A form tutor at a secondary school has a chatbot soften a pupil's report comment. The pupil's name and the reason for concern are still in it. One floor up, the special needs coordinator asks the same tool to summarise a support plan for tomorrow's meeting. The board, nine schools, adopted an AI policy last year in which neither situation appears.
AI governance in education is about that distance: the set of agreements by which a board deals responsibly with the AI use of teachers and staff: which tools are in use, what has been decided per tool and under which condition, whether a teacher sees that decision at the moment they type, and whether the board can show that it works. The general version is in what is AI governance; at school almost every decision turns on whether pupil data may go in.
What do teachers use AI for?
Three kinds of work. Lesson material: an assignment at three levels, a test with a marking scheme. Feedback: marking an essay, wording a report comment. Administration: making a letter to parents easier to read, writing up a parent meeting.
A fourth category nobody experiences as AI use: the AI features inside learning materials the school already has, such as adaptive practice programmes. And a fifth: the work finished at home, on a personal account, because the school laptop was already closed. A policy that describes only the first three misses the fastest-growing part.
Which pupil data goes in with it?
For lesson material usually nothing. For feedback and administration almost always something: a name, a class, a grade, and in the heavier cases the reason for a concern. A support plan holds data about the health and behaviour of a minor, which is special category data under the GDPR. A report comment saying "has struggled to concentrate since the divorce" is a file too.
The task can nearly always be done without. Softening a report comment works just as well with "the pupil" as with a name, and summarising a support plan is only responsible in an environment the school has an agreement with. Which data comes out of a prompt as standard is in what you can and cannot share with AI. The board's decision is then not "AI yes or no", but per tool: is it allowed, and with which data.
What do the privacy covenant and the data processing agreement settle, and what not?
The Dutch education privacy covenant is the agreement between the education sector and suppliers of digital learning materials on the handling of pupil data, with a model data processing agreement. The school is the controller, the supplier the processor, and no school has to agree those terms again.
That covers what the school procures, not a free chatbot a teacher uses on a personal account: no supplier signed the covenant, there is no data processing agreement, and often the school does not know the tool is in use. A personal and a business account differ widely for AI tools, and what belongs in a data processing agreement for an AI tool applies to a school in full. The covenant answers "which suppliers do we have agreements with", not "which tools are being used".
What does the AI Act ask of a school?
Two things. Article 4 of the AI Act has asked since February 2025 that every organisation using AI systems works on the AI literacy of the people who use them. For a school that means: teachers know which tool they are using, which data does not belong in it, and where they report a new tool. What sufficient means per role is in AI literacy in 2026.
The second weighs more but touches fewer schools. Annex III lists education as a high-risk area, including AI used for admission and for evaluating learning outcomes. A teacher who has a chatbot prepare a lesson is not using a high-risk system. A school that uses AI to grade essays or admit pupils does have to ask, with the provider of that system, because the classification hangs on the intended purpose. When an AI application is high risk works through that distinction; the precise application differs per school and per system.
How do you divide this between board and school?
| Question | Board | School |
|---|---|---|
| Which AI tools are in use | One register for all schools, fed by what is actually opened | Reports a new tool, including one inside a learning material |
| Is this tool allowed, and with which data | The decision per tool: allowed, limited use, no decision yet, not allowed, with the condition on pupil data | Applies the decision in its own practice |
| How does a teacher use AI in a lesson or a test | Frameworks and room | The teaching choice, per subject and per assignment |
| Do teachers know what has been decided | The notice at the moment someone opens a tool | The conversation in the department |
| Does it work | Figures per board and per school, without names | A reason to adjust something |
What happens at a board with nine schools?
The board's policy from the opening names two tools: Microsoft 365 Copilot is allowed, ChatGPT is limited use without pupil data. When the board first looks at which AI services are opened from the browser, a lesson material tool appears at two schools, a transcription tool used in parent meetings at one, and a presentation tool nobody had recognised as an AI tool at all nine. None of the three is in the register, and the transcription tool hears conversations in which concerns about a child are discussed.
That is not an incident: the teacher with the transcription tool saves an hour a week, and nobody had said it was not allowed because nobody knew the tool existed. A policy with two tools is missing three a year later. The work that follows: assess three tools, choose a status per tool, check whether an environment with a data processing agreement can replace the transcription tool, and show that alternative as soon as someone opens the old one.
BeeSensible links a catalogue of 865 AI tools to the use observed in the organisation: only the domain, at most once per day per device, with no user id and no content. A board administrator sees which tools have no decision yet and records that decision in the AI Tools module; if the board has switched notices on, a teacher sees the decision in the browser as soon as they open the tool, never as a block. This is what that register looks like:
