Schools that use Gradr can now switch on full EU sovereignty. Every company in the chain is then EU-owned, from storage to the AI that reads the students' texts. Here we explain what that means, how we have built Gradr and why we think it is worth it.
Over the past year, we at Gradr have been working to move everything that touches the platform to the EU. We have now come far enough that the AI models can be European too. Schools that use Gradr can now switch on full EU sovereignty. Every company in the chain is then EU-owned, from storage to the AI that reads the students' texts.
Here we explain what EU sovereignty means, how we have built Gradr and why we think it is worth it.
What does EU sovereignty mean?
Three concepts are often confused: data localisation, data protection and sovereignty. Data localisation is about where the server is. Data protection is about how personal data is processed. Sovereignty is about whose law ultimately applies to the data. A service can have the first two without having the third.
That is why it is not enough for the servers to be in the EU. Extraterritorial laws such as the US CLOUD Act and FISA 702 follow control of the company, not the location of the server. A European subsidiary of a US group can therefore be ordered to hand over data stored in Stockholm or Frankfurt. EU sovereignty covers four layers:
- Storage: where the data is, including backups, replicas and logs, which are often forgotten.
- Processing: every processor and sub-processor that handles the data, not just the hosting provider. That includes the AI model, the email service, error reporting and analytics tools.
- Jurisdiction: where the companies in the chain have their group headquarters, and therefore which laws they are ultimately subject to.
- Access: who can actually reach the data. The European Data Protection Board has established that access from a third country can in itself be a transfer under GDPR, even if the data never leaves the EU. A support technician outside the EU therefore counts.
The European Commission's Cloud Sovereignty Framework measures this in levels from SEAL-0 to SEAL-4. The highest level requires European control all the way down to the hardware, and almost no digital service reaches that today. So the question is not whether a service is sovereign, but how far down the chain the control reaches.
Many of the providers we know of have some kind of link to the US, often at a point in the chain nobody thinks about: a parent company, a sub-processor or a support tool. We have chosen to keep everything in the EU as far as possible.
How we have built Gradr
We have always had the same goal: the highest possible quality with the highest possible security. This is what we have thought about.
Storage. All data is held by EU-owned providers, including the backups.
Processing. We run most things ourselves, from email to monitoring. It takes more time, but it means no US cloud service sees student data when we send an email, log an error or check that the platform is healthy.
AI. This is where it is hardest. The best language models have long come from the US, and that is where development moves fastest. That is why schools in Gradr can use all kinds of AI models. Those who put sovereignty first can run Gradr with an EU-owned AI provider. We cannot promise that those models will always be as good as the best US ones, but the option is there for anyone who wants to give sovereignty that high a priority.
Access. Here we have learned how little it takes. A US support chat in the app is enough for personal data to be reachable from the US. The same goes for an error reporting tool that saves what the student wrote when something went wrong, or an analytics tool that tracks who clicks where. On their own they look harmless. Together they become holes in the chain. That is why we have gone through every tool that touches the platform, not just the big ones, and kept the number of people who can reach the data to a minimum.
Our developer Felix describes how the move itself was done technically in the devlog.
Why we have done it
Over the past year we have talked with a great many municipalities, school groups, IT managers, heads of digitalisation and security coordinators. We have heard the same thing again and again. People want systems they can trust, and they want to be able to explain to guardians, the municipal board and the data protection officer where the students' data ends up.
We know what things looked like ten years ago. Back then, an exam platform or a new assignment tool could come into a municipality without much trouble. A data processing agreement, a quick check, and you were up and running. That is no longer the case, and that is a good thing. When a service contains AI, it is scrutinised in a completely different way. The procurement officer wants to know which model is used and where it runs. The data protection officer wants to see the whole chain of sub-processors. IT security wants to know who can reach the data and how. And the municipal board wants to know that someone has really thought this through.
We have grown up under that pressure. We have never seen it as an obstacle, but as the requirements specification. That is why we have spent a great deal of time on this, more than most would expect for a company our size. We have mapped every provider, replaced those that did not measure up and built our own operations where there was no European alternative that was good enough.
It has come at a cost. Time that could have gone to new features has gone to things no user sees. But that is exactly what makes a difference when a municipality has to make a decision, and we would rather do the work before anyone asks than after.
Why it matters
Student texts are not just any data. In an essay for Swedish class, a 16-year-old writes about their upbringing. In religious studies, the student reflects on their own faith. In civics, they argue for their political views. These are exactly the kinds of assignments that GDPR counts as special categories of personal data. The student did not choose the provider themselves; the school did it for them. The trust lies with the school, and the school needs to be able to stand behind the choice.
Legally, the problem is that two legal systems collide. Under Article 48 of GDPR, an authority outside the EU may only demand personal data on the basis of an international agreement. At the same time, the CLOUD Act lets US authorities demand data from companies under US control, regardless of where the data is stored. The provider ends up caught between two laws, and the customer cannot contract the problem away.
What this looks like in practice became clear in June 2025. A lawyer from one of the world's largest US cloud providers was asked under oath in the French Senate whether he could guarantee that data about French citizens would never be handed over to US authorities. The answer was no.
The transfer decisions between the EU and the US do not provide stable protection either. The Court of Justice of the EU has struck them down twice, in 2015 and 2020, and the current decision has already been challenged. Meanwhile, the requirements in Sweden are getting stricter. The Public Access to Information and Secrecy Act requires that it is not inappropriate to disclose confidential information to a provider. The Cybersecurity Act requires security throughout the supply chain. The Government's cloud policy from 2026 asks municipalities and regions to assess their dependencies and avoid lock-in. With full EU sovereignty there is no US party in the chain, and then you do not need to follow those developments at all.
That is why we think schools should set high requirements, and set them for all systems, not just new ones. A system bought five years ago was reviewed with the questions of that time, and those questions are no longer enough when there is AI in the tools. Start with the systems that handle the most sensitive student data: the learning platform, assessment tools, student health and AI services. Ask where each sub-processor has its group headquarters, where the AI runs and which tools see data in the background. A system already in use deserves the same scrutiny as a new one.
What does the AI Act require?
AI that evaluates learning outcomes is listed as high-risk in Annex III, and the requirements apply from 2 December 2027. A central part is human oversight. In Gradr, it is always the teacher who assesses. The AI helps produce feedback, but it is the teacher who decides what reaches the student.
Frequently asked questions
Is it enough for student data to be stored in the EU?
No. What matters is where the provider and every sub-processor has its group headquarters, not where the server is.
Can Gradr run entirely on EU-owned infrastructure?
Yes. All data is already held by EU-owned providers, and with full EU sovereignty the AI also runs with an EU-owned provider.
Does the AI set grades in Gradr?
No. Gradr is a support in producing formative feedback, and it is always the teacher who assesses.
We are not done
We are proud of where we are today, but we are far from satisfied. What we have built over the past year is a foundation, not an end goal.
Every new feature, every new provider and every new AI model that comes into Gradr is measured against the same question: does this make the chain stronger or weaker? If the answer is weaker, it does not matter how good the feature is or how much time it would save us. Then we find another way, or we build it ourselves. That is how we have worked all the way here, and that is how we will keep working.
The European models get better every month, and we follow them closely. Our goal is that EU sovereignty will one day not be something you have to ask for. It should be the starting point. We are not there yet, but that is where we are heading, and we have no intention of slowing down.
We are a small team, and we know what every hour is worth. Even so, this is one of the areas where we have spent the most time, and where we will keep spending it. The requirements on AI in schools will get stricter, new risks will appear and the questions from procurement officers and data protection officers will get harder. That is a good thing. We do not want to be the provider that keeps up when the requirements are raised. We want to be the one that has already raised them.
The students' texts deserve it. The schools that trust us deserve it. And we intend to do whatever it takes to live up to that.
If you want to know more or review us, write to privacy@gradr.se.
