technology transfer to civilian applications. Ukraine’s strategy bets on integration: defence AI as a catalyst that builds the skills, practices, teams, companies and institutional confidence that become broader ecosystem assets.

The (Agentic) State as First Mover

Most national AI strategies assign the state three jobs: regulate, fund, and convene. Some add a fourth — use AI in government — as a separate workstream. Ukraine’s strategy does something different. It positions the state as the primary customer whose deployment at scale helps create the market that the private sector then serves.

The state acts as a “first mover, using public services and reconstruction processes to stimulate the digital transformation of the economy by generating demand.” This isn’t a supporting measure. It’s the core mechanism of the strategy. Public procurement becomes the primary instrument of AI industrial policy — not subsidies, not tax incentives, not research grants, but actual demand for AI-powered services delivered to citizens and businesses.

There’s a spectrum of how seriously different countries take this idea. Ukraine’s version of this idea is reflects a structural reality: While individual AI adoption, especially among younger users, is high, organizations are not systematically deploying AI. Enterprise AI adoption remains low and private capital to invest at firm level is highly constrained. The state is the one actor with both the scale and the mandate to generate sufficient demand, not just in defence but across a range of sectors.

This approach builds on the fact that the state, in Ukraine, already functions as a technology platform in ways most governments don’t. Diia, the government’s digital services platform, has over 22 million users. Ukraine ranks first globally in the UN E-Participation Index and fifth in Online Services. Prozorro, the electronic procurement system, has already demonstrated how public infrastructure can reshape markets.

When you follow this state-as-first-mover logic to its conclusion, you arrive at the strategy’s most distinctive concept: the Agentic State. Ukraine is the first country to formally define this term in a binding strategy document:

A model of public administration that involves the integration of artificial intelligence systems into administrative and operational processes to automate complex routine tasks, enabling the management apparatus to focus on strategic leadership, political accountability, and decision-making.

This is not a chatbot on a government website. It’s a thoroughgoing transformation of the front-end of service provision and internal functioning of government built around the capabilities of Agentic AI, with a governance model that mixes central platforms and leadership with bottom-up business transformation across government. (For more on Ukraine’s roadmap for deploying an Agentic State, see a standalone white paper: https://agenticstate.org/projects/ukraine-roadmap.html).

There’s a political dimension to this too. Ukrainian public support for AI is highest where it’s linked to transparency and fighting corruption. The strategy connects AI adoption to integrity: AI-powered analysis in public procurement through Prozorro, an AI Transparency Panel for disclosing how algorithms are used in public services. This isn’t decorative ethics. It’s a political adoption strategy: deploy AI where it earns legitimacy first.

Governance for Mass Adoption

Wave 2 — the transition from defence-led capability to whole-of-economy transformation — is the hardest bet in the strategy. It’s also where the most concrete commitments live, because this is where “deploy AI at scale” has to translate into sectoral reforms and commitments.

The mechanism the strategy chooses is sectoral missions. Rather than centrally prescribing what AI looks like in healthcare, energy, agriculture, or justice, it asks each priority ministry to define its own mission: a sector-specific ambition tied to outcomes that matter in that sector (not just “AI adoption” as a metric), a roadmap of named measures with deadlines, a core team with political mandate, and KPIs tied to actual sectoral results. The strategy’s own KPI on this is modest but real: ten priority sectors with approved AI action plans and roadmaps by 2030.

This is a deliberate distribution of ownership. The Ministry of Digital Transformation coordinates horizontally — sets standards, provides shared infrastructure, runs the WINWIN AI Centre of Excellence, supports ministries metrics, assessment, training. But the substance of what AI does in each sector is owned by the ministries themselves and their CDTOs and CAIOs.

A word on CDTOs: Ukraine’s networks of CDTOs (Chief Digital Transformation Officers embedded in every ministry and agency and many local and regional governments) creates a unique leverage point. In ministries, CDTOs have the status of deputy minister. Their selection is done jointly by the responsible line minister and the digital minister. This creates a matrix management structure familiar in the private sector but distinct in government. The strategy envisions that ministries (but not necessarily all public bodies) will get dedicated Chief AI Officers within the CDTO’s team to support AI deployment both within the ministry’s public services and within the broader area of governance of the ministry.

Education is where this gets most concrete. The strategy treats education as both a deployment domain in its own right and as the foundational precondition for societal AI adoption. The ambition is to make AI-based learning tools available to 100% of learners in general, vocational, pre-higher, and higher education by 2030 - while also recognizing that “available” doesn’t mean “required”, and that broad adoption can only happen when teachers, pedagogy, standards catch up with what the tech can do.

Alongside the education system, the strategy sets ambitious goals for training the adult population with two distinct skills tracks. The depth track aims to grow Ukraine’s specialized AI/ML workforce from around 6,100 today to 50,000 by 2030, with explicit attention to gender balance. The breadth track aims for 1 million workers with verified, stackable AI credentials by 2030 - at least half of them women.

The connecting logic with the deployment bet matters here. Mass AI literacy isn’t conceived as a parallel training program separate from deployment. It runs through the platforms where people are already engaging — over the past 6 years, Ukraine has run one of the most successful digital skills program in the world, achieving a level of population-level digital skills that exceeds the EU average level. Leveraging the platforms and programmes that already work for AI skilling is the best bet to get to a workforce and population that is broadly capable of working with AI.

What about Ethics and Compliance?

The regulatory architecture follows the same logic learn-by-doing approach. The two-phase approach — soft law from 2026 to 2028, binding legislation from 2028 to 2030 — is a deliberate capacity-building sequence.

Ukraine has for some years already been aligning to EU and European norms on AI governance (2024 Whitepaper). A Voluntary Code of Conduct, already signed by 14 companies, establishes norms before they become rules. Sectoral guidelines (ten are already issued) for civil servants, educators, legal professionals, media, and others create domain-specific expectations. The Council of Europe’s HUDERIA risk assessment methodology gets deployed across public sector institutions. An AI Sandbox Solutions helps companies to test compliance and for regulators to learn what works. All of this builds the muscle, in both industry and government, to actually implement binding regulation when it comes.

In the second phase, a Digital Regulator assumes full statutory enforcement powers. A Law on AI, aligned with the AI Act takes effect. Full alignment with the EU AI Act is completed by 2030.

The thinking behind this sequence is informed by what we’ve been watching across the EU. Even well-resourced member states are struggling with AI Act enforcement capacity. Companies face compliance uncertainty because there are no established best practices yet. The European Commission itself has signaled it wants to simplify implementation. Ukraine is essentially doing what several EU member states wish they could do — building compliance capacity through experimentation before the binding framework locks in. The commitment to EU alignment is real and non-negotiable, given Ukraine’s accession trajectory. But the pathway to alignment runs through institutional learning, not through premature codification.

This is, I think, the governance expression of the strategy’s central argument: you build capacity by doing, not by planning. That applies to regulation just as much as it applies to deployment.

What about sovereignty?

On sovereignty, the the goal is not complete technological self-sufficiency, but strategic autonomy. This means sovereign capabilities where they genuinely matter — defence, critical infrastructure, sensitive government data — and open, internationally integrated approaches everywhere else. Open-source ecosystems and international partnerships provide the base layer.

Ukraine doesn’t pretend it can build a domestic alternative to every global AI service. It focuses sovereignty where the strategic logic demands it and stays open where openness accelerates capability. And most importantly, the biggest enabler of Ukraine’s digital sovereignty is the capability for rapid (re)invention and new tech deployment the world has witnessed since 2022.

Key take-away: AI strategy needs a coherent theory of change

Ukraine’s draft AI strategy envisions AI adoption and deployment as a horizontal, whole of society capability the country can put to use in solving its stickiest strategic challenges. Today, that means winning its war with Russia. Tomorrow, that means rebuilding a society whose public sector delivers responsive, effective, transparent public services and guides a massive set of transformations and reforms. The day after tomorrow, that means commercializing the products and services Ukraine has innovated with into global export successes.

Whether Ukraine can execute on all of this is genuinely uncertain. The gap between the strategy’s ambition and its starting conditions is vast. But that gap is precisely the point. The strategy doesn’t pretend the conditions are favorable. It argues that the right response to unfavorable conditions is not caution but deliberate, structured deployment — because deploying is how you build the capacity that the conditions currently lack.

PS

In the interest of doing a quick pointed overview, there’s a lot of cool stuff I haven’t touched on, e.g, the approach to Ukrainian-language LLMs and AI tools and data governance, that definitely merit their own treatments. You’ll have to read the draft strategy for that ;).

Not a dashboard. Not a platform. Not an enterprise purchased from a vendor after an 18-month procurement. A blinking cursor in a command line interface called Claude Code. Everything flows through it: research, analysis, writing, data work, production. This essay is being written there (and in a markdown file). The work is better and faster, which wasn’t supposed to be possible.

I’ve been using Claude Code and Open Claw obsessively for the past month. The productivity multiplier isn’t 10-15%, it’s 5x. Some of that shows up as more output. Some of that shows up as crazy side projects (one is public, more to come). But a lot of it shows up as output that’s simply deeper, more thorough, and higher quality than anything I could have produced alone.

And so, interestingly enough, a command line has become the center of my professional life. (Judging from my anecdotal interactions and social media, this is true for a sizable community of knowledge workers, but it’s still a sliver of the global population).

My Workflows Today

Claude Code is the workbench. But the more interesting part is the swarm.

I run a small constellation of AI bots (Clawdbot/Moltbot/OpenClaw) on a few virtual private server — different bots with different permissions, different security configurations, different jobs. They run on heartbeats. They check things proactively without being asked, then ping me only when something actually needs my attention.

One scans social media for AI governance developments and logs findings to a research database. One owns my calendar and scheduling. I’m building a few into my side projects - one constantly parsing government procurement feeds in several countries, flagging violations, one upgrading my ‘Legislative Fuckup Finder’. And there’s an admin bot checking the own infrastructure: firewall rules, exposed ports, configuration permissions.

I’m quite conservative about the security posture — these bots have real permissions and real access, which means real attack surface. Each one has scoped access to only what it needs. And they’re not perfect yet - they break from time to time, need resets and help finding the right data. But even with a meaningful correction rate, the net gain is transformational.

And here’s what matters for this essay: I’m policy/product/vision guy with a terminal window and some API subscriptions. If I can build this, anyone with similar motivation can build this. That’s the individual empowerment story.

What does this mean for governments?

• The Upside: the promise of what the hype has been promising from AI ... since forever ... is here now, if only you can enable it. Massive automation, cost savings, better outcomes (and deep, justified embarrassment for those administrations who sit on their hands)

• The Downside and Risk: everything I just described about my own setup, imagine it pointed at government rather than for it.

Let me break my 2026 predictions into two parts - what is inevitable, what will hit governments regardless of what they do - and what the upside of action is:

What Is Inevitable

These aren’t predictions. They’re extrapolations from things already happening. The question isn’t whether but how fast — and whether governments adapt before they’re overwhelmed.

Inbound agents are already here

Individuals and companies are deploying AI agents against government systems today. DoNotPay pioneered the “robot lawyer” — fighting tickets, contesting bills, navigating bureaucracy. Visa appointment bots monopolize consular booking systems. In Australia, a parliamentary inquiry received fabricated submissions with citations to papers and authors that don’t exist. Freedom of Information request volumes are spiking in ways that suggest automation.

These aren’t hackers in the traditional sense. They’re targeting procedural chokepoints — consultations, appointments, adjudication pathways, democratic participation mechanisms. The core dynamic is simple: LLMs increase content elasticity; automation increases interaction intensity. Together, they create a new kind of pressure — bureaucratic denial-of-service — that government systems weren’t designed to handle.

Current legal and security frameworks assume Level 0-1 behavior. The serious harms emerge at Level 2-3. And those capabilities are already accessible to motivated individuals. (If I wanted to wreak havoc and gum up the works and create a bunch of scandals across 20 jurisdictions at once, I know exactly what buttons to press).

Public information will be read by bots

Every government website, every open data portal, every published regulation is now readable not just by citizens but by their agents. The question “is your website agent-friendly?” is about to become as important as “is your website mobile-friendly?” was a decade ago. It’s an entire new dimension of web accessibility (and indeed one of the positive stories here is about how browsing and research agents can help the visually impaired). Governments that haven’t thought about this are already being scraped, parsed, and automated against — whether they intended it or not.

Web-exposed services will be automated

Anywhere government exposes a process through a web interface — permit applications, benefit claims, business registrations, tax filings — agents will automate it. This is already happening with simple forms. It will expand to complex multi-step processes. The forms were designed for humans clicking through at human speed. They will face agents completing them at machine speed, at machine scale. (What happens when an agent makes an error in a submission?)

And eID is not the moat governments think it is: browser automation can drive logged-in sessions just as easily as a human sitting at a keyboard. If you have eID but complexity behind the identity wall, people will want to use agents to navigate that complexity.

The implication is stark: identity verification protects against impersonation, but it does nothing to limit the speed and scaleat which authenticated users interact with government systems. Every service behind an eID login is automatable by the person who holds that eID.

Shadow IT is about to explode

For the last few years, the shadow IT problem was people putting sensitive data into ChatGPT. That’s a confidentiality risk, and perhaps a quality-of-work concern. But the next wave is different.

Civil servants — like people in any large organization — will want to move actual workflows into agentic setups. Not just asking questions. Building bots. Automating processes. Creating the same kind of swarm I run, but inside government networks, on government data, touching government systems. And if corporate IT says no — which it will, because corporate IT always says no first — this will happen on the fringes. Unapproved tools, personal API keys, agents running on personal devices against government systems.

By mid-2025, up to 50% of enterprise employees were already using unapproved AI tools, with 38% sharing sensitive information without permission. Government will be no different. The challenge isn’t “how do we get people to adopt AI” — it’s “how do we govern the AI people are already using without telling us.”

The answer isn’t to ban it — that’s been tried with every previous technology wave and it never works. It’s to set guardrails from the center: top-down standards for security, data handling, and accountability, with bottom-up freedom to experiment within those bounds. The governments that get this balance right will harness the creativity of their workforce. The ones that just say no will drive it underground.

Billions will be wasted on old-school procurements

This is the one that should make finance ministers lose sleep.

Government procurements today are committing billions to build legacy architectures. Multi-year contracts for systems designed around assumptions that are already obsolete. SaaS licenses that will be redundant in months as capabilities commoditize. Integration projects that embed the wrong patterns.

When a civil servant can vibe-code a working case management tool over a weekend, the question shifts from “how do we procure faster?” to “why are we procuring at all?” The need for solid, well-built infrastructures will increase (after all, these agents need to be well governed and sit on top of systems of record) - but the argument for sinking resources into the application layer is going to get thin.

What Is Possible

This is what I think EVERY government should be build right now, in February 2026. (I may need to update this every month of the year)

Agents for accountability — and agents for capability

Government already publishes enormous amounts of data: procurement notices, legislative drafts, regulatory filings, budget documents, statistical collections. Today, oversight of these functions is episodic and resource-constrained. Auditors sample. Journalists investigate when they get a tip. Citizens participate when they have time.

With agents, monitoring becomes continuous. Not sampling — comprehensive. Every procurement checked for threshold manipulation. Every legislative draft scanned for internal contradictions. Every budget line tracked against actuals. This is easy on open data — and transformational for accountability. All of this can be set up on OpenClaw in an afternoon.

And here’s the interesting symmetry: the same agent capabilities that enable external oversight also enable internal excellence. Give civil servants the same tools I’m using. A policy analyst who can direct an agent to synthesize evidence across jurisdictions, build comparative tables, identify gaps in proposed legislation — and do it in hours instead of weeks. A procurement officer whose agent monitors every submission for anomalies and flags what deserves scrutiny. A caseworker whose routine paperwork is handled by an agent, freeing them for the complex cases that actually require human judgment and empathy.

This isn’t science fiction. It’s what I do every day. The gap between my productivity and a typical civil servant’s isn’t about skill — it’s about access to tools and permission to use them. The “fear tax” — civil servants not knowing what they’re allowed to do with AI — is the single most expensive drag on government capability right now.

Agentic interfaces for citizens and businesses

Many governments are now building citizen chatbots and adding in agentic capabilities (allowing users to actually transact). This is good, but it’s already table stakes. Governments need to think more fundamentally - what is the role of government-owned UI and frontend platforms in the context of a rapidly evolving paradigm for all user interactions?

Working with inbound agents

Anywhere government has an API — tax declarations, benefit applications, business registrations — there could be an MCP server enabling richer, more composable interaction. Instead of a citizen filling out a form, their agent negotiates the data exchange. Instead of a business manually checking regulatory requirements across agencies, their agent queries multiple government MCP endpoints and assembles a compliance picture.

This is the positive version of the inbound agents story. If government builds clean agent interfaces, it controls the interaction — authentication, rate limiting, data quality, audit trails. If it doesn’t, citizen-side agents scrape and hack their way through web interfaces designed for humans. The choice isn’t whether agents interact with government. It’s whether that interaction is governed or feral.

Find the high-agent-potential services and push hard

Not every government service needs agents. But some are screaming for them — high-volume, rule-bound, multi-step processes where humans currently do repetitive work that machines could handle better. Tax filing. Permit applications. Benefit eligibility checks. Business registration. Procurement compliance. These are services where the rules are codified, the data is structured (or structurable), and the volume justifies investment. And the win from ‘using agents’ isn’t just high autonomy - it’s also benefitting from the much greater composability of the service stack for building with agents that is rapidly emerging.

A thought: identify the ten services where agent automation would deliver the biggest immediate cost savings and quality improvements, and go all-in. Pick the services where the ratio of human drudgery to judgment is highest, where error rates are embarrassing, where backlogs are growing. Deploy agents there with real resources and real authority. Measure ruthlessly. Then expand from proven wins.

(Re)building the agentic govstack

Agents need plumbing. Unified identity, so agents can authenticate. Interoperable data exchange, so agents can pull from multiple registries in one workflow. Programmable payments, so agent-initiated transactions can settle. Machine-readable policy, so agents know what the rules are.

Where this infrastructure exists, agents layer on top. Where it doesn’t, agents either break or go feral — scraping and hacking around the gaps. And even the governments that built this infrastructure first are discovering it wasn’t designed for agent-scale demand. Data exchange backbones built for human-paced queries buckle under agent volumes. Compute capacity is laughably thin — most governments have a few hundred GPUs when they need tens of thousands. The sovereign compute question — where your agents run and who controls the hardware — is strategic, not just technical.

Most countries have neither unified platforms nor mature data exchange. They’re trying to deploy agents on top of fragmented legacy systems. That doesn’t work. The govstack is the foundation, and for most of the world, the foundation isn’t ready.

Real-time policy simulation

Before you pass a law, run it through agent simulations. What happens when a million agents interact with this new regulation? How does it behave under adversarial pressure? Where are the exploitable gaps?

Machine-readable law isn’t just an efficiency play — it’s a testing environment. The same capabilities that let agents game regulations let governments test regulations against gaming before they take effect.

Reflecting Back on the Vision for an Agentic State

Six months ago, we published the Agentic State vision — a 12-layer framework for how governments should build, govern, and operate AI agents. It remains, as far as I can tell, the most comprehensive structure for government AI adoption that exists. No competing model offers comparable breadth across service design, workflows, policy-making, compliance, crisis response, procurement, governance, data, technology, security, finance, and workforce.

The framework holds up. If anything, subsequent events have validated the urgency more than I expected.

But living the agentic reality daily has revealed structural gaps that theorizing couldn’t.

The biggest gap is directional. The 12 layers are inside-out: they describe what government should build. They don’t adequately describe what government faces from the outside — the inbound agents, the expectation ratchet, the shadow IT explosion, the procurement obsolescence. These are real and mounting pressures. Hopefully, they are an impetus to action.

The second gap is speed. The framework treats iteration speed as an efficiency metric. It should be treated as security posture. If adversarial agents adapt tactics in days and government deploys countermeasures in months, government loses every cycle. Speed isn’t one variable among many — it’s the meta-variable that determines whether any layer’s architecture remains viable.

And the uncomfortable recognition: the disruption is coming faster than I projected. The things I described as future possibilities in October 2025 are things I’m doing personally in February 2026. If one policy person with a terminal window can build what I’ve built, the scale of what’s coming for government is larger than the paper imagined.

The Agentic State isn’t optional. Citizens already have agents. The question is whether government catches up deliberately — with governance, accountability, and public benefit designed in — or gets dragged there by the pressure of a world that’s already moved.

For more, take a look at www.agenticstate.org

Monday evening

Last Monday evening, like everyone else in Estonia, I saw the news about the a slip-up in the recently changed gambling tax law that accidentally made Estonian online casinos tax-free for 2026. The error was almost comically simple: an amendment meant to raise taxes on all remote gambling only referenced “osavusmängud” (skill games), leaving out “õnnemängud” (games of chance). Estimated cost to taxpayers: around €2 million per month until the Riigikogu passes a fix.

My first instinct was to test whether AI would have caught this error. I pulled up the law and ran it through Claude and Gemini. Both identified the problem immediately.

Tuesday morning: what about next time?

The next morning, walking to the office after the kindergarten round, I found myself pondering: what about next time? This law had passed through the sieve of the Ministry of Finance’s lawyers, Riigikogu committee staff, Finance Committee members, and the President’s office. Smart, well-meaning people working under time pressure. The system had set them up to fail.

Could I build something that would catch these errors systematically? And why wait for the state to do it? Riigikogu draft bills are available as open data via API. (Credit to parliament for that; many legislatures around the world don’t offer this kind of access.) Maybe I could wire up something simple that runs every draft bill through AI analysis?

A few hours later, during a free moment, I sketched out rough specs and threw them into the vibe coding platform Lovable.dev. A few minutes later, I had a working prototype.

The rest of the day was spent in 10-minute sessions between work and family, tuning the tool. The main headache was getting the tool to pull full texts without the Riigikogu API timing out. I also experimented a bit with different prompts and models.

By bedtime, the beta version was ready.

Wednesday: from prototype to TV broadcast in 12 hours

The next morning at 7am, I posted the link to “Apsakaleidja” (Estonian for “slip-up finder”). Link here: lawlab.ilves.ai/en. The timing was good for viral spread. Two hours later, the Prime Minister had shared it. The first news stories appeared.

By 7pm that evening, I was in Ringvaade (one of Estonia’s major nightly tv studioes), demoing the app live on air. The morning shows followed the next day.

But the most gratifying feedback came from familiar officials who reported that, inspired by the news, they had already found new slip-ups in draft bills they were working on through Apsakaleidja or their own prompting (and hopefully fixed them too :) ). That’s exactly what I’d hoped for.

The reaction has highlighted the positive side of Estonia’s digital society. Yes, it’s embarrassing that this particular slip-up happened at all. But I was heartened by the rapid, positive reaction - not just in media, but within government too. What matters isn’t that we never make mistakes. What matters is that we learn from them and iterate. Fast.

What’s next?

The tool works, but it’s still a rough beta. I see that at minimum it needs:

Better data pipeline. Apsakaleidja’s analyses are currently based only on draft bill texts. For proper comparison, it should also pull existing law from Riigi Teataja. There’s some complexity here - draft bills don’t map 1:1 to existing laws. The gambling tax amendment was technically “hasartmänguseaduse ja teiste seaduste muutmise seadus” (an act amending the Gambling Act and other acts), touching multiple laws at once.

Smarter prompting logic. The current logic could be more layered and model-agnostic. And the system’s assessment of the criticality of found slip-ups is somewhat random.

Government draft bills. Analysis should be done at an earlier stage of the bill’s lifecycle, when the government is preparing drafts.

Beyond that, all sorts of features could be added: emailing legislation authors when problems are found; letting users evaluate bills against different frameworks; historical trend analysis.

I know colleagues in the public sector are preparing similar tools for internal use. But I like the idea of a tool that belongs to civil society. I’m already talking to a couple of civic organizations who could help maintain the project long-term.

Want to help out? If you’re an experienced prompter/automation pipeline builder or a legislative lawyer, a few hours of your time would be a big help!

The bigger picture

The above is certainly a story about the potential of artificial intelligence, but from multiple angles:

First, of course, the potential of AI in governance and public sector processes. Imagine a state where AI agents empower all public sector functions, from frontline services to policymaking.

A few months ago, I sketched out a vision of the AI-agentic state with the world’s leading AI and digital government technology leaders, available here: www.agenticstate.org.

But behind this specific application (AI-based error checking) lies another moral too:

I’m not a developer. I’m not a data scientist. I’m not a lawyer. In this context, I was operating as an ordinary citizen. Yet vibe coding tools and AI as an assistant enabled me to build Apsakaleidja in a couple of hours alongside my regular work.

Imagine a society where doctors, artists, teachers, students, lawyers, journalists, neighbors create software, play with data, and amplify their impact as easily as they use ChatGPT or Google search today.

What would the collective impact be if just 1% of the population could use tools like Claude Code or Lovable? This, alongside the “agentic state,” is the “agentic society” - where important societal missions and solutions to everyday problems get an enormous, exponential boost.

Esmaspäeva õhtu

Eelmisel esmaspäeva õhtul nägin ülejäänud rahvaga uudist “apsakast” - hasartmängumaksu seaduse veast, mis tegi kogemata netikasiinode 2026. aasta maksuvabaks. Viga oli peaaegu koomiliselt lihtne: seadusemuudatus, mis pidi tõstma kõigi kaughasartmängude makse, viitas ainult “osavusmängudele”, jättes välja “õnnemängud”. Hinnanguline kahju maksumaksjatele: umbes 2 miljonit eurot kuus, kuni Riigikogu paranduse vastu võtab.

Minu esimene instinkt oli testida, kas tehisaru oleks selle vea kinni püüdnud. Tõmbasin seaduse lahti ja lasin selle läbi Claude’i ja Gemini. Mõlemad tuvastasid probleemi kohe.

Teisipäeva hommik: mis saab järgmine kord?

Järgmisel hommikul, jalutades kontorisse pärast lasteaia ringi, jäin mõtisklema: mis saab järgmine kord? See seadus oli läbinud Rahandusministeeriumi juristide, Riigikogu komisjoni töötajate, rahanduskomisjoni liikmete ja presidendi kantselei sõela. Targad, heatahtlikud inimesed, kes töötavad ajasurve all. Süsteem oli nad läbikukkumisele määranud.

Kas saaks ehitada midagi, mis püüaks sellised vead süstemaatiliselt kinni? Ning miks peaks ootama, et riik seda teeks? Riigikogu eelõud on API kaudu avaandmetena kättesaadavad. (Tunnustus parlamendile selle eest; paljud seadusandjad maailmas sellist ligipääsu ei paku.) Äkki saaksin kokku ühendada midagi lihtsat, mis laseb iga eelnõu läbi tehisintellekti analüüsi?

Mõni tund hiljem, vabal hetkel, visandasin umbkaudsed spetsifikatsioonid ja viskasin need loovkoodimise platvormi Lovable.dev Mõni minut hiljem oli mul töötav prototüüp.

Ülejäänud päev kulus töö ja pere kõrvalt 10-minutilistele sessioonidele tööriista timmimiseks. Peamine peavalu oli saada tööriist tõmbama täistekste ilma, et Riigikogu API timeouti läheks. Katsetasin ka veidi erinevate promptide ja mudelitega.

Õhtuseks uneajaks oli betaversiooni valmis.

Kolmapäeva: 12 tunniga prototüübist tele-eetrisse

Järgmisel hommikul kell 7 postitasin lingi “Apsakaleidja”le (lugejale: lawlab.ilves.ai). Ajastus oli viraalseks levikuks hea. Kaks tundi hiljem oli peaminsiter selle jaganud. Ilmusid esimesed uudislood.

Õhtul kell 7 olin Ringvaate stuudios, demomas rakendust otse-eetris. Hommikuprogrammid järgnesid järgmisel päeval.

Aga kõige rõõmustavam oli tuttavate ametnike tagasiside, kes raporteerisid, et olid uudisest kannustatuna Apsakaleidja või enda promptimise kaudu juba leidnud uusi apsakaid enda koostatavates eelnõudes (ning loodetavasti neid ka parandananud :) ). Just seda ma lootsingi.

Reaktsioon on toonud esile Eesti digiühiskonna positiivse külje. Jah, on piinlik, et antud apsakas üldse aset leidis. Aga mind rõõmustas kiire, positiivne reaktsioon - mitte ainult meedias, vaid ka valitsuse sees. Tähtis pole see, et me kunagi vigu ei tee. Tähtis on, et me neist õpime ja itereerime. Kiiresti.

Mis edasi?

Tööriist toimib, aga see on endiselt toores beeta. Ma näen, et minimaalselt oleks vaja:

• Paremat andmetoru. Apsakaleidja analüüsid põhinevad praegu ainult eelnõu tekstidel. Korralikuks võrdluseks peaks tõmbama ka kehtiva seaduse Riigi Teatajast. Siin on natuke keerukust - eelnõud ei vasta 1:1 kehtivatele seadustele. Hasartmängumaksu muudatus oli tehniliselt “hasartmänguseaduse ja teiste seaduste muutmise seadus”, puudutades korraga mitut seadust.

• Targemat promptimisloogikat. Praegune loogika võiks olla mitmekihilsem ja mudeliagnostiline. Ning süsteemi hinnang leitud apsakate kriitilisuse osas on veidi juhuslik.

• Valitsuse eelnõusid. Analüüsi tuleks koostada eelnõu elutsükli varasemas etapis, kui valitsus eelnõusid ette valmistab.

Peale selle võiks lisada igasugu funktsionaalsusi: seaduste autoritele meilitamine, kui leitakse probleeme; kasutajatele võimalus hinnata eelnõusid erinevate raamistike alusel; ajalooliste trendide analüüs.

Tean, et kolleegid avalikus sektoris valmistavad ette sarnaseid tööriiste seesmiseks kasutuseks. Aga mulle meeldib mõte tööriistast, mis kuulub kodanikühiskonnale. Räägin juba paari vabakonna organisatsiooniga, kes võiks aidata projekti pikemas perspektiivis ülal hoida.

Kas soovid appi tulla. Kui oled kogenud promptija/automatiseerimispipeline-ide ülesseadja või õigusloomejurist oleks mõnest sinust tunnist palju abi!

Laiem pilt

Eelnev on kindlasti lugu tehisintellekti potentsiaalist, aga seda mitme külje alt:

Esiteks, muidugi, tehisaru potentsiaalist riigivalitsemises ja avaliku sektori protsessides. Kujutagem ette riiki, kus tehisaru agendid võimestavad kõiki avaliku sektori funktsioone, esmaliini teenustest kuni poliitikakujunduseni välja.

Mõni kuu tagasi visandasin maailma juhtivate AI ja digiriikide tehnoloogiajuhtidega nägemus AI-agentide riigist, mis seda visiooni avab, saadaval siin: www.agenticstate.org.

Ent konkreetse rakenduse (AI põhine vigade kontroll) taga peitub ka teine moraal:

Ma ei ole arendaja. Ma ei ole andmeteadlane. Ma ei ole jurist. Antud kontekstis toimetasin tavakodanikuna. Ent loovkoodimise tööriistad ja AI abiline võimaldasid mul töö kõrvalt paari tunniga Apsakaleidja valmis ehitada.

Kujutagem ette ühiskonda, kus arstid, kunstnikud, õpetajad, üliõpilased, juristid, ajakirjanikud, naabrid loovad tarkvara, mängivad andmetega ja võimendavad oma mõju sama lihtsalt kui kasutavad täna ChatGPT-d või Google-i otsingut.

Milline oleks kollektiivne mõju, kui vaid 1% elanikkonnast kasutaks igapäevaselt tööriistu nagu Claude Code või Lovable? See oleks “agentide riigi” kõrval “agentne ühiskond”, kus olulised ühiskondlikud missioonid ning lahendused argielu tavaprobleemidele saavad meeletu, eksponentsiaalse boosti.

Compiling some info on the Estonian position on European rules on end-to-end encryption (E2EE) and law enforcement access to user data.

In general, Estonia has always been pro-encryption, both in theory and in practice.

• DigiDoc, the government-provided default app for digital signing, includes a functionality for encrypting a digital container with the public key of any Estonian ID-card. This is used routinely to send sensitive but unclassified information or e.g. by sysadmins to distribute new passwords.

• Estonians have been working on post-quantum cryptography for trust services since the 2000s (this was the motivating factor behind Guardtime’s KSI blockchain)

• There is a decent amount of cryptography expertise, notably in Cybernetica and at TalTech.

• And Skype was the granddaddy of end-to-end encrypted communications services.

Estonians recognize that without good encryption, the digital society we have built over the last 20 years breaks down. And that encryption must be widely available and integrated into every layer of how technology works.

Estonian holds no illusion about backdoors. Once they’re there for the good guys, they can be used by the bad guys too. And a small country recognizes that the illusion of government and companies using specialized services while consumers lack good encryption won’t work in today’s ongoing mashup of enterprise and consumer tech.

For all these reasons, Estonia has always been highly skeptical of EU-level legal mandates to “break” E2EE with backdoors, regardless of whether at alghorithmic level or elsewhere (e.g. on device).

Back in the 2010s, when there were active debates on the matter in the European Council, Estonia’s clear position was “no” (I worked on some of this). EU Commission VP Ansip, Estonia’s former PM, also had a clear “no” as Commissioner.

On the EU’s draft Child Sexual Abuse Material regulation, the Estonian government’s official position, adopted in late 2022, says:

Estonia does not support creating backdoors in end-to-end encryption solutions. However, we can support the use of privacy-preserving technologies that allow encrypted content to be analyzed without decrypting it, so that the reliability, security, and integrity of encryption-based digital services are maintained.

The explanatory note further comments:

Such a right to extensive monitoring and control of private messaging would significantly limit users’ rights to privacy and the protection of private life, the confidentiality of communications, freedom of speech and expression, and sexual autonomy. In addition, it would create a dangerous trend in the European legal order by opening the possibility of requiring backdoors in encryption. This, in turn, would increase the likelihood of cyber incidents, because the very existence of such access would effectively invite cybercriminals to look for ways to gain access to other information being exchanged. Creating backdoors reduces the security and integrity of services, since it is not technically possible to make a backdoor available only to service providers themselves or to security and law-enforcement authorities, such that only they could access certain encrypted information.

You can read the full position and explanatory notes here (machine translation is your friend), files at the bottom of the page.

All very sensible. But you will notice there is a “backdoor” of sorts for “privacy-enhancing technologies.” Again, the explanatory note:

Estonia is exploring the use of various privacy-preserving technologies (privacy-enhancing technologies) that allow encrypted content to be analyzed without decrypting it. At the same time, we emphasize that these technologies must not reduce the reliability, security, or integrity of digital services that rely on encryption. For example, one possible such technology is (fully) homomorphic encryption. The potential and use cases of these technologies are still being mapped out.

Helpfully, Estonia has not left open-ended it’s definition and understanding of PETs. The government commissioned a comprehensive study on PETs (already underway when the CSAM position was formulated), completed in 2023 (overview of the study phases and the study itself).

This lets us ask a straightforward question - is the form of end-point scanning envisioned by summer 2025 draft compromise on CSAM aligned with the Estonian concept of PETs (that Estonia might per its position accept as a limitation on E2EE)?

Since there’s a lot of text to analyze, I asked ChatGPT and Grok to take a look, here were their takes:

ChatGPT 5 (link to full analysis):

Although the compromise meaningfully tightens scope, safeguards, and removes grooming, Article 10’s E2EE pathway remains incompatible with Estonia’s stated approval baseline (no backdoors/circumvention; only PETs that analyze encrypted data without decryption). On that strict reading, Estonia could not endorse the draft as‑is without revising its 2022 position or securing additional text changes.

Grok 4 (link):

In essence, the draft enables measures Estonia views as backdoor-like, not PET-based, barring approval under current positions.

fin

Seletuskiri Csam

706KB ∙ PDF file

Download

Download

Poe Csam

343KB ∙ PDF file

Download

Download

Both draw on lessons from my time as Estonia’s government CIO:

1. The End of Bureaucracy: How Estonia, Latvia and Lithuania reinvented governance — and what Germany must do to catch up (published by the Friedrich Naumann Foundation - link).

2. Factsheet: Government Digitization, Efficiency & Performance in the US and Estonia (with Stanford Libraries, Vabamu & the Abundance Institute - link).

I’m deeply grateful to co-authors and interviewees who supplied data and ideas.

I find that the papers make at least two interesting contributions to the policy and academic agendas on digital government:

1. Defining what “good” looks like - in terms of outcomes, not tech. Both papers provide quantitative comparative data on the magnitude of government performance and efficiency improvements unlocked by digital transformation of government. Measured in terms of overhead cost, burden on users or performance metrics, many government services in Estonia are not 20-30% but 3-4 TIMES cheaper/faster than in the US or Germany.

Why is this important? The new US and German administrations have set goals of efficiency (e.g. DOGE) and increased state capacity (“Handlungsfähigkeit”), but they have neither defined those terms nor set many measurable targets. These papers identify metrics and KPIs for user satisfaction, cost and efficacy that should be the basis for clear and measurable goals in the US, Germany and elsewhere.

I don’t claim that the organizational, cultural, governance (and technical) developments needed in both countries are straightforward or easy. Reforming big, complex, federal systems is harder than making progress in Estonia (or Singapore). But that’s the point: the upsides at stake are so great as to motivate difficult efforts.

2. From tech specs → universal uptake: why saturation matters. These papers shift the focus of digital platforms from technical characteristics to uptake and functional integration across both public and private sectors as key drivers of societal benefit. Estonia’s X-road and e-ID are (no longer) unique in their technical characteristics. What continues to stand out, however, are their near universal uptake and the massive volume of transactions they enable. To match Estonia’s usage, transaction volumes for Germany’s mobile digital identity would have to grow 1000-fold (!).

Why is this important? Many digital governments build digital infrastructures and services without a clear business plan for uptake (or the intent to open up API-level access to the private sector). The result: uptake of 10-20% instead of 90%+ and unrealized economies of scale and network effects.

In addition, there are some topics the papers say little about explicitly but that hover in the background:

AI in government. The entire stack of enterprise software and processes is being reimagined in front of our eyes - with government reinvention to follow. Two conceptual points emerge from these papers:

(1) Automation, not (just) AI. Many government services and benefits are deterministic, applying general thresholds or conditions to specific cases. Estonia and others have achieved end-to-end automation of services without AI. Generative AI dramatically lowers the costs of automation and opens new domains and services for automation, but only reinforces the need for the good data governance and thoughtful process (re)engineering used in previous rounds of automation.

(2) We’ve gotten good at measuring the digital maturity of government services, but not the degree of automation. The canonical KPIs for automation and AI in government don’t exist (yet - but I’m on it and would love to collaborate!).

Ukraine has established itself as the standout pupil of e-Estonia, in some areas surpassing the teacher. Ukraine has vaulted to the top of global digital government rankings - but still has a long road to go on delivering broader public goods. This is what I’m working on today - leveraging technology and the innovation machine we’ve seen at play on the battlefield (Brave1, Army of Drones) for reconstruction, good government and growth.

And what can Estonia(ns) take away from these papers? The worst thing that could happen to Estonia is to become a “normal” European country. Estonians are at their best (their most collaborative) when they pursue big goals and best-in-class outcomes. In the words of Steve Jobs: Stay hungry, stay foolish.

I had a lot of fun writing these papers - I hope you enjoy them useful! I welcome your feedback, ideas for further work or opportunities to present the findings.

For other writings and posts, see ilves.ai/publications

Are they building an American panopticon and weaponizing personal data for political fights? Privatizing the data and giving it to the highest bidder (or xAI)? Handing it over to the Russians? That is the speculation in much reporting on the matter, including this recent Atlantic piece by Ian Bogost and Charlie Warzel on “American Panopticon.”

Or is DOGE fixing the federal government’s decades of technical debt, deduplicating data and laying the rails for automated services and AI in the public good?

Image: Elon in an xAI data center, image credit: xAI

Honest upfront disclosure: I don’t know. I have no first-hand information and I’m not privy to the inner workings of this administration. If you’re looking for facts on what DOGE is actually up to, read no further. This essay is about what DOGE (and Congress) could do to ensure citizen control over their data, transparency into the workings of government, and a positive feedback loop of technological and human controls that restore trust.

It bothers me that a raft of reporting on the matter jumps immediately to the worst possible conclusions. They report on DOGE pooling data across once-siloed federal databases - from IRS returns to veterans’ mental-health notes - and intimate that the unstated goal must be a panopticon, both politicized (sensitive data on anyone at the fingertips of political operatives) and privatized (government data sold off to the highest bidder).

This repeats the all-too-common fear mongering I have seen in American (British, German, Japanese…) discussions on government and tech for more than a decade.

Of course I share the central worry: when the state collects and shares *everything* it knows about its citizens without guard-rails, the results can be Orwellian. But there are also compelling public interest arguments FOR pooling and sharing the government’s data. What matters is *how* you integrate, *who* watches the watchers, and *what* rights people have over the resulting system.

Let’s start by reminding ourselves of the upsides that are at stake:

1. Better services. Data reuse powers convenient, automated services. In the world’s best digital government, people file their taxes in minutes, complete all government procedures online and report 80%+ satisfaction with public services.

2. Better policymaking. Decisions based on accurate up-to-date data. Think back to COVID - while the US cobbled together its stats from week-old PDFs and faxes, others made real-time decisions with data fed directly from electronic health records, saving lives.

3. Catching fraud, crime. The US loses 100s of billions to fraud (and fails to disburse allocated funds to those who need it). In many countries, the fraud level is close to zero.

4. Training AI. Large language models built on noisy web scraps hallucinate; models fine-tuned on verified, structured public-sector data could power safer clinical decision support, smarter traffic management or personalised education.

Can we all agree that, in principle, we want this in America? Then let me paint a picture of how this could be DONE RIGHT, drawing for inspiration on the experience of Estonia (a country whose government IT and data practices I know a thing or two about*).

Let’s call it a Panopticon for the people: one that delivers a convincing answer to the question of quis custodiet ipsos custodes (who watches the watchers): everyone!

And while most aspects of doing tech in government right start with the executive, this is something where Congress (or State Legislatures) could take much of the lead:

Rule 1: Everyone “owns” their data.

In Estonia, the answer to “who watches the watchers?” is EVERYONE. A tool (on the government app and portal) lets me track every instance of my personal data being accessed or shared by the government, visible in a single dashboard. I can also give (revokable) permissions for government held data to be shared with private parties (e.g. giving the bank automatic access to my tax returns when applying for a mortgage).

Image: Estonia’s data tracker and data consent, mobile views, screenshot by author.

The logs on the left show a recent prescription renewal, while the consents on the right are (1) to the patient records solution used by the author’s primary care physician and (2) for a mortgage the author applied for in 2022.

This is the gold standard - but it’s also a heavy lift. It requires integration into or standardization of every government system holding personal data, and it requires consistent digital identity to pull together the logs across those systems.

🇺🇸Where to start: In short, doing this across all government services in the US is not where you’d start - but individual agencies (e.g. IRS, DHS-CBP) could start rolling this out for logged-in users right now. (Use Login.gov + state-DMV OIDC under the hood, hashes, not SSNs, as join keys).

In the mean time, there are other rules that are far easier to introduce, even in a messy federal system:

Rule 2: Proactive transparency (no FOIA request needed!)

The Estonian government practices pro-active transparency, with an extensive catalogue of 30+ items that must be posted online without request, including budgets, salaries, audit reports, environmental data. The metadata for ALL government documents (memos, incoming and outgoing letters, personnel orders) is published in searchable online registers, along with every purchase order and transaction (no matter how small).

This isn’t just good for transparency, it saves costs. The bane of FOIA requests, common to officials in most democracies, is largely absent, as the data are already public.

Image: screen grab of payments made by the Prime Minister’s office (circled: a €23.18 payment to one of our major brewers. I guess someone had beer at the office 🍻)

🇺🇸Where to start: Pick a few agencies and make all the records public in a pilot RIGHT NOW. Congress can start updating the Freedom of Information Act to mandate proactive disclosure of a much larger set of data, including correspondence.

Rule 3: Publish the Map (ideally, before you lay the tracks)

In Estonia every public database has a living, machine-readable schema published on RIHA, an asset catalogue of data and information systems. Anyone can see *what* data fields exist, *why* they are collected and *which* law authorises them.

Every machine-to-machine request on Estonia’s X-Road data-exchange layer must have a legal basis (that is transparently visible in RIHA and in the legal gazette - here’s an example of the Social Protection Information System’s terms of reference).

That’s not bureaucracy for its own sake; it is a machine-enforced separation of powers. Legislation decides *purpose*, the executive runs and updates the systems and code, auditors watch the logs, with a tight coherence between what the law says and what the code does.

It is impossible to smuggle a hidden column called “political_affiliation” past that sunlight. For a good example, you can look at the entry for Estonia’s population register (you’ll want to use a browser that auto-translates Estonian into English).

(Note: Estonia is consolidating its RIHA system and open data portal in a few weeks, so some links may not work permanently)

What are the control mechanisms? If the operator of an information system doesn’t provide this information, they are not allowed to connect to any other database over X-road (which basically renders the system non-functional). Also, as an added bonus, the Estonian government opensources its code (here), though this is more of a cost-saving measure.

🇺🇸Where to start: require every federal system of record to file a public “data prospectus” as a rider to the existing Privacy Act System of Records notice. No prospectus, no production database. And for any data pooling, require the API key to be issued only after the Office of Management and Budget (or a revamped Privacy and Civil Liberties Oversight Board) publishes the clause under which data travel. If the purpose disappears—say, a pandemic emergency ends—the key expires.

Rule 4: Log Everything and Give the Logs to the People

Most modern IT systems are transparent and auditable at the machine level. Every keystroke and transaction creates a digital paper trail - but the question is *who* gets to see the logs.

In addition to the individualized transparency of rule 1, Estonia makes the (anonymized) logs of all data exchange publicly available at logs.x-tee.ee. This can in turn be visualized to produce a (real-time!) mapping of all data flows in and around government.

Image: screengrab of visualization tool on logs.x-tee.ee

To avoid any risk of tampering, the hashes of all logs are incorporated into the widely-witnessed KSI Blockchain. The same tech is already used by the US DoD to ensure the integrity of its most critical software supply chains.

🇺🇸Where to start: pretty much anywhere. If DOGE is going to build any new systems (or construct a multi-agency data lake), they should make all of the logs and records around data flows available in a real-time, machine-readable form.

Rule 5: A special (but well-guarded) lane for law enforcement access

Estonia is not naïve about national security or law enforcement. The country borders Russia and is renowned for its successful counter-intelligence and anti-corruption efforts. We allow law enforcement lawful intercept of X-Road queries, but only with a court order and only for the specific datasets authorised. Intelligence services cannot go on phishing expeditions, and the subject—the citizen—is notified post-factum once secrecy is no longer required.

Image: data from the Estonian Ministry of Justice on law enforcement surveillance and legal intercept.

🇺🇸Where to start: Requires federal law enforcement (FBI, Justice Department, DEA, Secret Service, etc…) to publish legal intercept, investigation and transparency request statistics annually and set up a Congressionally mandated watchdog with authority to review all cases.

Bringing it all together

The measures I have catalogued above are mostly technical, but they don’t function in a vacuum - they create the data and rails for human accountability - audits, laws and public pressure - to work, iteratively. The result, in Estonia, is a high degree of public confidence in how the government uses data and tech. The percentage of Estonians who refrain from using government digital services over concerns of privacy or security is tiny (about 0.3%, the lowest rate in the OECD).

I don’t mean to suggest Estonia is perfect. The country has its fair share of data leaks and governance failures. And the architecture of data sharing was designed in a pre-big data era. (If I were starting today, I’d put more emphasis on big data lakes with access control baked in vs. a federated x-road style sharing solution. Central storage is no longer the enemy; opaque storage is. Anything pooled must inherit the same rules-of-engagement: purpose-tagged queries, citizen-visible logs, independent permit keys).

Nor (disclaimer) are the rules above complete. I haven’t really touched on how big data for AI changes the privacy picture (reidentification risk, technical mitigations (row-level access policies, DP noise for mass extracts, secure enclaves for model training)) or procedural safeguards for automated public services.

But the country has gotten the basic OS of a people’s panopticon of accountable, transparent government by and large right.

Want to know more? Conveniently, most of the rules discussed above (and many other principles of how government services work) are contained in one piece of legislation, Estonia’s **Public Information Act**. Link to an up-to-date official English translation here.

🇺🇸Could it work in America?🇺🇸 I don’t think you could recreate overnight everything Estonia has built, but why not start? America invented accountable government for and by the people, maybe it’s time for an update?

*About the author. Luukas Ilves grew up in the US but has spent most of his adult life in Europe. He has served as Estonia’s government chief information officer and in various roles for the Government of Estonia, European Commission and Council of Europe. More

AI will improve our healthcare, spur our research and innovation and boost our competitiveness. We want AI to be a force for good and for growth. We are doing this through our own European approach – based on openness, cooperation and excellent talent. But our approach still needs to be supercharged.

This week, the European Commission releases its first step in that direction, its “AI Continent Action Plan.” I’ve written my first, spontaneous take on the Action Plan, highlighting the kinds of steps the EU could take if it were truly serious about competing head-on with the US and China on AI production and adoption. There is nothing ground-breaking here - just observing (and copying) some of the measures the US, China and others are already taking.

The contrast in velocity and direction between Europe’s AI policy and the new Trump administration (which took office many months after the new Commission) stands out. Agree or disagree with Trump’s politics, the tempo and momentum are remarkable: Within 2 weeks, a first executive order had removed Biden-era AI safety restrictions and made federal land and resources available for data center construction. 9000 (!) submissions have been made to the request for information to prepare a new AI strategy, expected in October. (Ditto the differences between the DC and Brussels bubbles, see note below).

Back to Europe: I’m organizing my comments by the five pillars of the Action Plan (infrastructure, data, sectoral adoption, skills, regulatory simplification), with further comments at the end (for topics not covered in the plan).

1. AI Infrastructure

The EU's plan focuses on three key areas:

• AI Factories: 13 factories selected with EUR 2 billion investment under the EuroHPC initiative, enhancing collaboration and computing capacity for model training and sector-specific applications (e.g., health, manufacturing). Key actions include deploying supercomputers and establishing a single-entry point for users by Q2 2025.

• AI Gigafactories: Proposes up to 5 large-scale facilities for frontier AI models, requiring over 100,000 processors, with the InvestAI Facility aiming to mobilize EUR 20 billion via public-private partnerships. Calls for interest were launched in April 2025.

• Cloud and Data Centre Capacity: Addresses gaps through the Cloud and AI Development Act, aiming to triple EU data center capacity in 5-7 years, with a public consultation launched in April 2025.

Purely in terms of numbers, this pales in comparison with the public and private investments being made in the US and China. US investments in data centers and compute are probably nearing a trillion dollars. China, faced with a shortage of Nvidia chips, is being quite strategic about how it manages and distributes its resources. EU plans seem unambitious in comparison.

The academic consortium model funded as AI Factories has failed to keep up with the training and deployment needs of private sector companies. The AI factories are great tools for researchers and startups building products, but they don’t provide the scalable GPU compute for massive training runs or the ongoing inference needs of massive foundation models. To wit, pretty much all of the useful compute for US AI startups is coming from cloud providers.

Europe also needs to think about a critical dilemma - in the current geopolitical environment there is a strong push to reduce dependencies on the US, yet most of the tooling to build AI scalable applications (not just hyperscalers but many specialized companies) is American (or Chinese). Policy steps that limit access to these tools risks throwing out the AI baby with the geopolitical bathwater.

What if the EU were serious about AI infrastructure? Some ideas:

• Consolidate EuroHPC into a serious undertaking with teeth and set up a national compute bank analogous to NAIRR in the US

• Direct resources planned for AI Gigafactories toward the private sector cloud/AI compute market, with the aim of building up big private sector compute providers with guaranteed (subsidized) demand from EU researchers and startups anchoring the business case. Aim to achieve the AI Gigafactory goals in 2027, not 2030.

• Create a pathway for US companies (hyperscalers and startups alike) to “Europeanize” themselves, securing access to the EU market even in case of a major US-EU trade breakdown in exchange for clear steps to ensure data sovereignty, locate R&D and personnel in Europe, bring innovations to market here.

• Allow Member States to repurpose their remaining RRF and Cohesion funds to support AI infrastructure (especially considering the recent devastating report by the European Court of Auditors - LINK).

• Coordinate AI infrastructure and chips investment, research, supply chains with key Allies (eg. UK, Canada, Japan, South Korea).

2. Data for AI

The Action Plan recognizes that access to large, high-quality datasets is essential and proposes:

• A Data Union Strategy: To be presented in Q3 2025, focusing on ensuring access to high-quality data through a Data Union and reducing bureaucracy.

• Data Labs: Set up within AI Factories to pool and curate data, enhancing interoperability, with examples like the Alliance for Language Technologies (ALT-EDIC) launched in March 2025.

All good, but the differential in data access is just humongous. U.S. AI developers can tap rich troves of data (from social media, e-commerce, cloud services, etc.) with relatively few restrictions (mostly commercially negotiated). China in 2023 established a National Data Bureau in 2023 to coordinate data policy and has been building data trading markets.

What we need is an honest reflection on why data sharing and pooling is so hard in Europe, along with a real rewrite of GDPR (Axel Voss has made good points about what a rewritten GDPR could look like, both in 2021 and this year) and perhaps copyright rules. I get that this is hard - individual data protection rights and copyright are both compelling interests - but at present European societies simply are not able to pull effectively on their collective resources.

What if the EU were serious about Data for AI? Some ideas:

• Publish a European data inventory/map of all available public and private sector data sets and feeds, sort of a data portal on steroids with metadata, legal status and contact points. AI could help put this together and keep it current.

• Data generation grants or tax credits to entities that make their data available for public good.

• Rewrite GDPR. Prioritize availability of AI in the public interest. Provide a safe harbor against copyright and data protection claims under certain circumstances

3. AI Innovation and Adoption in Strategic Sectors

The Action Plan recognizes that Europe’s situation is dire - only 13.5% of EU companies have adopted AI in their core business. Those numbers are much higher in the US and China, driven respectively but cutthroat US competition in many markets and significant public stimulus in China. Adoption requires not just horizontal policy but gettings hands dirty in the details of key sectors.

The EU's AI plan includes:

• A forthcoming Apply AI Strategy to promote AI use in key industries (e.g., manufacturing, healthcare, public services), with the public consultations starting this week.

• Supporting ongoing work of European Digital Innovation Hubs to supporting SME and public sector AI adoption.

• Testing and Experimentation Facilities and the GenAI4EU Initiative: EUR 700 million for research and deployment, with facilities to drive innovation.

What’s missing is a sense of urgency - if corporate Europe doesn’t move soon, European companies’ relative decline in labor productivity vs China and the US will leave them vulnerable to further rounds of acquisition and slow retrenchment in many industries (especially when we consider alongside AI the progress being made outside of Europe in synthetic biology and research into new forms of energy).

What if the EU were serious about AI adoption in strategic sectors? Some ideas

• The Commission should name 5-10 strategic AI adoption missions with 2030 KPIs linked to Horizon Europe challenges, with prizes and funding for major products, e.g.:

  • AI in health diagnostics: 90% of cancer screening read by AI by 2030

  • AI in mobility: 25% of city traffic lights adaptive by 2027.

  • AI in agriculture: real-time crop analytics (or detection of EU subsidies compliance) on 50% of EU farms by 2030.

• Mobilize massive AI-enabled public procurement programs. The EU talks about innovation procurement but in practice applies it to 1-2% of procurements. Europeans could build a voluntary coalition of Member States to require that every large public contract (e.g. over 5m EUR) in transport, education, health, utilities to consider AI-native alternatives. Additionally, governments could buy AI via outcomes-based contracts (“reduce patient wait times 20%” instead of product specs) and set up challenge funds. Finally, the procurement itself could be conducted by Pactum AI.

• Make sectoral AI adoption key political tasks of the relevant Commissioners and DGs (not just Henna Virkkunen’s job). Sectoral strategies need to combine deregulation, data access and investments. Prepare omnibus reform packages in key sectors, actively seeking out inputs from tech companies trying to build new business models in critical sectors (not just playing the usual Brussels game of letting the lobbyists find their way to the Commission’s door).

• One useful output measure (perhaps not a KPI in itself) - how many category-defining sector-specific AI models are made in Europe? E.g. GPT-Health (multilingual, multimodal voice and visual recognition for medical data), GPT-Farm (optimized for life sciences, biochemistry), GPT-Energy (real-time grid control and forestating).

• Create a European AI deployment fund of funds to overcapitalize EU VC and credit institutions in support of AI infrastructure, startups and corporate reorgs (learning from the mistakes and pitfalls of previous EIB funding).

Beyond concrete measures, successful AI adoption requires confronting the EU mantra that SMEs are beautiful and traditional companies should be protected. Past rounds of technological disruption show that two kinds of companies move fastest - new market entrants (startups unencumbered by past ways of doing things) and big players (with the internal resources to invest and reorganize).

Creative destruction does in fact mean that many companies will go bankrupt and positions will be terminated. The trick is to make sure there is something new - in Europe - to replace the old, not to coddle the old.

4. AI Skills and Talent – Incremental Steps vs. US/China Talent Magnetism

Skills and education have always been tough for Brussels, as the most of the funding and policymaking for education and workforce sits at national level. The Action Plan’s steps are modest:

• A plan to enhance Legal Migration for Skilled Workers, including incentivizing European AI experts to return home and reach out to overseas talent.

• An AI Skills Academy: To launch in Q2 2025, with support for increasing the volume of bachelors, masters, PhDs and fellowships in key tech, as well as apprenticeships and a pilot Gen-AI degree. The AI factories are also to function as research hubs.

• Promoting AI Literacy: Through Digital Innovation Hubs for education and training.

Admittedly, much more is being done at the national level in some Member States. But none of it matches the massive boost the US gets from sucking up the world’s combined AI talent (much of Europe’s top AI talent is working in the US) or the volume of domestic talent China has been nurturing.

What if the EU were serious about AI? Some concrete ideas focused on areas where the EU does have some authority:

• Set an EU-wide goal for specific skills categories and types of engineers and track outcomes on a public AI Talent Scoreboard measuring the number of AI MSc’s, PhD’s from each Member State, median salaries, net migration etc. Similarly, create a skills and adoption dashboard with CEDEFOP and Eurostat to measure what individual countries are doing on the AI skills of professionals.

• Fund 1000 AI Chairs with long-term, flexible chairs, modeled after ERA Chairs, and pitch it at top academics in the US. Furthermore, co-fund national initiatives to bring research and commercial talent home.

• Create a model AI skills policy based on existing national best practices (e.g. Estonia’s plan to have all secondary students using GenAI in the next year) and help organize co-financing for Member States that implement best practices.

• Focus on AI skills in professions and specialized domains (e.g. AI in Law, AI in medicine, etc…) by introducing a hard pivot toward AI for Erasmus and Erasmus+ programs and reviewing all of the existing funding for cross-border academic collaboration, student exchanges etc to introduce an AI focus.

The big game here is to maximize the impact of national AI skills initiatives. This is primarily a leadership question (one that Henna Virkkunen, previously Finnish Minister for Education, should be well placed to tackle - it requires national education ministers, employers’ federations etc to get in one room and commit to joint actions at Europan level. Perhaps through a coalition of the willing?).

5. Regulatory simplification

The Commission made waves in February when it withdrew its plans for an AI Liability Directive and appeared to signal a willingness to amend the AI Act (and GDPR) to facilitate AI adoption. The actual steps now taken are much more modest and about supporting compliance, not rethinking the regulation:

• An AI Act Service Desk staffed by a team from DG Connect will answer legal and technical questions about AI Act Implementation.

• Further public consultation on potential measures to simplify AI Act Implementation.

This is worth doing, but doesn’t erase the fact that a European startup building an AI medical device or an autonomous drone will soon face detailed legal checks that an American or Chinese counterpart might avoid until much later in product development.

Like previous EU documents, the Action Plan reinforces the mantra that “AI made in Europe” is trustworthy and high-quality​, but the claim that this is really competitive advantage remains speculative. Remarkably, no other global player has adopted similar rules, with the Trump admin rolling back analogous (but softer) guidelines from a Biden executive order.

Why? Existing liability rules are pretty good and we have way overestimated the risk from AI systems. Fixing Europe’s regulatory mistakes will require an honest rethink of current (quite fresh) regulatory rules, not just fiddling with implementation.

What if the EU were serious about regulatory simplification? Some ideas (short of entirely binning the AI Act and GDPR):

• Rewrite the AI act to create clearer exemptions and exceptions for B2B use cases and turn the AI risk tiers into actual gates, with a legal safe harbor for lower-risk tiers (instead of just listing open-ended obligations).

• Create a pre-certified AI compliance toolkit with pre-audited, open-source documentation templates, testing protocols and compliance checklists and guaranteed approval of even high-risk AI use cases within a limited timeframe (e.g. 30 days).

• Deregulate safe AI applications in well-understood domains (e.g. route optimization in public transport or AI visual inspection in manufacturing) with a white list.

• Unlike the GDPR’s lead authority model, the AI Act permits any national competent authority (NCA) to act, even if the provider is established elsewhere. This should be replaced by some form of passporting. Ideally, at least one EU Member State would step up with a single window for AI companies that combines all AI Act, GDPR, Product Safety and sectoral compliance in one process.

• A focus on sectoral de-regulation (see above).

Further topics

A few key areas where the plan is silent:

• A global role. The EU plan is almost entirely inward focused. Yet with China remaining an autocracy and the US spurning global trade, there is a gaping hole of global leadership the EU could help fill, especially if its approach is truly collaborative, focusing on…

• Open Source AI and open innovation, which have been the breakthrough surprise of the last year (e.g. the DeepSeek moment). Collaboration overcomes secrecy and silos, yielding models and products that are functionally superior, cheaper and with less lock-in and dependency. A preference for open source makes particular sense for an underdog challenger (even China, hardly coming from behind any more, has embraced this approach in its AI competition with the US). The EU’s open source mandate (and funding) should be much stronger (and h/t to Bart de Witte, for whom none of this is a surprise).

• R&D funding for AI needs to be upped drastically, perhaps not with new money but reviewing existing grants and awards. EU funding, once awarded, usually lets you quietly chug along for the duration of a 3-5 year research project. This means there is lots of research in the pipeline that should be completely rethought given advances in generative AI.

• AI in Defense and Security: Unlike the U.S. and China, the EU’s plan largely ignores military and intelligence applications of AI.

A few general observations on process and the debate:

• The Brussels bubble is timid. There are almost no “radicals” (strong accelerationists, AI adoptionists) in the EU policy debate. Even the good industry initiatives (e.g. the AI EU AI Champions’ Initiative) keep their recommendations in vague and general terms.

• Ambitions need to be defined in terms of outcomes, not just inputs and steps. That’s the only way we can avoid misinvestment aimed at satisfying political constituencies instead of the outcome-driven R&D that e.g. DARPA is great at.

• European policy-makers love supply-side policy measures, but they need to focus on adoption and demand. Unless we see a sudden AGI takeoff (AI-2027-style) the big value add from GenAI is going to come from the application layer and sector-by-sector implementations. Without losing sight of its R&D and frontier capacities, this is where Europe should focus its creative energy.There are lots of non-AI specific policies that are also super critical. To simplify - implementing the recommendations of the Draghi and Letta reports, especially with regards to capital/funding and a functional single market.

Finally, does my call for deregulation and swifter adoption mean I’m calling for Europe to abandon its focus on ethical AI and fundamental rights? No. But my proposals here and elsewhere to recognize some key truths:

1. The best way to build ethical AI is to be building AI at all. The best way to be competent at regulating and pushing new markets and tech toward better outcomes is to have your hands dirty, experimenting, learning, being empirical about risks.

2. If Europe is left in the dust moralizing about European values will have little value. Companies are already going around the European market instead of choosing to comply with European rules. This will only worsen.

ABOUT ME

I’ve spent most of the last 15 years making tech policy in Europe, both in Brussels and Estonia. Most recently, I was Estonia’s undersecretary for digital transformation (responsible for digital infrastructure, cyber security and government technology in what many call Europe’s most digital country). I am currently supporting Ukraine’s digital policy and a non-resident fellow at ICDS, a defence think tank in Estonia. Everything written here is my own opinion.

Transcript

So, first of all, apologies for the really ugly slides—I'm not a great designer. I want to talk about this word that you hear a lot, in different permutations: digital sovereignty, tech sovereignty, data sovereignty. I'd like to demystify it and then try to think in practical terms about how we deal with a world where these words matter.

Sovereignty - a 17th century idea

It’s important to go back to the roots of “sovereignty” and how we’ve talked about it for the last couple of centuries. In a broader sense, it's older even than the 17th century, but our modern concept of sovereignty really comes from that era.

We have the idea of the sovereignty of the state: the Sovereign or Authority that governs a piece of territory, with the modern definition of the state being the entity that exercises a monopoly on the legitimate use of force in a certain area.

Beyond the collective sovereignty of states, there's also the sovereignty of the individual—another 17th-century idea—that every man (it was a sexist era) is sovereign over his own “castle,” a kind of personal domain at home. From that came important concepts about privacy and inviolability of the personal home, which underlie many of our modern liberal ideas of human autonomy.

All of these notions of sovereign states and sovereign individuals were very much grounded in a physical world. Now, we live in a world with a virtual layer overlaid on the physical layer—call it the “metaverse” or something else—so we’re no longer just living in a world of atoms, but also a dematerialized world of electrons. Those electrons do not easily fall into the borders of any classic notion of sovereignty.

We’ve been dealing with these problems for a while now:

For instance, over a decade ago, the U.S. tapped into Google and Facebook data—arguably for good reasons like catching terrorists—but that upset people, because private communication and domestic conversation was suddenly less private.

There’s also the possibility that a company like Huawei, or theoretically Nokia or Ericsson, could insert itself into a country’s telecom infrastructure and siphon off data.

Another example is when a government wants to see people’s communications, but Apple defends its users’ privacy via strong device encryption. Then there are the ways social media platforms—often headquartered outside a given country—can play a major role in democratic discourse, such as allegations that TikTok influenced Romanian elections.

It’s also worth noting that this discourse is only going to intensify as technology continues to evolve. We already see this emerging in the realm of neurotechnology, where direct access to the human brain raises questions about cognitive or neurological sovereignty:

Two thinkers who have tried to grapple with all this are Balaji Srinivasan and Bruno Maçães

Balaji, a Silicon Valley investor, has promoted the idea of the network state: that many of the functions of a state may no longer be tied to territory, and instead can be distributed via digital networks. He was inspired partly by Estonia’s approach to digital identity, data embassies, and e-residency.

Bruno, a former Portuguese foreign minister, takes a somewhat similar premise but goes further, suggesting that we already live in virtual worlds that are essentially built by the U.S. or China. According to him, anyone else is just playing in one or both of those worlds—an unsettling prospect if you’re not the U.S. or China.

The issue is all the more stark when we look at developments in artificial intelligence. Most big AI models are coming out of the U.S. and China. The rest of the world might compete for a small share, but the lion’s share of AI patents, AI startups, and large-scale AI development is dominated by these two giants.

If a serious third competitor emerges, it may not be a traditional state at all; it could be an autonomous artificial intelligence. We already see the beginnings of AIs that can replicate themselves, plus distributed computing and storage paid for by cryptocurrencies—censorship-resistant technology that, in theory, could let an AI roam unbound.

Where does this leave the rest of us? If we’re not the U.S. or China, and if we’re not potentially an emergent AI, do we have any hope of real digital self-determination? The question also applies to many Americans and Chinese who are not part of their government or their massive tech companies.

Let’s start by looking at three approaches that don’t fully solve the problem:

1. Control through regulation. Some say the GDPR has been successful. In many ways, it’s had important positive effects. But the implementation has also placed a significant brake on innovation in Europe, and some companies are now choosing to avoid or delay coming to the European market due to regulatory complexity. Rules alone can’t fully control or govern technology.

   

2. Specialized “bubbles.” Estonia has a nice visualization of its digital ecosystem and X-Road. Within that niche, they’ve done an amazing job. Through open-source solutions and standardized approaches, Estonia controls a certain limited piece of the technological stack. It works for a small country, but it doesn’t address all dependencies or scale to Europe’s wider political and economic reality.

   

3. A fully end-to-end “Euro stack.” Europe has started talking about building its own entire technology stack, from semiconductors to applications, to reduce dependency. But Europe is starting from far behind, and trying to replicate everything all at once would likely waste huge resources without guaranteeing success. That doesn’t mean we shouldn’t do it at some levels, but we have to be strategic about where we invest.

That brings us to the question of digital or tech sovereignty. We’re clearly not comfortable with an environment dominated by U.S. and Chinese platforms, but focusing on the ideal of full sovereignty can distract us from more achievable goals, namely practical autonomy and empowerment.

Looking at the “stack” of technologies, Europe has a foothold in networking (Nokia, Ericsson), partial entry in chip manufacturing, and maybe some software or IoT presence (like Germany). But if we tried to replicate everything end-to-end, it would be extremely difficult. Another factor is the fast-moving landscape: in past revolutions, most profit eventually shifted from the infrastructure to the application layer, so it’s not obvious that building and owning foundational infrastructure alone will guarantee long-term returns.

That poses a dilemma. Economically, it might make sense to emphasize AI applications rather than replicate entire tech stacks. But if you rely on someone else’s core infrastructure, they can tilt the playing field against you. There’s no easy solution, so let me offer a few practical ideas—opinionated, certainly, but hopefully constructive.

1. Build up competence. Understand how these technologies work on both individual and collective levels. Estonia is pushing universal AI education—an excellent step. Everyone should be experimenting, deploying, and building. But we have to keep pace, as AI tooling changes by the week.

2. Focus more on data integrity and availability than on confidentiality alone. Confidentiality matters, but once data has leaked, you can often continue operating. Losing integrity or availability can be catastrophic. Also, the notion that data itself is “the new oil” is overrated. The quality of data and what you build on top of it often matter more than the raw data. The big value lies in what you do with data, not merely hoarding it.

3. Avoid lock-in. While we might not be able to build and own every piece of the stack, it’s crucial to adopt open standards and open-source solutions wherever possible and be able to switch providers if needed. That’s a big part of practicing digital “sovereignty” in a practical sense.

4. Worry less about regulation as a straitjacket. Regulation like the GDPR is there for good reasons, but don’t become paralyzed. Enforcement often targets either big tech companies, severe security lapses, or blatantly unethical behavior. Everyone else usually gets a small fine or just a warning. So follow the spirit of the law, implement decent security, and then go for it. Don’t let fear of hypothetical compliance issues hold you back from innovating.

5. Aim high with big goals. Marc Andreessen wrote an essay titled “Why AI Will Save the World,” laying out a positive vision of AI’s benefits. You don’t have to agree with every point, but the key is to adopt big, bold challenges. Rather than obsess over the layers of compliance, aim to solve a large-scale problem. In the process of solving it, you’ll naturally figure out data governance, skill development, and vendor management. A concrete mission can drive the solutions you need.

Finally, let’s address a brief question about Ukraine. Ukraine offers a lesson that necessity is the mother of invention. The war forced the Ukrainians to innovate rapidly—especially in military AI and drone technology—faster than perhaps anyone in the world. Now they’re trying to channel that same energy into civilian reconstruction and development. If you’re an innovator with a good idea, they welcome you. Estonia has historically been good at this, but maybe we’ve become too comfortable. We need that drive to see big goals through even when friction and obstacles arise.

So, in conclusion, we won’t resolve the entire digital sovereignty dilemma overnight, but these approaches can help us reach an equilibrium in which we can do what we need to do. In short: Don’t worry—be happy, use AI! Thank you.