In short
- Each AI call is a data transfer, and it is often the only one that leaves the region.
- Choose the model the way you chose the hosting: in region or inside your own network.
- An allow-list should refuse a call to any other endpoint, not log a warning.
The programme office has done everything the security team asked. The platform runs in the country, in an approved data centre. Sign-in goes through the organisation's own identity service. The files sit in local storage. Then the office asks to switch on the AI assistant, and the security review takes one more look.
Its question is short: when a programme director asks the assistant why the capital budget is behind, where does the text go? The answer, once someone traces it, is a model provider's endpoint in another country. The budget lines, the delay explanations, the names of the owners and the risk descriptions all travel with the question.
The AI request is refused, and the office goes back to writing its board pack by hand. The hosting was sovereign. The AI was not.
What goes wrong
Sovereign AI is usually discussed as national infrastructure: data centres, chips, home-grown models. For a programme office the question is narrower and more urgent. Where does the text go when someone uses the AI, and who decided that?
The AI is the new way out. A programme platform reviewed five years ago had one data flow that mattered: the database, and where it lived. An AI assistant adds another. Every question, summary and draft sends a slice of the record to a model endpoint. The hosting contract covered the database. It rarely covered the prompts.
The model is built into the product. Many tools are built around one model provider. The provider is a design decision, not a setting, so a client cannot point the product at a model its security team has approved, or at one running inside its own network. The choice is to accept the vendor's provider or to switch the AI off.
A setting can move the data without anyone noticing. Where the endpoint is configurable, it is often configurable without a guard. An administrator changes a model name, a default is updated in a release, and requests start going somewhere new. Nothing refuses the call, so nothing tells anyone.
The AI reads more than the person asking. An assistant that runs with a system account can summarise projects the user was never allowed to open. For a programme that spans ministries or business units, that turns the assistant into a wider door than any page in the product.
The training terms are vague. Whether prompts are kept, for how long and for what purpose sits in a provider's terms that nobody in the programme office has read.
The result is one of two bad outcomes. Either the AI is banned and the office loses the hours it could have saved, or it is allowed by exception and nobody can say where the data has gone.
What good looks like
Sovereign AI in a programme office is a set of plain controls. Five moves, which apply to any product.
Treat every AI call as a data transfer
Write down what text an AI request carries, where it goes and under which terms, in the same register where you record the database's location. A question about a project carries that project's name, numbers and notes. A board-pack draft carries the programme. Once the AI calls are on the data-flow map, they are reviewed with the same care as any other transfer.
Choose the model the way you chose the hosting
If the data must stay in the country, the model must sit where the data may go: an approved provider's endpoint in region, or a model running inside your own network. That choice belongs to the client and its security team, and the product has to accept it. A vendor that offers only its own provider has made the sovereignty decision for you.
Make the allow-list refuse
Name the endpoints and providers the AI may reach. Any request to something else should fail, with a clear message, rather than go through with a warning in a log. Failing closed means a misconfigured setting stops the AI for an afternoon. Failing open means it sends the programme abroad until an auditor notices.
Let the AI read only what the person may read
The assistant should act with the permissions of the person asking, and no wider. Its answers then match what the person could find by opening the pages. Changes it proposes should go through the same approvals as a change typed by hand. I set out that pattern in agents that act, and the person who signs.
Test on the model you will actually run
A model inside your network, or an approved one in region, is often not the largest model on the market. Ask the vendor which tier of model the product is committed to and test the assistant on that model, with your data, before you sign. A feature that only works on the most expensive model abroad does not meet a sovereignty requirement.
In Lodestar, Polaris works on the model you choose: any model through your own API key, or a model hosted inside your own network, which is how air-gapped deployments run. Once an allow-list of approved providers and endpoints is set, any other AI call is refused rather than sent, and we do not train models on client data. More detail is on the security page.
Questions for your next review
- When someone asks our AI assistant about a project, which endpoint receives the text, and in which country?
- Who in our organisation approved that endpoint, and where is the approval recorded?
- If the endpoint setting were changed tomorrow, would the next AI call be refused or sent?
- Have we tested the assistant on the model we would be allowed to run, rather than the one in the demo?
A programme office can keep its database in the country and still lose control of its data one question at a time. Sovereignty has to cover the AI calls too, or it covers less than it claims.
Short answers
What is sovereign AI for a programme office?
For a programme office, sovereign AI means the organisation decides where its AI requests go: to an approved provider's endpoint in its region, or to a model inside its own network. Any other endpoint is refused, the AI reads only what the person asking may read, and no data is used to train models.
Does an AI assistant send programme data out of the country?
It can. Each AI request carries part of the record, such as project names, budgets and risk notes, to the model's endpoint. If that endpoint is abroad, the data leaves the country even when the platform itself is hosted locally. Check where the endpoint is and who approved it.
Can Lodestar run its AI without an outside AI provider?
Yes. Lodestar can run its AI on a model hosted inside the client's own network, which is how air-gapped deployments work, or on any model through the client's own API key. Once an allow-list of approved providers and endpoints is set, any other AI call is refused rather than sent.
