When the State Buys Artificial Intelligence

Most contracting authorities approach AI as just another information system. That is understandable. Public procurement law does not distinguish between them, and the tender templates available to authorities were written for supplies of hardware, licences and custom software.

An AI system, however, behaves according to the data it was trained on, and it changes over time. A model that performed excellently in acceptance testing may work noticeably worse a year later because the input data, the legislation or the behaviour of the people it affects has changed. If the contract does not account for this, the contracting authority is usually the last to find out, often only from a citizen’s complaint.

This raises a whole set of questions for which a conventional tender has no box to tick. Who is responsible for the quality of the data the authority itself provides to the supplier? May the supplier use that data to improve its product for other customers? How will the institution explain to a citizen a decision shaped by a model it cannot see into? And what happens five years later, when it wants to re-tender the system and discovers that the model fine-tuned on its own data is effectively in the hands of the original supplier?

These are not academic concerns. With AI, dependence on a supplier builds up faster and runs deeper than with ordinary software, because the system’s value grows with every month of operation on a specific institution’s data. What is not captured in the tender usually cannot be fixed later without the state paying more.

At the international level, these questions have been receiving attention for longer than is generally assumed in Slovakia. As early as 2015, the OECD adopted its Recommendation on Public Procurement, built on twelve principles ranging from transparency and integrity to the professionalisation of procurers and risk management. In May 2019 it was joined by the Recommendation on Artificial Intelligence, according to the OECD the first intergovernmental standard in this field, updated in May 2024; today it has 47 adherents, including the European Union. OECD recommendations are not legally binding, but Slovakia, as a member state, has adopted them and made a political commitment to them.

The OECD’s analysis Governing with Artificial Intelligence from June 2025 maps around two hundred cases of AI use in government and, in the area of procurement, identifies obstacles that are all too familiar from Slovak practice: outdated legal frameworks, fragmented and hard-to-access data, a shortage of skills in the civil service, under-investment in infrastructure, the risk of bias, and the tension between transparency and fair competition. The same analysis also shows the value AI can bring to procurement itself. According to the OECD, Brazil’s audit system Alice cut audit times from 400 days to 8, and Ukraine’s ProZorro system uses AI to classify contracts and detect risks.

In its Apply AI Strategy of October 2025, the European Commission identified public procurement as one of the main instruments for bringing artificial intelligence into public administration. It also sees procurement as the place where it must be ensured that the solutions purchased are functional, ethically defensible, compliant and interoperable from the outset. The strategy openly acknowledges that procurers need to strengthen their own knowledge of AI and that Europe must reduce its dependence on non-European suppliers.

The most practical European output is the set of model contractual clauses for the procurement of AI, known as MCC-AI. They were prepared by the community of practice on the procurement of AI within the Public Buyers Community platform managed by the European Commission. The first version was published on 29 September 2023 and an updated version on 5 March 2025. They come in two variants: one for high-risk systems and a lighter version for other systems that may affect health, safety or fundamental rights. They cover data governance, transparency, human oversight, the contracting authority’s right to audit, and indemnification for infringements of intellectual property and data protection rights.

The potential of public procurement law itself also remains under-used. In its 2021 guidance on innovation procurement, the Commission discusses in detail preliminary market consultations, competitive dialogue, innovation partnerships and pre-commercial procurement. Slovak law provides for these instruments. Yet contracting authorities rarely use them, often for fear of audits, and prefer an open procedure with a precisely specified solution that may already be outdated by the time the contract is signed.

Finally, the experience of countries that started earlier is instructive. The UK Office for AI published its own guidelines for AI procurement as early as 8 June 2020, developed in cooperation with the World Economic Forum and the government’s digital service and central procurement bodies. Public Services and Procurement Canada maintains a list of pre-qualified AI suppliers (the AI Source List), which shortens public institutions’ path to proven solutions. Amsterdam was among the first cities to develop its own contractual standards for algorithmic systems, and its experience has fed into today’s European templates. What all these examples have in common is that they did not begin with a new law. They began with standardised documents, a methodology and investment in the people who work with them.

If I had to sum up where AI procurement most often breaks down, it would not be violations of public procurement law. Audits can catch those. It is decisions that are formally flawless but substantively wrong.

The first is a tender that describes a technology instead of a problem. “Supply of a chatbot using a large language model” is a subject matter that is easy to describe but makes it impossible to measure whether anything has improved. Without a defined outcome, there is nothing to evaluate bids against, and a year later nothing to show whether the money spent delivered any benefit.

The second is data. The authority often does not know exactly what data it has, in what quality and under what conditions it may share it. This uncertainty then carries over into a contract that is either silent on data or contains general wording that leaves unclear who owns the outputs and what happens to the data when the contract ends.

The third is price. Award criteria under which the lowest bid effectively decides the winner are particularly treacherous for AI. The cheapest bid is often cheapest precisely because it does not account for monitoring, documentation, explainability and support during operation, which is what will later prove most expensive.

The fourth is a contract that ends with delivery. For a system that changes over time, acceptance is only the beginning. Without agreed performance monitoring, an incident procedure and a right to look into how the model works, the institution has almost no leverage when problems arise.

And finally, the missing exit. Portability of data and the model, open interfaces and cooperation during handover to a new supplier are topics no one wants to discuss when choosing the winner. Five years later, however, they will decide whether the next tender is a genuine competition.

None of these mistakes is made when the bids are opened. All of them are made during preparation, at the stage when they are cheapest to eliminate.

Consider a model scenario that does not describe any specific case but is assembled from elements that recur in various forms. A state institution procures a system that helps sort citizens’ applications and proposes their priority. The tender goes through without objections, the lowest price wins, acceptance tests go well and the project is presented at a press conference as a digitalisation success.

A year later, the legislation changes, and with it the structure of the applications. The model, which no one systematically monitors, begins to underrate certain types of applications. The first complaints arrive months later. The institution wants to know why the system decided as it did, but the contract gives it no right to look into the model or its operating logs. The supplier offers a fix, for a fee of course, by way of an amendment.

Three years on, as the contract nears its end, the institution discovers that it cannot transfer the model fine-tuned on its own data to another supplier and that a new tender would mean starting from scratch. The original saving from the lowest price has long since been spent several times over on amendments and staff time, not to mention citizens’ trust, which no spreadsheet can quantify.

The OECD and the European Commission agree entirely on one point: the biggest obstacle to good AI procurement is neither technology nor law, but capacity. A procurement officer who prepares a tender for office supplies one week and for an artificial intelligence system the next cannot be an expert in everything. Nor should anyone expect that.

In AI procurement, a single tender brings together questions of public procurement, contract law, personal data protection, cybersecurity, data governance, fundamental rights and the institution’s own substantive agenda. Each of these areas has its own people in the civil service, but they rarely sit at the same table while the tender is taking shape. More often they comment on it one after another, in rounds of comments, when the key decisions have already been made.

This, in my view, is where an institution’s leadership has the greatest influence. Not by writing technical specifications itself. What matters is whether it treats AI procurement as an administrative task for the procurement department or as a strategic decision that deserves time, the right people and an independent outside perspective.

On 9 September 2026, the European Commission presented its proposal for a European Public Procurement Act. The three existing directives on public contracts and concessions are to be replaced by a single directly applicable regulation. For technology procurement, it brings one ground-breaking change: quality criteria are to account for at least 30% of the bid evaluation, and 50% for labour-intensive contracts, with a lower weighting allowed only where quality is already secured through the specifications or contract conditions. The proposal also strengthens strategic procurement, including a European preference, and provides for public procurement data spaces. The legislative process will take time; by current estimates, adoption is not expected before the end of 2027, and the regulation is to apply roughly two years after adoption. The direction, however, is clear.

Milestones of European AI regulation fall within the same period. Transparency obligations under the AI Act have applied since August 2026, and following the adoption of the Digital Omnibus package, obligations for stand-alone high-risk systems have been postponed to 2 December 2027. In Slovakia, market surveillance of AI is to be carried out by the Ministry of Investments, Regional Development and Informatization from 1 January 2027.

For contracting authorities, this adds up to simple arithmetic. Preparing and running a tender for a non-trivial IT solution typically takes a year, implementation another year, and the operating contract is concluded for three to five years. A system whose procurement starts being prepared today will spend most of its life in an environment where both the new procurement rules and the full obligations under the AI Act apply.

“Draft today’s tender to yesterday’s rules, and you are drafting tomorrow’s amendments.”

One thing I want to stress here. Regulation sets the minimum. An institution that follows the recommendations of the OECD and the European Commission will meet most regulatory requirements as a natural consequence of good preparation. It does not work the other way round. Formal compliance with the rules has never yet guaranteed anyone a good purchase.

Whenever an institution plans to procure a solution involving artificial intelligence, I recommend that its leadership ask itself a few questions before approving the tender documents. They are not the only ones that need answering, but they reliably show whether the project is ready:

If an institution cannot answer any one of them in a single sentence, that is no reason to stop the project. It is a signal that the tender is not yet ready to be published.