AI data sovereignty for companies: contractual, or architectural?

The current conversation on AI sovereignty is about infrastructure: where the models run, where the data sits, who owns the compute. It misses the layer a business actually depends on. The most valuable thing your business has is what it knows, and the default path of this era hands that to the AI labs: inside their models, in formats you don’t own, under terms and roadmaps you don’t set.

Sovereignty, for a company, is owning what it knows. The principle behind it is separability: the model that reasons must be separate from the memory it reasons over, the same way the bank that holds your deposits must be separate from the one that trades. In a separable system, any model you permit can work on your knowledge and help improve it, but none can hold it.

There are two ways to get there, and they are not the same.

The first is a promise. Turn the AI’s memory off, so the model reads your knowledge each session rather than keeping it between sessions, and rely on the lab’s Terms of Service for the rest. As long as those terms specify no training and bounded retention, you have contractual sovereignty.

It is real, and it is available today. On a team or enterprise plan, memory is an administrator’s switch, not a request to the lab. If you signed an annual contract last week, nothing here asks you to break it; the one thing that contract was quietly buying, a memory the lab keeps for you, is the thing you switch off. The model keeps working for you.

Where the knowledge goes instead is a library you own: organizational memory written as plain text, at addresses any AI can read.

Is a promise enough? That is the fundamental tension of the AI era. If you turn the memory off but your knowledge still passes through the model, what inferences are kept? How much of the transcript feeds it? What does the cache hold, and for how long? Your contract answers some of this. No customer can check the answers, and a rogue swarm is not bound by anyone’s Terms of Service.

Contractual sovereignty rests on a promise. Architectural sovereignty rests on structure: you move the look-up outside the model. An independent operating system reads your library by address and assembles the context for each question before the model is called. The model never sees the repository. It is handed only what the question needs, and what any provider can keep is that one question’s context, never the library.

Starling calls this a Domain Language Model: stateless, interchangeable AI reading and writing to a sovereign repository. It is not a new lock. The knowledge is plain text you own and the standard is free, so the DLM can be replaced too.

Two answers, because there are two parties. Starling never trains on your memory, and that is structure, not a policy: your knowledge lives outside the model, and a task is sent only the context it needs.

The model providers do not train on what Starling sends them either, but that is their contractual commitment rather than Starling’s architecture, and it comes with bounded retention rather than none. Anthropic and OpenAI may keep inputs and outputs for up to 30 days for abuse monitoring, and Google for up to 55 days, then delete them. The current terms, per provider, are on the models page.

Sovereignty isn’t a trade-off. It’s an upgrade. Build on memory you own, and the next model upgrade is a swap, not a rebuild. Stored AI memory hardens and drifts, and a Monday pivot can take weeks to land; governed canon stays current, so every model works from what’s actually true. And the model never holds your business, only the slice a task needs, assembled before it’s ever called.

The urgency is practical, not theoretical. In June, export controls took two frontier models offline, and the work built on them stopped. More than a third of organizations with over $1 billion in revenue say they would need more than twelve months to move off a critical provider. Turn the AI’s memory off now, and your knowledge goes wherever you do.

Business can make separability the default now, the way HTTPS became the default: it is the better way to build, and companies can adopt it today. The labs that scrape and retain need rules, and those rules need to be verifiable. Business goes first; government makes it stick.

What each provider keeps, and how Starling limits it, is on our Trust & Security page. The eight terms every model should be held to are the AI Code of Conduct.