40 Questions to Ask Before Signing a Sovereign AI Vendor

Every AI vendor in a sovereignty conversation says the same three words: "we're EU-hosted." Almost none of them can tell you, unprompted, where their embedding model runs, whether their vector database has ever been migrated, or what an audit trace actually looks like when you ask to see one.

That gap between a claim and evidence is exactly what a scored checklist is built to expose. We put together 40 questions across six categories, worth 80 points, with a threshold of 60 to even be considered ready for regulated deployment. You don't need all 40 to start a useful conversation with a vendor. You need to know which ones disqualify fast, and what a real answer sounds like compared to a rehearsed one.

Why "trust us" isn't a vendor answer

A confident yes is not evidence. It's a claim, and claims are free.

The rule we'd apply to every one of these 40 questions is simple: when a vendor answers with certainty, ask for what's behind it, a contract clause, an architecture diagram, an audit report, or a live product screen.

Score every answer on a 0–2 scale: 2 for specific and evidenced, 1 for vague or partial, 0 for refused or dodged. A vendor who won't commit to a number on a single question usually won't commit to one on any of them.

Sovereignty: 10 questions, 20 points

This is the section that narrows the field before you've discussed a single feature or a single euro. If a vendor can't answer these with named regions and contractual language, nothing that follows matters much.

1. Where is language-model inference hosted? Which region?
2. Where is the embedding model hosted?
3. Where is the vector database hosted?
4. Can we self-host each of these components? At what cost?
5. Which subprocessors do you use? In which jurisdictions?
6. Where are backups stored?
7. Where are audit logs stored?
8. What is your DPA position on transfer to third countries?
9. Do prompts or responses leave the EU under any circumstance?
10. Under what conditions would model inference happen outside the EU?

Question 9 is the Atlantic test. If a prompt is processed on a US-hosted API even briefly, your data crossed the ocean regardless of where the results land afterward.

Governance: 8 questions, 16 points

Sovereignty tells you where the system runs. Governance tells you whether you can actually inspect and audit it once it's running.

1. Do you use our data to train models? Provide contractual language.
2. What audit trace do you produce per interaction? Show a sample.
3. Can we export the audit trace on demand? In what format?
4. What certifications do you hold? Show current audit reports (ISO 27001, SOC 2, or equivalents).
5. Are you EU AI Act ready? Show which obligations you support.
6. What happens when confidence is below threshold? Show us the fallback.
7. How do you handle prompt injection?
8. How do you handle model updates? Do you notify us?

Question 2 is the whole section in miniature: "show us" is the actual test. A vendor who can only describe a trace in words hasn't built one.

Technical fit: 8 questions, 16 points

Sovereignty and governance get a vendor through the door. This section tests whether the system actually works the way you need it to, and whether you're locked into today's model choice for the life of the contract.

1. Which LLMs do you support? Can we choose or swap?
2. Can we run our own LLM through your gateway?
3. What is your default response latency? P95?
4. How many languages do you support with source-grounded output?
5. Do you cite sources on every answer? Or is it configurable?
6. How is content indexed? What sources are supported natively?
7. How is content updated? Batch or real-time?
8. How does your system integrate with existing Drupal, Mautic, CMS?

Question 5 matters more than it looks. Citations aren't a nice-to-have feature, they're the mechanism that makes an answer checkable at all.

Commercial: 6 questions, 12 points

This is where a sovereignty story either holds up under a contract or turns out to be marketing with a compliance badge attached.

1. Is pricing usage-based or fixed?
2. Under a large attack or usage spike, could our cost balloon?
3. Show us a customer running your system at similar scale.
4. What is your exit process? What data goes with us?
5. What is your typical time-to-production for a new customer?
6. What support is included at each tier?

Question 4 is worth pausing on. A sovereignty story with no defined exit is a dependency story wearing a compliance badge.

Team and support: 4 questions, 8 points

Everything above can be answered in a sales call with a slide deck behind it. This section starts testing whether there's an actual operating team behind the answers.

1. Who runs incident response? Where are they based?
2. What is your 24/7 coverage model?
3. Can we escalate to your engineers, or only to support?
4. How do you handle customer feedback that a specific response was wrong?

Reference check: 4 questions, 8 points, ask a live customer

This section can't be answered in a sales call at all, it requires talking to someone who has actually run the vendor's system in production.

1. How long did production deployment take?
2. How does audit trace look in practice?
3. What's one thing you wish you'd known before signing?
4. Would you sign with this vendor again?

Question 3 is the one that surfaces what the sales deck leaves out, and question 4 is the hardest one for a reference to fake a good answer to.