Sovereign AI for schools. Why your data has to stay in your building.

The requirement the banks already made non-negotiable

Sovereign AI has become a serious procurement conversation this year — for banks, for defence, for critical national infrastructure. The requirement is specific and unromantic: the data stays inside the customer's own walls, the data does not leave, and there is no dependency on an account in another country that can be switched off in an evening. A bank writes that into how it buys. It has the leverage to insist, and it does.

A school has the same requirement and none of the leverage

A primary school holds some of the most sensitive personal data in the country — children's records, safeguarding notes, SEND assessments. Under UK data-protection law and the ICO's Children's Code, that data is meant to be handled with more care than a bank's, not less. Yet most AI tools arriving in schools quietly assume the data can leave the building to be processed somewhere else, because that is the cheaper thing to build. The school gets whatever the vendor ships by default — because a school does not have a bank's procurement department to push back.

“Free” is a bill you haven't been handed yet

These tools are often free at the point of use, which is exactly the problem. Free AI you don't own is a bill you haven't been handed yet — payable in dependency, in data, and in the day a directive, a price change or a shutdown takes away the thing your week was built on. If you don't own the model, you are renting your own data back. Vendor lock-in is a security risk with a monthly invoice, and in a school the invoice is settled in children's data.

Built the other way round

Aevum is built the other way round. A shared base model is trained once, by us, on Dawn — the supercomputer at the University of Cambridge, through the UK's national AI Research Resource. From there the model diverges locally, in the building, learning from how the school actually uses it. The children's records never leave to get marked. To be plain about it: this is not “no training” — it is no training on your data. The model learns its craft on national infrastructure and does its job in private.

Reasoning a teacher can inspect

Ownership and privacy are not the whole of it. A teacher — or a safeguarding lead — needs to see why a tool reached an answer, not just the answer itself. An answer a teacher can't inspect is an answer a teacher can't use. Building for reasoning you can inspect is harder than shipping a black box. It is also the only version that belongs anywhere near a child's education.

Three requirements, one sentence

Own the model. Keep the data. Inspect the reasoning. The same three requirements the banks already insist on — pointed at the sector with the most at stake and the least power to ask for them. Britain is building sovereign AI for the institutions with the most leverage. We are building it for the classroom. That is the gap Aevum exists to close.

Everything is in development. When each piece is ready, you will get one email with the link. No countdowns, no weekly nudges.