We have now had this conversation with enough schools to notice something. Almost nobody opens with a question about the AI. They open with a question about the data. If you are building for education and your pitch starts with model quality, you are answering a question nobody in the room asked.
Where does it go, and who else touches it
The first question is always about location, and it is never satisfied by the word cloud. Schools want to know which country the data sits in, which legal entity processes it, and which subprocessors are in the chain behind that entity. Dutch education has been through this before. The sector spent years negotiating collectively with the large platform vendors over exactly these terms, and school boards remember how that went. They learned to read a subprocessor list. When a vendor cannot produce one, the conversation is effectively over.
Schools evaluate the contract before they evaluate the model.
Does the model learn from our pupils
The second question is whether student input becomes training data. A lot of vendors answer this ambiguously on purpose, because the honest answer is complicated once you are reselling someone else's model. The answer a school needs is a flat no, in the contract, covering the vendor and every model provider underneath. Anything softer than that gets read as a yes with extra steps.
Who is accountable when it is wrong
AI in a classroom will produce something wrong. Everyone in the room knows this. What they are testing is whether you know it too, and whether you have designed for it. The good answer is not a higher accuracy figure. It is a teacher who can see what was generated, change it, and overrule it, and a system that treats that correction as normal rather than as an error state. Schools trust products that assume they will be checked.
What happens when we stop
Procurement people ask this and product people hate it. Can we export everything, in a usable format, and is the data actually deleted afterwards. It is a fair question and it is a cheap one to answer well. If leaving is easy, arriving is less frightening. Vendors who make exit difficult are telling schools something about their confidence in the product.
Why we built PrivateEdGPT the way we did
These five questions shaped the product rather than the FAQ page. PrivateEdGPT runs on European models with European hosting, no training on school data, with the teacher in the loop by default and export built in from the start. None of that is a feature list a child cares about. It is the reason an administrator lets the thing through the door, which is a precondition for a child ever seeing it.
Take these with you
- 1Schools evaluate the contract before they evaluate the model.
- 2An ambiguous answer on training data is heard as a yes.
- 3Design for being overruled by a teacher, not for being right more often.
- 4Easy exit makes adoption easier, not harder.