Clinical performance is only part of the investment decision. European health services need to understand how AI changes tests, staffing, infrastructure and recurring expenditure before converting a promising tool into a funded service.
An AI product can save a clinician time without releasing money from a hospital budget. It can reduce one type of investigation while adding another. It can even be good value for patients while increasing annual expenditure. These are different economic outcomes, and confusing them makes a technology harder to commission sustainably.
For European health systems, the central procurement question is therefore what changes in the care pathway after the software is introduced. The licence price matters, but so do the people who review results, the appointments that follow and the existing services that remain in place.
A useful business case connects the clinical decision to those consequences. It also names the organisation that pays and the organisation that benefits. Otherwise, the financial case can depend on savings that are real somewhere in the system but unavailable to the department expected to fund the technology.
A cardiac example separates a model from a trial
HeartFlow’s coronary analysis provides a well-documented illustration. The company describes a process combining deep learning, trained analysts and computational fluid dynamics to estimate blood flow from coronary CT images. This is an AI-supported analytical service with human quality checks, rather than a standalone algorithm that directly examines a patient.
NICE’s guidance, updated for costs in 2021, estimated potential savings of £391 per patient compared with the modelled diagnostic pathway, assuming access to appropriate CT facilities. This was a modelled estimate using specified costs and assumptions, not a measured saving guaranteed for every hospital.
The FORECAST randomised trial tested a strategy of CT coronary angiography with selective fractional flow reserve analysis against standard care in 1,400 patients with stable chest pain. At nine months, mean cardiac costs were £1,605 in the experimental group and £1,491 with standard care. The difference was not statistically significant. Invasive angiography was reduced, but the trial did not demonstrate lower overall costs.
These findings should not be treated as interchangeable estimates or as a simple contradiction. They involve particular comparators, testing strategies, prices and follow-up. For a buyer, the practical message is to reconstruct the local pathway before importing a headline saving.
A 2025 observational study adds evidence from national implementation across 27 English hospitals and 90,553 patients undergoing coronary CT. Availability of fractional flow reserve analysis was associated with fewer subsequent invasive angiograms and noninvasive cardiac tests over two years. However, when the analysis itself was counted as a diagnostic test, the total number of diagnostics was higher. This was a comparison across implementation periods, rather than randomised allocation. It shows why counting all tests equally obscures changes in resource use: analysing an existing scan and arranging another patient visit impose different demands.
Capacity and cash require different plans
NICE’s evidence standards framework asks digital health suppliers to present budget impact analysis relevant to the intended setting. Its user guide distinguishes monetised effects on cash from effects on capacity, and indicates that technologies with higher financial risk may need further economic analysis. Meeting the framework does not itself constitute NICE endorsement.
Consider a hypothetical tool that saves several minutes of review time per case. Multiplying those minutes by a salary rate estimates the value of capacity released. It does not establish that payroll expenditure will fall. The same staff may instead review more cases, spend longer with complex patients or avoid work extending beyond the planned day.
Those can be worthwhile gains. A financial proposal should identify which outcome is intended and how the service will realise it. Additional throughput requires enough capacity in the next stage of care; reduced overtime requires a change in overtime actually worked. Labelling both as cash savings would hide the operational decisions needed.
The analysis should also follow additional findings. If an AI tool prompts more investigations, the costs of those investigations belong in the evaluation alongside the possible benefit of detecting disease earlier. A higher detection rate alone cannot determine whether the resulting pathway is affordable.
The commercial model determines what the buyer receives
AI can be purchased as software, as part of a wider platform or within a clinical service. A 2024 NHS Commercial Solutions framework award included both outsourced and onsite radiology reporting, with AI-assisted and traditional reporting within the relevant lots. The notice documents a procurement route and supplier appointments; it does not establish that every reporting service uses AI or delivers savings.
This distinction changes the responsibilities being purchased. A software licence supplies a capability that the hospital must incorporate into its own service. A reporting contract purchases an agreed output with associated staffing and service arrangements. The buyer should compare the complete offers rather than treating an algorithm fee and a clinical reporting fee as equivalent prices.
Payment units also distribute risk differently. A fee per processed case links expenditure to activity. A fixed subscription creates a predictable commitment but a higher effective cost per case if use is low. An organisation considering either structure can calculate the volume at which one becomes cheaper, while checking whether support, integration and upgrades are included.
None of these models is intrinsically best. The appropriate choice depends on demand, the ability to manage implementation and how confidently the intended service can use the purchased capacity.
Assurance is part of delivery but does not establish value
In England, the Digital Technology Assessment Criteria cover clinical safety, data protection, technical security, interoperability, and usability and accessibility. NHS England revised the approach in February 2026, including a shorter form. Its guidance says DTAC works alongside other required approvals and does not replace them.
That process addresses whether a digital product can be introduced with appropriate assurance. It does not answer the separate question of whether the product offers good value in a specific service. Procurement teams need both sets of information, without asking a security assessment to stand in for an economic evaluation.
The delivery budget should assign responsibility for the interfaces, local configuration, staff familiarisation and ongoing support that make the technology usable. A low software price can be misleading if substantial work is excluded from the offer. Conversely, a higher bundled price may cover services that would otherwise have to be purchased separately.
The useful comparison is a common service specification with an explicit implementation period and recurrent costs. Suppliers then compete on the same operational requirement, and the health service can see which costs remain on its own side of the contract.
Contracts need to remain useful when the technology changes
The European Commission-supported Public Buyers Community published updated model contractual clauses for AI procurement in March 2025. The materials include versions for high-risk and other AI systems, with guidance on adapting them. They are procurement tools, not a certification that a product complies with every applicable rule.
The accompanying commentary makes their limits explicit: the clauses are not a complete contract and leave matters such as payment, delivery deadlines and liability to the wider agreement. It also discusses dataset rights and the potential to reduce dependence on a supplier.
For healthcare buyers, this suggests a practical negotiating agenda. Agree how product changes are communicated, what evidence accompanies a consequential update and who pays for additional integration work. Establish how performance information can be examined during the contract, and what records and usable outputs remain available when it ends.
A contract that is easy to start but expensive to leave can restrict future competition. Exit arrangements therefore belong in the initial commercial assessment, alongside the price and expected benefits. Buyers can ask suppliers to price transition support explicitly and describe the format in which records would be returned, making future switching costs more visible.
A funded service needs a decision after the pilot
A pilot should produce a decision that the budget holder can act on. Before it starts, the parties can specify the target population, comparator, activity assumptions, implementation costs and period over which effects will be assessed. They can also distinguish improvements that justify continued spending from those that would merely justify another experiment.
The subsequent decision may be to expand, narrow the intended use, change the price or stop. Each can be rational when based on the observed pathway and an affordable service plan. Continuing because installation has already consumed effort is not an economic justification.
For Europe’s healthcare AI industry, a credible commercial case therefore goes beyond demonstrating that software works. It explains what the customer must change, how those changes affect the whole service and which benefits can actually fund continued use. That is the connection between a promising technology and a durable market.
