Eight questions to ask any AI supplier

Choosing an AI supplier commits you to far more than a model: a law that applies to your data, a place where processing runs, a pricing policy, exit conditions. This is where sovereignty comes into play, a word with a bad reputation in tech. People associate it with inward-looking rhetoric, with national solutions chosen on principle, sometimes at the expense of quality. There is another reading, more useful when it comes to signing.

For an organisation, being sovereign means keeping the ability to choose, and to change its mind.

Four levels where the choice is made

This ability plays out on four levels. Data, and the law it lives under. Infrastructure, meaning the place and the operator running the processing. Models, and the ability to switch them. And finally competence, which is often forgotten: understanding the models you use, knowing how to evaluate them on your own documents.

An organisation that can measure what a model is worth to it stays free to adopt another tomorrow.

Reversibility in practice

In practice, reversibility rests on four commitments. The index and enrichments must be exportable. Enriched documents and findings must live in open formats. An internal evaluation set must make it possible to test a new model within days. And the contract must plan the exit from the start, with its timelines and formats.

These commitments are checked before signing. That is what the eight questions below are for.

The eight questions

1. Under which law does my data live, including at your subcontractors?

   On 10 June 2025, before the French Senate’s committee of inquiry on public procurement, rapporteur Dany Wattebled asked Microsoft France’s director of public and legal affairs whether he could guarantee, under oath, that French citizens’ data entrusted to Microsoft through the public purchasing agency Ugap would never be transferred to US authorities without the agreement of the French authorities. His answer, recorded in the Senate minutes: “No, I cannot guarantee it, but, once again, this has never happened” (our translation). An honest answer, describing a legal reality: an American company remains subject to American law, wherever its servers are.

2. Which models do you use, and where do they run?

   Sovereignty is about where processing runs, which law applies to the data, and the ability to switch; the flag of the lab that trained the model comes second. At Oxebo, for example, open-weight models of American and Chinese origin run on Scaleway’s French infrastructure.

3. Can I choose the model myself?

   Some uses justify a cutting-edge proprietary model, and an organisation is perfectly entitled to choose one. The decisive criterion lies elsewhere: being able to switch models without rebuilding everything else.

4. How quickly can you switch to another model?

   An organisation that owns its corpus, its index and its evaluation sets can switch models in a few days, test the new one on its own questions, and compare the results before switching. Conversely, knowledge absorbed into a model retrained on its data, or into a supplier’s proprietary memory, ties the organisation to that model.

5. What happens to my index and enrichments if I leave?

   European law itself pushes towards reversibility: from 12 January 2027, Article 29 of the Data Act prohibits cloud providers from charging any fee for switching provider, and France, through a decision by the telecoms regulator Arcep of 20 February 2025 followed by a ministerial order of 17 November 2025, has already set the cap on data transfer fees at zero euros. A data-focused strand of the Digital Omnibus, still under negotiation in summer 2026, could change some aspects: the contract, for its part, must plan the exit from the start.

6. How do your prices change when model prices change?

   On 29 September 2026, at its DevDay conference, OpenAI launched a $500-a-month plan and announced that the usage included in its $200 plan would drop, on 30 October, from 20 to 10 times that of the Plus plan, according to The Next Web; the same day, the company presented GPT-6.1 Sol, a model it describes as close to its flagship at a fifth of the price per token. An organisation tied to a single supplier absorbs the increases and misses the decreases; an organisation free to switch models benefits from both.

7. What happens if a model becomes unavailable overnight?

   In June 2026, Anthropic suspended access to two of its models for nearly three weeks to comply with US Department of Commerce export controls, according to the company’s statement; access was restored from 1 July, after these controls were lifted, according to Al Jazeera. An administrative decision can therefore cut access to a model overnight, whatever the goodwill of its publisher.

8. How do you evaluate models on my own documents?

   Every model carries the biases of its training, whatever its origin: that is why you evaluate each model on your own documents and your own questions. Open weights, which can be downloaded, hosted and audited, are precisely what make that evaluation possible.

None of these questions rules out a model or an origin. All of them check the same thing: that you will be able to change your mind.