Does AI have to be procured, and what actually decides it in Sweden

In Sweden, the commonest objection to an AI pilot in the public sector is that it has to be procured first. The objection is rarely put as a question and almost never with a figure. It works as a stop, and it seldom survives a closer look.

The act lists six procedures, and none of them is AI

The Swedish Public Procurement Act sets out exhaustively, in chapter 6 section 1, the procedures that may be used: the open procedure and the restricted procedure, the competitive procedure with negotiation and the negotiated procedure without prior publication, competitive dialogue and innovation partnership. None of them is tied to a technology. What is required is decided instead by the value, which under chapter 5 section 3 is calculated on the total amount payable under the contract. Buying AI support is therefore treated like buying other goods and services, and is decided by what is bought and what it is worth.

That is worth saying plainly, because the opposite is assumed so often. The question "do we have to procure AI" has no answer of its own. The question that does have an answer is what the contract is worth and what it covers.

What decides is the value, not the technology

The direct award threshold is the point below which a purchase can be made without advertising. According to the Swedish National Agency for Public Procurement, it sits at SEK 700,000 under the Public Procurement Act, and has done since 1 February 2022. Under the act governing the utilities sectors and the act governing defence and security procurement it is SEK 1,200,000. It is the EU thresholds that are recalculated every two years, not the direct award limits: those are stated as amounts in the act itself and change only through an amendment to it.

The figure comes with two qualifications. The agency is explicit that the threshold is not an annual value: it is calculated as the total amount the contracting organisation estimates will be paid over the term of the agreement, including the value of any option or extension clauses. And a procurement may not be split in order to fall below a threshold. Neither qualification changes the main point. A bounded trial in one part of an organisation often sits below the threshold, and then advertising is not the question.

What stops most pilots is not procurement law. It is that nobody has done the arithmetic.

While the contract value is unknown it is impossible to determine which procedure applies, and a question that is really arithmetic sits there as legal uncertainty. An estimate across the full term is a bounded piece of arithmetic, and it dissolves the uncertainty.

The existing agreement rarely covers it

The second common assumption is the opposite of the first: that nothing needs to be bought, because the capability is already included in a licence the organisation pays for.

Digg, the Swedish Agency for Digital Government, states plainly that it is not possible to obtain generative AI within existing supplier agreements for the office suite. Digg also notes that it remains difficult to rely on larger public sector framework agreements for generative AI. Both pages are in Swedish.

It is worth noticing how the two assumptions fit together. One organisation can hold both at once: that a new tool requires a full procurement, and that the tool already in use requires nothing at all because it arrived with the package. The first assumption stops what has been thought through. The second lets through what has not.

What the rules actually add

The EU AI Act regulates the system and the roles around it, not the procurement procedure. It adds obligations for whoever puts a system into use, and more extensive ones where the system is classified as high risk. It creates no new procurement procedure and no new monetary threshold. The requirements therefore land in the specification, within the procedure that applied anyway.

Obligations on the system do not disappear because the purchase is small, and they do not arise because the purchase is large.

Two supports exist for anyone who would rather not draft every clause from scratch. Digg and IMY, the Swedish Authority for Privacy Protection, have jointly produced guidelines on generative AI in public administration, last updated in October 2025, covering adoption, ethical use, copyright and procurement. The European Commission's Public Buyers Community also publishes model contractual clauses for the procurement of AI, in a full version and a lighter version with accompanying guidance. Both are guidance and model terms rather than legislation, and neither replaces an organisation's own requirements.

The assessment in any given case turns on the history of the contract and on the framework agreements the organisation is party to.

Three questions before you start

1. What is the contract worth across its full term? Estimate the whole agreement rather than the pilot month, and include the value of option and extension clauses. Without that figure none of the questions below can be answered.
2. What are you buying? A licence, a consulting engagement, a developed function, or a service where a supplier processes your data. Those four look different in a contract, and they look different in the roles the AI Act sets out.
3. What are you already using? Whatever is already in the organisation is part of the picture, however it got there. An inventory before the procurement is cheaper than one after.

What is worth requiring in the contract

1. Where the data is processed and by whom. Including subprocessors, and including what applies during support and troubleshooting.
2. Whether input is used to train the supplier's models. The question should be answered yes or no in the contract, not in a product brochure.
3. What documentation and logging come with it. You need them to discharge your own obligations, and they are hard to negotiate afterwards.
4. What happens to the data when the agreement ends. Deletion, return and time limits, written so that they can be followed up.

None of these is specific to AI. That is precisely the point. What this takes is the same procurement competence as a cloud service, and anyone waiting for a special AI procedure is waiting for something that is not coming.