The Suniworks journal
Sovereign AI: eight checks before choosing a supplier
Check data flows, access, retention and exit arrangements before choosing a sovereign AI supplier. Includes a downloadable decision worksheet.
Updated on .

A supplier can host its application in Switzerland while sending PDFs to another service for text extraction. It can exclude training on your documents while retaining excerpts in its logs. Before choosing a sovereign AI service, examine the data journey, access controls, dependencies and exit arrangements.
For a Swiss SME, the useful question is specific: can we approve this use, with these documents and this configuration? Approval for public manuals does not automatically extend to confidential customer contracts.
Preparing for a supplier meeting? Download the sovereign AI decision worksheet: six pages covering eight evidence requests, three tests and blank fields for your context. Enter your email to access the PDF.
What does a sovereign AI offer actually promise?
In this guide, we examine sovereignty as your business’s ability to retain control over its data, operating arrangements and choice of replacement service. This is a decision method proposed by suniworks, not a label or certification.
Three commonly advertised features answer different questions:
- Swiss hosting identifies a location for the components covered by that commitment. You still need to identify those components and their supporting functions.
- A Swiss-developed model describes its origin. It does not describe the service built around it.
- An open-weight model can offer deployment and inspection options, subject to its licence. It does not make the entire application independent.
If your first decision is between local processing and a hosted service, our guide to local AI for Swiss SMEs covers that choice. Here, the focus is on evidence to request from a supplier, whatever architecture you select.
Write down your context first
Choose a real task: finding a procedure, drafting a sales response or querying technical documentation. Specify which documents are needed and which remain excluded. Add the authorised users, confidentiality constraints and processing destinations your business accepts.
This turns a meeting from an assessment of an impressive demonstration into a discussion about your actual purchase: “Show us what happens for this task, with the features we would enable.”
Requirements can differ between projects. Processing outside Switzerland does not, by itself, establish that an offer is unacceptable; a Swiss server does not establish that it is suitable either. The assessment depends on the data and applicable commitments. The Swiss FDPIC explains that data protection law also applies to AI (French source). For personal data, the obligations still need to be assessed in your circumstances.
Eight evidence requests before you sign
Each answer should identify the offer, enabled features and document version reviewed. Keep the written commitment separate from what you observe in a test and what remains unknown.
1. The organisations and their responsibilities
Ask who signs the contract, operates the service, hosts the data and delivers each purchased function. Link every organisation to its role and country of operation. A brand name does not describe the whole chain.
For outsourced personal-data processing, the FDPIC explains that the controller retains responsibility (French source). Reviewing the contract and subprocessors therefore belongs in the purchasing record alongside the technical assessment.
2. A document’s complete journey
Request a diagram from upload to answer: text extraction or OCR, document search, model, external tools, logs and backups. Ask for the location and recipients at each step. A diagram limited to the model leaves supporting processing unexplained.
3. Content-reuse rules
Ask what can be done with files, prompts and responses: training, evaluation, human review or other purposes. Excluding training does not answer questions about retention or access. The terms should apply to your subscription and selected settings.
4. Support access
Ask who may open a folder to investigate a problem, who authorises access, how long it lasts and what record is available. A list of roles does not show that permissions work. Testing two accounts with different permissions can help check part of that behaviour.
5. Retention by data category
Separate files, conversations, search indexes, logs and backups. For each category, request a retention period, deletion trigger and any exceptions. Removing a conversation from the interface does not establish immediate erasure of every copy.
6. Changes to models or suppliers
Ask which dependencies exist and how changes will be communicated. What can you reject? Which options remain if the new arrangement falls outside your accepted scope? A supplier list describes the service at a point in time; it does not guarantee stability.
7. Behaviour during an outage
Does the service stop, queue the request or switch elsewhere? For a switch, ask for the destination and triggering conditions. A test agreed with the supplier in an authorised environment can examine the planned scenario without causing a production outage.
8. Exit and recovery
Request a trial export and open it outside the product. Check recoverable items, formats, fees, timing and the treatment of remaining copies. A readable archive may not preserve permissions, relationships between documents or working instructions.
A fictional example: customer contracts stay on hold
An SME wants to use an assistant with confidential customer contracts. It requires processing destinations to be approved before use. The offer states Swiss hosting and no model training, but excludes OCR from the location commitment. Fallback is undocumented.
In this specific context, management could continue a pilot using entirely fictional documents while keeping real contracts out of scope until OCR and fallback are clarified. It is not finding the supplier compliant or non-compliant. It is recording that two steps needed for its decision remain unresolved.
With public documentation or different constraints, the decision could change. The worksheet therefore includes a blank context field rather than a conclusion to apply to every project.
Turn answers into an approval record
For each open point, record the evidence required, the person responsible and a due date. Then agree a few tests using fictional data: user separation, primary-service unavailability, and export and deletion of a small folder.
These tests produce observations in a defined configuration. They are not a complete audit or proof that improper access is impossible. The UK NCSC’s technical guidance emphasises understanding dependencies and documenting assets. It serves here as technical guidance, not a Swiss legal requirement.
Keep a dated decision covering approved use, permitted and excluded data, unresolved conditions, accepted risk and the decision-maker. Revisit it when features, subprocessors or processing destinations change.
Download the worksheet for your next meeting
The sovereign AI decision worksheet includes the eight evidence requests, three tests with expected outcomes and the fictional example. Complete the final page on screen with your own context and decision, then save your copy.
The download is available immediately after the email form, without automatically subscribing you to marketing emails.
For help assessing an architecture or preparing a pilot, see our approach to AI integration and data control.