AI in public procurement – the legal and strategic framework

I was invited to deliver a presentation on ‘AI in public procurement – the legal and strategic framework’ at the Nordic Public Procurement Forum 2026 organised by Nohrcon in Copenhagen. In case of interest, my slides are available here.

The more I talk about this topic, the more I sense there is a fork on the road ahead of us.

One way (the steep and hard way), it goes towards properly assured AI use in the public sector, including the use of AI in procurement. The other (the downhill slope leading to a cliff edge) leads to chaotic AI implementation without appropriate guardrails. Right now, there is no map in the legislative and strategic framework and there is a high risk that the seemingly easier path takes to a dangerous situation.

There is an opportunity for the proposed Public Procurement Act to make a difference and clearly not only show the way to the right path but also, and more importantly, to fence off the slope to the cliff. Will we see meaningful change in the process of negotiation of the new EU rules? My suggestions are available here (‘AI's absence from the proposed EU Public Procurement Act – suggested improvements’ , 5 Oct 2026).

Procurement of digital solutions set to become significantly more complex in the EU--time to move away from 'regulation by contract'?

I was invited to deliver a lecture on “Public procurement and European industrial policy: Strategic dependencies, European preference and legal limits” at Nova in Lisbon in a programme led by Filomena Vieira (slides available here). The lecture was in a module on digital procurement, which gave me the opportunity to reflect on recent EU digital law proposals bound to impact procurement of digital solutions. For me, the question was not only how much more complex digital procurement is likely to become, but what (if any) changes in approach to the current ‘regulation by contract’ model of procurement-based digital gatekeeping could arise from the proposals. Short answer: there is not much change other than in relation to (defeasible) centralised requirements by the Commission. Longer answer follows.

Procurement under the EU AI Act as the regulatory benchmark

To date, the EU AI Act has set the path for the ‘regulation by contract’ of digital technologies acquired by the public sector. As I have shown elsewhere, EU law has not directly regulated public sector AI adoption, but rather relied on public procurement as a regulatory tool. The EU AI Act contains a series of requirements for high-risk AI use cases that public buyers are required to operationalise through their tender processes and public contracts.

Setting aside the issue of the uncertainty of what AI solutions are covered by the EU AI Act’s ‘high-risk regime’ (due to the potentially rather narrow definition of AI systems in the Commission’s guidance, and the equally unclear applicability of the ‘filter exemption’ in Art 6(3) to use cases otherwise captured by Annex III, also as a result of the Commission’s guidance), in themselves, the obligations arising from the ‘high-risk’ regime are rather open-ended—which creates significant space for legal uncertainty and difficult practical operationalisation. Despite the endorsement of model contract clauses for high-risk AI procurement by the Commission, there is still no default or fallback set of terms for AI procurement under EU law and each public buyer is meant to set its own requirements by contract. This leaves a significant number of challenges unaddressed—which under-resourced public buyers can struggle to meet.

In my view, this is an unsatisfactory approach to digital regulation and, more generally, evidences the shortcomings of the ‘regulation by contract’ approach. I was thus interested in assessing if the same approach underpinned other recent proposals with an impact on the procurement of digital solutions. Without aiming to be comprehensive, I looked into the Cloud and AI Development Act, Chips Act 2.0, Industrial Accelerator Act and the proposed EU Public Procurement Act. And, within those, I focused on issues of digital sovereignty (broadly construed) and geopolitical management of access to EU procurement markets as a related concern.

Cloud and AI Development Act (CADA)

Of the several proposals, CADA is the most closely linked to the EU AI Act by subject matter. From a procurement perspective, CADA is interesting because it creates a system of tiered assurance that could relieve procurement of part of the digital gatekeeping role it has been assigned. CADA sets four levels of EU assurance. The requirements for each level are cumulative and concern issues such as the need to be established in the EU, the localisation of infrastructure and assets in the EU, data localisation, and aspects of personnel dedicated to the provision of relevant connected services. The elements that are more modular across assurance levels concern cybersecurity requirements, as well as the risk of interference by legal entities established in third countries or third country governments. Notably, while level 1 can be self-assured by the tech provider and would only, at most, require registration with a national competent authority; assurance at levels 2-4 would require independent audit and separate verification by such a national competent authority. Under this model, public buyers seeking to procure cloud services (for exclusive use…) would be able to simply state the level of assurance required and then check against the relevant decisions.

Relatedly, CADA, in principle, requires public buyers that ‘procure cloud computing services for their exclusive use’ to at least procure solutions with a tier 1 of assurance, with the requirement heightened to levels 2-4 for those public buyers whose activities have been identified as contributing to the preservation of public order or which are active in the areas of national security, internal security, external border management, defence, justice or law enforcement, including the prevention, investigation, detection and prosecution of criminal offence. However, in a return to the decision-making decentralisation characteristic of regulation by contract, this is subject to exceptions; public buyers would be able, on an exceptional basis, to decide not to procure cloud computing services recognised as having a Union assurance level 1, 2, 3, or 4 where (i) there is no recognised cloud computing services provider in the relevant EU-wide register and there is no reasonable alternative, (ii) there has been an unsuccessful attempt to procure assured services in the previous year, or (iii) ‘applying the requirements of [CADA] would require the contracting authority to procure services at disproportionate cost.’ CADA does not set a threshold of cost disproportionality for those purposes.

More centrally in the ‘regulation by contract’ approach, CADA contains a second regime of relevance for procurement, which seeks to promote the taking into consideration of Union added value. In that regard, CADA requires that ‘In public procurement procedures for innovative cloud computing services and AI systems, contracting authorities shall include, as part of the quality evaluation of the tender, non-price award criteria that allow them to evaluate the tenderer’s contribution to the development of a European cloud and AI ecosystem.’ The scope of application of this obligation is unclear, as ‘innovative cloud computing services and AI systems’ remain undefined—which seems to leave unspecified the key trigger for the obligation to use non-price award criteria relating to the tenderer’s contribution to the development of a European cloud and AI ecosystem.

Notably, those non-price award criteria must (a) be linked to the subject matter of the contract; (b) not confer unrestricted freedom of choice on the contracting authority; (c) be expressly set out in the procurement documents or in the contract notice; and (d) ancillary and not decisive in the award of the contract. I struggle, in particular, with the last requirement, especially if the issue of Union value is to be used as a tie-breaker or in any other way to differentiate between ‘EU’ and ‘non-EU’ solutions. CADA also provides a list of the considerations contracting authorities should take into account in relation to those non-price criteria, such as the tenderer’s contribution to strengthening the digital technology supply chain in the Union, or whether it has integrated technologies developed in the Union. From a procurement perspective, another difficulty is that most of those considerations relate to tech provider-level, rather than offer level, considerations—which makes their treatment as award criteria legally challenging.

Overall, then, CADA does not move very far away from the ‘regulation by contract’ approach. While the process to assure cloud solutions in pre-defined tiers could be helpful, its effectiveness is conditional on market conditions (availability, relative cost) and, ultimately, it is likely that a significant number of cloud solutions will only be subject to level 1 self-assurance, which does not address the issues of verification of tech providers’ self-declarations. Additionally, in relation to the promotion of Union value added, CADA leaves it all for public buyers to figure out (save for problematic constraints on how to do so).

Chips Act 2.0

The Chips Act 2.0 includes some rules of relevance to procurement concerning infrastructures, equipment or systems in sectors of high criticality and in other critical sectors for the purposes of the NIS 2 Directive (Annexes I and II). In those cases, public buyers may require economic operators to submit, together with their tender documents, a security of supply declaration relating to the sourcing of semiconductors incorporated into the infrastructures, equipment or systems covered by the contract. Importantly, this request is not self-standing; the Chips Act 2.0 foresees that ‘Where a security of supply declaration is required from economic operators, contracting authorities and contracting entities may specify, in the procurement documents, requirements relating to security of supply relating to the sourcing of semiconductors… To that end, contracting authorities and contracting entities may include technical specifications, selection criteria, award criteria or contract performance clauses.’ This is also followed by the same caveats that those requirements must (a) be linked to the subject matter of the contract; (b) not confer unrestricted freedom of choice on the contracting authority; (c) be expressly set out in the procurement documents or in the contract notice; (d) ancillary and not decisive in the award of the contract; and (e) designed in line with the Union’s international commitments. This is merely a reminder of the possibility of ‘regulation by contract’.

Going beyond this approach, the Act empowers the Commission, upon identifying a potential supply chain risk, to issue recommendations to contracting authorities and contracting entities regarding the supply of semiconductors by domestic undertakings for the procurement procedures mentioned above. Where the Commission is not satisfied that those risks are sufficiently being taken into account, the Commission can then adopt an implementing decision specifying for which specific infrastructures, equipment or systems contracting authorities and entities shall require a security of supply declaration from economic operators. While this goes some way to centralising decision-making and the exercise of regulatory power, once again, compliance with such requirements is subject to exceptions where (i) there is no reasonable alternative, (ii) there has been an unsuccessful attempt to procure assured services in the previous two years, or (iii) applying the requirements ‘would require a contracting authority or contracting entity to acquire goods, services or works having disproportionate costs or would result in technical incompatibility in their operation and maintenance’. In this case, the Chips Act 2.0 sets a 25% cost difference as the relevant disproportionality threshold.

The only exception to the possibility of deviating from Commission-issued requirements concerns the risk-prone sectors listed in Annex V or Annex VI. In these sectors, the Commission can issue guidance on how to carry out security of supply risk assessments and recommend possible mitigation measures. If the Commission considers that this does not lead to sufficient risk management, it can then adopt an implementing act on specific risk mitigation measures that sectors or subsectors shall take for specified semiconductor products, namely mitigation measures related to: (a) performance of a risk assessment; (b) procurement of semiconductors; (c) dual sourcing; (d) stockpiling; [or] (e) diversification of supply of semiconductors or related components. There is no exception to the obligation to apply this implementing act (though it is conceivable that such possibility is included in the act itself).

All in all, the Chips Act 2.0 also follows a ‘regulation by contract’ approach to the intended management of risks in semiconductors supply chains. Only in relation to risk-prone sectors can the Commission exercise binding regulatory discretion and, even then, its application is likely to require a fair amount of follow-on technical discretion from public buyers.

Industrial Accelerator Act (IAA)

Focusing solely on the specific requirements for procurement, for specified sectors (steel, concrete and mortar, aluminium, and electric vehicles), the IAA contains two relatively-clear rules. First, public buyers must exclude tenders submitted by economic operators owned or controlled by an entity established in third countries which have not concluded an international agreement with the Union guaranteeing such access. Second, they must apply the Union origin requirements and low-carbon requirements laid down in the IAA.

However, once again, these obligations are excepted where (i) there is no reasonable alternative, (ii) there has been an unsuccessful attempt to procure assured services in the previous two years, or (iii) applying the IAA requirements ‘would require a contracting authority or contracting entity to acquire goods, services or works having disproportionate costs or would result in technical incompatibility in their operation and maintenance’, with the IAA also setting a 25% cost difference as the relevant disproportionality threshold.

The IAA, then, also stays close to the ‘regulation by contract’ approach in relation to the exception from its main obligations, at least in relation to the possibility to deviate from them in relation to their likely cost implications.

A question the IAA, CADA and Chips Act 2.0 do not address is the extent to which contracting authorities can (or should) seek to retain most of their requirements where compliance is subject to an exception (ie are the requirements all-or-nothing). In the absence of guidance, it is also unclear whether, for procurements not covered by their requirements, public buyers should (or can) seek to take inspiration from the (skeletal) legal regimes they develop in creating their own contractualised requirements for levels of assurance falling short of the legal benchmark. In itself, it seems that the proliferation of EU digital laws can lead to the development of templates and standard terms that could overshoot the mark in some procurements. In that regard, the scant guidance accompanying the EU AI Act model clauses for non-high-risk AI procurement could be a harbinger of the vacuum public buyers would (continue to) face where relevant legal instruments are not (fully) applicable.

EU Public Procurement Act (PPA)

A final proposal worth considering concerns the current review of the procurement rules themselves, as they will provide the legal framework underpinning the exercise of ‘regulation by contract’. Although the PPA is a rather complex instrument, a focused analysis on the guiding principles of its treatment of the issue of third country access / European preference, as well as the management of risks to national (cyber)security also show a clear alignment with the ‘regulation by contract’ model.

In outline, the PPA includes an explicit regime of access to EU procurement markets that largely consolidates the complex and fractioned legal regime left by the Kolin and Qingdao line of case law. Generally, covered third-country operators benefit from equal treatment and non-discrimination (Arts 4(2) and 70), subject to an extremely problematic proposed power for the European Commission to restrict their access where there is insufficient reciprocity (Art 72). The PPA allows for the exercise of an ‘EU preference’, which can take the form of requirements for EU content, exclusion of ‘non-EU’ tenders and evaluation adjustments — though, for all intents and purposes, ‘EU content’ extends to that of countries with which the EU has international obligations (Art 73).

Non-covered third-country operators are left to the discretionary approach by contracting authorities, but cannot benefit from equal treatment (Art 4(2) a contrario + Kolin/Qingdao as saved by Recital (33)), subject to power of the European Commission to close access or impose ‘EU preference’ mechanisms ‘where this is in the interest of the Union’ (Art 75). However, this is subject to exceptions: (i) emergency procurement, (ii) unavailability, (iii) unsuccessful procedure in prior two years, (iv) disproportionate costs — with no stated benchmark (Art 76), as well as subject to special ‘European preference’ rules in sectoral legislation (art 77). In short, this is a model of ‘regulation by contract’, even where the Commission may seek to impose centralised European preference requirements.

Differently, the PPA contains an expanded scope for rules based on national security concerns (Art 66). These include: (a) a general duty to to ensure the protection of the security and public safety interests of the Union or one or more Member States for any public procurement procedure identified as presenting or including a risk for security or public safety; (b) an indication of sources of risks that covers both issues arising from the subject-matter of the contract and from economic operators (which opens the door to geopolitical considerations); (c) the mandatory exclusion of operators identified as high-risk suppliers pursuant to the proposed Cybersecurity Act 2; and (d) an empowerment to the Commission for the adoption of delegated acts for specific categories of goods, services or works where such elements address an identified specific security and public safety interest of the Union. In this case, it seems that the centrally-established requirements would be entirely mandatory for public buyers, without possibility to except them.

Still on cybersecurity matters (Art 68), the rest of the PPA is somewhat less centralised, including rules on the: (a) mandatory screening of compliance with the requirements of the Cyber Resilience Act; (b) possibility for additional requirements where digital elements are procured or used for national security or defence purposes; and (c) possibility to impose other cybersecurity requirements, without prejudice of the NIS 2 Directive. The last two measures are clearly just an empowerment and would still leave it to public buyers to set the relevant cybersecurity requirements as ‘regulation by contract’.

Overall, especially in relation to the third country access regime and cybersecurity issues, the PPA remains closely aligned to the ‘regulation by contract’ model (and, on a quick reading, the same applies to most of the core issues the PPA seeks to address in relation to strategic procurement more widely). This shows limited potential for a change of model where the PPA contains rules that reflect elements of digital regulation.

Final reflection

This overview of some of the recent proposals on EU digital law with an impact on procurement, as well as the proposed new procurement regime, shows that the model remains one of ‘regulation by contract’ where public buyers will be expected to manage an increasingly diverse and ever more complex array of technical, socio-technical, geopolitical, economic and other risks and considerations. In my view, this is not the best way forward. Overburdening procurement as not only a tool for the enforcement of horizontal/sectoral regulation, but a tool to formulate such ‘regulation by contract’ can only (continue to) generate difficulties and lead to growing ineffectiveness of a function that is key to the functioning of the public sector. There is a need to rethink the regulatory architecture underpinning the (formulation and) delivery of the EU’s policy goals. If procurement is the only (or the primary) regulatory infrastructure being engaged, the system will be ineffective and it is likely that it will collapse.

Initial assessment of proposed EU Public Procurement Act

I have submitted feedback to the public consultation on the proposed EU Public Procurement Act (currently open until 24 December). My submission (F33581451) provides an initial high-level assessment of the proposal for a Public Procurement Act. It focuses on selected issues.

The first part focuses on the choice of legal instrument and proposes a split approach to the legal form underpinning the process of review: a short regulation on digital ecosystem, a transfer of the third country access rules to the revision of the International Procurement Instrument, and a resubmission of a proposal for a new directive, subject to there not being enough support for an alternative approach not to revise the procurement rules and instead make significant investments in boosting procurement capacity.

The second part engages in some more detail with the content of the proposed Public Procurement Act, flagging selected issues that would require reconsideration in the legislative process leading to new EU procurement rules, regardless of the choice of legal instrument. The submission is by no means comprehensive and has focused on those issues that are either most problematic, or that would require significant development.

The full text of the submission is available here. Comments most welcome, as always.

AI's absence from the proposed EU Public Procurement Act -- suggested improvements

If we live in the age of artificial intelligence, the proposed EU Public Procurement Act is anachronistic in its omission of AI. AI simply does not feature in the Public Procurement Act — or at best, only as an after-thought or ‘something to think about at a later time’. Given the continuous expectation that procurement will operate as a key governance tool for public sector AI use, that is not good enough.

As is now trite, there are two major interactions between AI and procurement: procurement is the gateway for AI acquisition by the public sector (‘AI procurement’); and AI can be (and is increasingly) used to manage procurement (‘AI for procurement’). Here, we can distinguish the direct use of AI for procurement and the use of AI to monitor procurement (eg to tackle corruption or collusion risks). The latter could be thought about as a meta application of AI for procurement.

Despite the growing understanding of these two/three layers of issues and the emergence of a quickly-expanding academic literature, as well as policy and guidance literature from across a wide variety of jurisdictions, the Public Procurement Act does not include meaningful consideration of AI.

Missing priority

To be fair, other than in relation to some meta applications of AI for procurement, AI also did not feature in the key documents influencing the Commission’s approach to the proposal. The report by the European Court of Auditors and the Council’s conclusions that kick-started the review solely referred to AI in passing (in relation to automated translation by ECA, and about exploitation of public procurement data by the the Council). The Parliament’s resolution on procurement similarly solely stressed that the transition to a digital-first approach would “reinforce the possibility of public scrutiny in the procurement process, especially considering the potential offered by disruptive technologies such as artificial intelligence” and called the Commission to update its tools for monitoring competition in public procurement. As far as I know, only the opinion of the European Committee of the Regions has gone beyond consideration of these meta uses of AI for procurement, and stressed that “it is necessary to provide a specific legal regime for the use of AI in the area of public procurement, thereby guaranteeing a framework of legal certainty” (para 38).

Similar considerations were captured in the public consultation underpinning the proposal. As the Commission’s impact assessment reflects, “several respondents see AI and advanced analytics as a way to improve procurement governance, through automated checks, risk scoring, fraud detection, and carbon-footprint verification. At the same time, they call for clear rules on the use of AI and algorithmic tools in award processes, to maintain transparency, fairness and legal certainty”.

The Public Procurement Act’s approach

This has not translated into any provision in the proposed Public Procurement Act, not even in relation to the (meta) use of AI for procurement. All there is, in the Legislative financial and digital statement (part 4.3), is a consideration that several elements of the envisaged new digital architecture (such as the Commission’s eProcurement platform, the electronic eligibility service, or the procurement data spaces) could interact with the EU AI Act. However, there is a repeated statement that “No AI-based decision-making is foreseen for the core legal obligations. If automated risk detection, ranking, recommendation or fraud analytics are introduced, an AI Act risk/compliance assessment will be carried out.” So, if anything, the proposal suggests that the digital architecture will not include ways to embed AI for the purposes of any of the key aspects of a procurement procedure (ie ‘core elements’). Is this a tenable position? Is this a ‘future-proof’ approach to regulating procurement in the EU, especially on the basis of a single rulebook? What would happen when Member States start (or continue) deploying AI in relation to ‘core obligations’? Would that be permitted? Under which conditions? How would divergence in the (applicable by default) administrative law of the Member States be negotiated?

Regulatory assumptions?

More broadly, it seems that, at best, the Commission operates on the assumption that the EU AI Act would provide the regulatory architecture for the use of AI in procurement (if there was any), but this is only explicit in relation to the meta use of AI for procurement and pending some future impact assessments (not least, because there is no detail whatsoever on what those components will look like in the many delegated and implementing acts required to develop them). Even then, the use of AI for procurement is not really considered in full, such as eg the Public Procurement Act’s omission of the more immediate potential obligation to disclose human-AI interaction under the transparency requirements of Art 50(1) EU AI Act even where no AI-based decision-making takes place—as soon as tenderers or public buyers eg interact with a chatbot.

The procurement of AI is, in itself, not even referred to in the proposal. Perhaps this is based on the Commission’s view that its (meagre) work on this issue to date suffices.

However, this ‘silent approach’ to reflecting the AI/procurement interactions in the Public Procurement Act ignores the fact that, other than transparency, the EU AI Act does not create obligations for the use of AI in procurement and, at best, provides a very weak level of regulatory assurance for AI procurement, and that the model clauses for AI procurement are nothing but an empty shell.

taking (more meaningful) action

All in all, the Public Procurement Act does not contain any provisions on AI procurement, does not fully consider the use of AI for procurement, and only hints at the future consideration of meta uses of AI for procurement monitoring purposes. This is at odds with the pressure for procurement to regulate AI acquisition by the public sector, as well as the pressure to deploy AI to manage procurement processes.

In my view, there are at least three improvements the co-legislators should introduce in the legislative process the Public Procurement Act is undergoing.

First, the Public Procurement Act should reform the EU AI Act and, in particular, Annex III to include the use of AI for public procurement (both directly, and in relation to its meta monitoring) as a ‘high-risk’ AI use case. This would bring procurement within the scope of the EU AI Act, but would still be largely ineffectual due to the exemptions in Article 6(3) EU AI Act. Therefore, the Public Procurement Act should also specify that Article 6(3) EU AI Act is inapplicable to the use of AI for procurement. This would go a long way in addressing the governance of AI use by public buyers.

As a complement—or a rather weak substitute, if modifying the EU AI Act in this way is not (politically) feasible—the Public Procurement Act could include an explicit provision on AI use for procurement to make public buyers directly responsible for the consequences of AI use for procurement, regardless of any responsibility or liability that can be sought from the providers of the used AI solutions—and this could be coupled by an obligation to develop, publish and constantly update a policy on the use of AI for procurement. This should include specific approaches to testing for bias and anticompetitive effects potentially arising from AI use (for discussion, see here).

Second, to mirror this, the use of AI by economic operators participating in procurement procedures would also need regulating. At a minimum, this should be reflected in the rules on exclusion (which need a significant overhaul more generally, in particular to make all exclusion grounds mandatory if the instrument is to remain as a Regulation) and an explicit exclusion ground should cover the misuse of AI in the participation in procurement procedures or in the execution of public contracts.

Finally, AI procurement needs to be given legislative coverage. As there is no chance that a full regime can be included in an Act, and much less that this would be sufficiently dynamic and adaptable, the Public Procurement Act should include a new article mandating the Commission and the Member States, jointly, to develop and constantly update a set of mandatory (not model) clauses for AI procurement. These would need to be a lot more detailed and implementable than the current clauses, and it seems to me that the future clauses need to be jurisdiction-specific (not merely translated into relevant languages) for them to adequately bridge the gap between public and private law rules (including liability) and jurisdictional variations. It would also be important to consider comparative approaches and to reflect on and learn from, eg recent developments in the US with the adoption by the General Services Administration (GSA) of a new clause for incorporation into government contracts of artificial intelligence systems (552.239-7001, Basic Safeguarding of Data within Large Language Model Artificial Intelligence Systems).

These interventions would still need refining and careful drafting and, most likely, would probably still be insufficient to fully assure AI procurement and the use of AI for procurement and its oversight, but they would provide a firm statutory grounding on which to continue the regulatory conversation.