The question that unlocks AI adoption in government

Most government artificial intelligence (AI) projects hit the same wall.

The prototype and pilot may test well. But moving to real-world use has become much harder once you ask what happens if the system makes a mistake, and who would be responsible?

For a national statistics office, the question is existential. 

Official data underpins economic policy, business decisions, and public trust. It can become an institutional crisis when AI assistant that presents an incorrect figure as official data. 

When Ukraine's State Statistics Service (Ukrstat) approached us about supporting an AI assistant for public data, we knew this was the problem they had to solve before anything else.

Ukrstat's answer was surprisingly simple. They decided the AI would never generate a number.

Governance begins with saying no

StatGPT converts a plain-language question into a database query, then retrieves the answer directly from the official statistics system. 

The user gets the real figure, with a link to the source — covering more than 2,000 indicators across over 100 statistical series.

The AI helps people find data. It doesn't produce data. The distinction sounds subtle, but it fundamentally changes the governance model.

StatGPT does not fabricate missing information. It does not estimate, interpolate, or perform analysis on behalf of the institution. 

If a user asks an ambiguous question, it returns no result — which is a visible, fixable problem, not a confident wrong answer.

Every number originates in the official statistical database and remains fully traceable.

That architectural decision solved a governance problem.

Instead of asking whether the model could always be trusted, Ukrstat designed a system in which trust did not depend on the model in the first place. 

That clarity is what got the tool approved, built, and deployed — in the middle of a war.

Standards, not servers

It is tempting to view StatGPT as an AI project but it is the visible outcome of a much longer institutional reformthat began years before anyone mentioned AI.

The data underneath StatGPT was already disciplined before the assistant launched. 

Ukrstat had migrated its entire statistical output to SDMX — the international data-exchange standard used by Eurostat, the OECD, the World Bank, and the IMF. 

"Doing something that isn't compliant with SDMX is reinventing the wheel," as Deputy Head Artem Rudko puts it. "We decided to store all statistics in the correct format from the foundation."

For Ukraine, this wasn't purely a technical choice. SDMX compliance is an explicit requirement on the path to EU accession — Eurostat only accepts data in compatible formats. 

When Mariana Kotseva, Eurostat's Director-General, attended the portal launch and called what Ukraine had done "an example for the entire European statistical community," that wasn't diplomatic courtesy. 

It was a recognition that the standards work, not the AI layer.

Most governments get the sequence wrong when they pursue AI interfaces before agreeing on what a clean, comparable dataset even looks like. 

The prerequisite for trustworthy AI isn't AI readiness. It's a narrower, less glamorous thing — can your numbers speak the same format as everyone else's before you let a model speak for them?

Reforming institutions, not buying AI

Ukrstat is a conservative institution. In my experience working with public agencies across Ukraine and the wider region, it is only appropriate. 

Statistical offices are expected to be conservative because consistency and reliability are essential to their credibility. It is also what makes the StatGPT story genuinely instructive.

As Ukrstat head Arsen Makarchuk has explained, the agency studied an approach the IMF had already prototyped for its own statistical data, ran an open procurement, and adapted the result for Ukrainian conditions — adding Ukrainian-language support and the restrictions that keep the assistant within its intended scope. 

That makes Ukrstat the second statistics body in the world, after the IMF itself, to deploy this kind of retrieval-only assistant — which matters, because it means the governance model has been tested across two very different institutional contexts.

What made that leap possible was a partnership between Ukrstat, EPAM, and East Europe Foundation (EEF) through the EGAP Program, supported by Switzerland.

EPAM — one of the largest technology services and software engineering companies in Central and East Europe — won the open procurement not as a software contractor but as a strategic consultant with international GovTech experience. 

EPAM built the Sigma platform that replaced 45 legacy systems with a single unified environment covering the entire statistical lifecycle, but the harder work was what came alongside it: introducing product thinking, agile methodology, and modern data management practices into an institution that had operated on a very different logic for decades.

The EEF, through EGAP Program supported by Switzerland, was to connect these parts. 

In practice, that meant something I hadn't fully anticipated when we started: functioning as a translator between two institutions that speak fundamentally different languages. 

EPAM's teams think in sprints, backlogs, and product iterations. 

Ukrstat thinks in regulations, precedents, and the reputational risk of a single wrong number reaching a policymaker. 

The practical bridge was a hybrid working model — at different stages, between ten and seventy EPAM specialists were embedded alongside Ukrstat teams, sometimes leading a workstream, sometimes acting as mentors, with two-week delivery cycles that let the agency course-correct without the kind of big-bang procurement that usually ends in a system nobody uses.

What I found was that the moments of real friction were about pace. 

A private technology partner naturally wants to move fast and test in practice; a statistics agency cannot afford to treat its public outputs as a beta. 

Finding the right tempo — fast enough to maintain momentum, slow enough that the institution remained in control of its own transformation — was where our programme's presence made the most practical difference. 

We weren't mediating conflict. We were managing the speed at which change was safe to absorb.
But matching Ukrstat's expectations with EPAM's capabilities also meant looking beyond delivery. 

However innovative the solution, it would ultimately belong to the state, with Ukrstat responsible for maintaining and developing it, training staff to work with it, and eventually funding these needs from the public budget. 

That meant EPAM's role could not stop at delivering technology. It also included training, consulting, and knowledge transfer so that Ukrstat could continue working with the system independently. 

For us, that was a critical part of sustainability — building the capacity to own, maintain, and develop it.

Having facilitated partnerships like this across the region, I believethe state has to want the reform for its own reasons, not because a donor or a technology company convinced them it was a good idea. 

What made Ukrstat different was that the pressure to modernise was internal — driven by the need to align with European standards, to survive wartime conditions with aging infrastructure, and to serve a public that increasingly expected digital access to official data. 

We helped channel it into something buildable.

The clearest proof that this worked is a structural one. In December 2022, Ukrstat created the Statistical Data Processing Centre — a new state entity staffed almost entirely from the private sector, with around twenty-five data engineers, analysts, and machine learning specialists working under conditions closer to a tech company than a civil service department. 

The active transformation phase is over; the institution now develops and maintains the platform on its own.

As EPAM put it: "we launched irreversible processes of digital transformation and handed them to a new structure within Ukrstat." 

What made this work was the combination: a state institution reforming from the inside, a private technology partner capable of rebuilding complex systems without outpacing what the institution could absorb, and a program partner whose job was to sustain the process until the institution no longer needed it. 

I believe this three-way model — government, business, and international programme working as genuine partners rather than client and vendors — is the one worth replicating for anyone serious about digital reform in this part of the world.

Three things other governments can take from this

First: define what your AI cannot do before you build it. That constraint is not a limitation — it's the governance argument that gets a project approved. 

A "never generate a number" rule is a one-sentence answer to the question that kills most government AI pilots.

Second: standardise before you build. A well-designed AI layered on locally-formatted, non-standard data is still a weak tool. 

Migrating to SDMX did more for StatGPT's credibility than any model choice did — and it was the part of this project that took years, not the part that made the headlines.

Third: recognise that institutional reform is part of AI implementation. Governments do not become AI-ready by adopting AI. 

They become AI-ready by reforming the institutions, governance, and data that AI depends on.

Building that capability — through the right combination of state ownership, technical partnership, and sustained programme support — is what allows governments to deploy AI responsibly at scale.

The next frontier of digital governance isn't about chasing the flashiest demo. 

StatGPT is not remarkable because it uses AI, but because a traditionally risk-averse public institution trusted it enough to put it in front of every citizen.

That trust was built long before the first prompt was ever entered. And it is entirely replicable.

--------------------------------------

The author is the president of East Europe Foundation (EEF) in Ukraine, which is a non-profit organisation that supports civil society, democratic governance, and community resilience.

Since 2008, EEF has directed $30 million to national and local initiatives, implementing more than 100 programs: From instituting e-government, supporting civil society organisations, to equipping school shelters and producing educational products about landmine safety.