
Published
Author
AI is now rapidly being integrated into the IT systems already used by the public sector. But when AI capabilities are priced through consumption models, tokens, and credits, a fundamental procurement law challenge arises: how do you procure something whose scope, cost, and future use are difficult to describe transparently?
The principles behind public procurement are based on transparency, equal treatment and predictability. A contracting organisation must be able to describe what is being purchased, what requirements apply and how the bids will be evaluated. At the same time, guidance on procured contracts states that the general rule is that contracts and framework agreements may not be changed without new procurement, unless the change falls within special exceptions or clear change and option clauses in the original procurement documents. Change clauses must be clear, precise and unambiguous and state the scope, nature and conditions of possible changes.
What makes AI particularly challenging is that the technology is increasingly integrated into existing operational systems. Vendors of financial systems, ERP systems, CRM systems, case management systems, HR systems, and other platforms are developing AI capabilities as a natural next step in their products. For the customer, this may appear as a logical further development of an already procured system. From a procurement law perspective, the situation is more complicated.
If the AI functionality was not included within the original procurement's scope, it cannot simply be added as a new cost-driving component. This is especially true when the functionality changes the system's utility, working methods, or financial conditions. The public sector can thus find itself in a paradox: the organisation wants to benefit from the innovation in the platform it already uses, but must at the same time ensure that every new call-off complies with LOU.
The greatest uncertainty often does not lie in the AI function itself, but in the pricing model. Many AI services are based on consumption of tokens, credits or similar units. Even if the customer receives a price per token, it is rarely clear how many tokens are consumed for different types of queries, documents, workflows or automations. Furthermore, the vendor can change the models' capacities, underlying computational logic and packaging over time.
This means that the customer may know the unit price but still lack the ability to calculate the actual cost. Such a model is difficult to reconcile with the requirement for transparency in a procurement. To be able to compare bids, not only price lists are required, but also an understandable connection between use cases, volume, capacity, limitations and total cost over time.
Many AI services are based on the consumption of tokens, credits, or similar units. Although the customer receives a price per token, it is rarely clear how many tokens are consumed for different types of queries, documents, workflows, or automations.
The Public Procurement Agency describes how requirements can be formulated as detailed requirements, functional requirements or through standards, and that functional requirements should specify what result is to be achieved. For AI this is easier said than done. Ideally, the customer should describe what AI capabilities are needed, what quality is required, what risks should be managed and how the benefits are to be measured. At the same time, the technology develops so quickly that today's requirements risk being outdated already when the contract starts to be used.
The consequence is that the evaluation model risks resting on assumptions rather than verifiable facts. One vendor may offer a low initial cost but high consumption risk. Another may offer a higher fixed cost but greater predictability. A third may include AI in the license package but limit capacity in a way that only becomes apparent in practical use. For the procuring organisation, it becomes difficult to determine which tender actually provides the best value.
The challenge is not limited to the price. DIGG recommends that generative AI is purchased based on the needs of the operation and the rules for public procurement, and that organisations consider licence and contract terms before acquisition and during use. At the same time, the AI Regulation applies alongside other regulations. IMY describes the AI Regulation as a risk-based regulatory framework where certain AI systems must meet requirements such as risk assessment, transparency, and documentation before they are put into use. For the public sector, this means that procurement must create conditions not only for purchase but also for lawful and responsible use.
Public organisations facing system procurements therefore need to map out how AI may be used during the contract period already in the preparation phase. It is not enough to ask if the system has AI support. The organisation needs to understand which AI capabilities can be activated, how they are priced, which volumes drive costs, which contract terms can be changed, and what obligations arise from the use.
The public sector should strive for pricing models where the vendor commits to clear and comparable parameters over a longer period. This may involve price caps, defined use cases, volume ramps, indexing models, transparent calculation bases, and an obligation to notify changes in AI models or consumption logic. Without such mechanisms, it becomes difficult to create the predictability that LOU requires.
1. Which AI capabilities are included in the original system delivery and which require separate call-off?
2. How is the consumption unit defined, and can the vendor demonstrate how it is consumed in typical use cases?
3. What cost ceilings, volume levels, and price adjustment mechanisms can be agreed for the entire contract period?
4. What changes to AI models, features, or packaging can the vendor make without a new agreement?
5. What documentation is required to meet requirements for transparency, data protection, information security, and the risk-based obligations of the AI regulation?
AI will become a natural part of future ERP systems. But for the public sector to benefit from the technology, procurement models that are as transparent as they are flexible are required. Today, this is often far from reality. Therefore, every public organisation facing a major system procurement should analyse how AI should be handled in requirements specification, evaluation, contract, and pricing model. Otherwise, AI risks becoming the most attractive but least predictable part of the public IT deal.
Are you facing a system procurement? Have a dialogue with us about how you can prepare the procurement with clearer needs, requirements and decision support that takes AI into account.