The demonstration lasts three minutes. The assistant summarizes a contract, retrieves a clause and drafts an impeccable response. Then come the questions that slow the sale: where do the documents go? Who can access the logs? How much will the service cost at scale? And how do you switch providers? Behind the debates on artificial intelligence sovereignty, notably at the RAISE Summit in Paris, a less spectacular battle is playing out: the battle over the purchase order.
Looking ahead to September 2026, this perspective is becoming central. This analysis draws on documented developments in the sector, notably the rise of open models and the adoption of the EU AI Act in 2024; the potential developments discussed remain forward-looking. The aim is not to attribute unverified announcements or commitments to summit speakers, but to examine what their stated ambitions mean for buyers.
Sovereignty Starts with the Requirements
The RAISE Summit brings together AI, infrastructure and investment players in Paris. This intersection exposes a contradiction: Europe wants to develop its own capabilities, while companies are looking for tools that deliver immediate results. The two objectives can converge. They do not automatically align.
A model designed in Europe may run on American infrastructure. A model with accessible weights may be hosted in France while relying on foreign hardware components. Conversely, an international service may offer European data residency without resolving every question about jurisdiction or government access. Sovereignty is therefore not a binary label, but a set of controls to verify.
For a procurement department, the first task is to define what must remain under its control: data, operations, updates, service continuity or the ability to leave. A bank, a manufacturer and a communications agency will face different risks and have different requirements. Buying “sovereign” without specifying those requirements amounts to comparing promises.
Useful Performance, Not Today’s Rankings
Public rankings provide an initial reference point, not a purchasing verdict. A model with excellent reasoning scores may disappoint when handling poorly digitized French documents. Another, less versatile model may be sufficient for classifying customer requests. The right question is not “which is best?” but “which performs our task successfully under our conditions?”
Testing must cover a representative sample: tables, incomplete emails, industry acronyms and ambiguous wording. Accuracy must be measured, but so must omissions, faithfulness to sources and the ability to acknowledge missing information. For a document assistant, an elegant answer accompanied by a fabricated reference is still a failure.
Speed matters too. A conversational tool cannot afford repeated delays; an overnight analysis can accommodate slower processing. Buyers must therefore examine actual response times, capacity limits and availability commitments. Above all, the entire system must be tested: the search engine, document retrieval and instructions can matter as much as the model.
The Headline Price Hides a Chain of Costs
The price per million tokens makes comparisons easier, but does not describe the final bill. A query may involve an extensive conversation history, several documents and multiple successive calls. Retries, automated checks and multimodal processing add further expenses. An attractive unit price can therefore come with costly usage.
In-house hosting does not eliminate these costs: it shifts them. Computing capacity must be funded, software maintained, access secured and sometimes scarce expertise made available. Infrastructure that is reserved but little used can cost more than a shared service. Conversely, steady volumes can make dedicated operations attractive.
The relevant metric is the cost per correctly completed task. It includes human review, integrations and errors requiring correction. A low-cost system that demands exhaustive checking may lose its advantage over a more expensive but reliable solution. Contracts should also clarify temporary discounts, overage charges and the terms governing price revisions.
“Hosted in Europe”: What Exactly Does That Cover?
Data location is often presented as a box to tick. In practice, several flows coexist: documents sent to the model, generated responses, technical logs, backups and monitoring data. A guarantee covering primary storage does not necessarily answer questions about support or telemetry.
Buyers should request a map of these flows, a list of subcontractors and retention policies. Is the data used to train or improve models? What access is possible from other countries? Who holds the encryption keys? The answers must appear in enforceable commitments, not just a sales presentation.
The GDPR already governs the processing and transfer of personal data. The EU AI Act adds obligations that vary according to systems and their uses, with a phased implementation timetable. Neither automatically makes a European offering compliant. Likewise, a security certification must be examined within its precise scope: it does not necessarily cover the entire AI application.
Dependence Is Best Measured When It Is Time to Leave
Technical lock-in is not just about the model. It resides in proprietary interfaces, document repositories, evaluation tools and mechanisms for calling external software. After months of integration, replacing an API can become a major undertaking, even when two services appear to offer similar features.
Open-weight models offer flexibility: a choice of hosting, adaptation and, depending on the license, redistribution. They do not, however, guarantee access to training data or long-term maintenance. Their licenses, hardware requirements and the availability of expertise remain decisive. “Open” therefore does not automatically mean independent.
Four Pieces of Evidence to Request Before Signing
- A reproducible test of a business task, with success criteria and error measurement.
- A total cost estimate across several usage volumes.
- Contractual documentation of data flows, access and retention.
- An exit exercise: export the data and test an alternative solution.
This final test is decisive. An exit clause only has real value if the organization knows how to retrieve its documents, configurations and histories in usable formats. Planning an alternative does not mean duplicating everything: it means retaining credible bargaining power and the ability to maintain continuity.
What next? A plausible outcome of the RAISE Summit debates is not so much an automatic victory for domestic providers as a more professional approach to procurement. Companies could spread their use cases across international services, European offerings and models run in controlled environments. This approach will bring costs and complexity. But it establishes the right criterion: AI that is genuinely under control is not just AI you can choose; it is also AI you can audit, replace or shut down.
