Estonia's data protection inspectorate has issued a non-binding circular telling educational institutions what to check before adopting an AI-enabled application, and saying that AI output should not be the sole basis for a decision about a learner or a member of staff

What the circular actually covers

The document is a circular addressed to educational institutions, and it covers applications generally before it reaches AI. Its first half deals with whether an application processes personal data at all, the legal basis for a school's processing, the processor agreement with the vendor, sub-processors, and where the data is stored and whether it leaves the European Economic Area.

Its definition of an application is broad: software or a web service used to run teaching, administrative or support processes, allowing users to learn, communicate, manage documents, assess knowledge or create content. It names Moodle and Microsoft Teams as examples of the two deployment types it has in mind.

The AI half opens with a working description, not a legal definition. An application running on tehisaru, the Estonian term used throughout, is described as one able to learn from data, analyse it and generate answers or make decisions, rather than merely executing pre-programmed instructions. The circular's point is that such applications attract the ordinary data protection requirements plus considerations specific to AI.

The human oversight passage, and what mood it is written in

This is the part that travels furthest in summaries, so it is worth reading in the original. Under the heading Inimjärelevalve kui prioriteet, human oversight as a priority, the circular says the role of AI is above all to support the decisions of a teacher or other school staff member and not to replace them entirely, especially where decisions affect learners, and that solutions should therefore be preferred in which AI acts as an aid to the decision and the final decision is taken by a staff member after substantive evaluation of the AI output.

A second sentence, in the section on the reliability of AI outputs, carries the same idea: AI output should not be the sole basis for decisions concerning learners or staff. The Estonian is "ei tohiks", the conditional, not "ei tohi". Our reading of that grammar, and we mark it as ours: the inspectorate is stating a strong expectation, not asserting a flat prohibition in this document.

One qualification belongs here, and it cuts against that reading. The inspectorate's own news item announcing the circular, published the same day at 15:37, paraphrases both points in the indicative rather than the conditional: that AI generated content peab, must, be checked before use, and that human oversight peab, must, be retained for decisions concerning learners. The circular's own sentences use tuleks and ei tohiks for the same two propositions. We treat the circular as the document of record, because that is the instrument schools are pointed to, and we record that the authority's own summary of it is firmer than its text.

That distinction is not a technicality, because a genuine prohibition on solely automated decisions with legal or similarly significant effects already exists in Article 22 of the GDPR and binds regardless of what any circular says. What the circular does is tell schools where the inspectorate expects the line to fall in an educational setting, which is a good deal broader than Article 22's own threshold.

Risk assessment, profiling and the DPIA sentence

The circular's risk passage names three uses by way of example: profiling of learners, automated assessment, and making recommendations. On higher-risk AI use of that kind, the Estonian reads that possible risks are to be assessed first, and then, in the following sentence, that where necessary a further risk analysis or a data protection impact assessment should be carried out.

The mood shifts between those two sentences. The first uses tuleb, the indicative. The second uses tuleks, the conditional, and qualifies it with vajadusel, where necessary. Read precisely, the circular is not saying that profiling of learners always triggers a DPIA. It is saying to assess the risk first and to reach for the fuller instrument where the assessment shows it is needed.

Anyone who needs a firmer answer than that has to go back to Article 35 of the GDPR and to the inspectorate's own DPIA threshold material, which this circular does not reproduce and which we did not open for this piece.

What the circular restates without creating

Several of its expectations are duties that already bind Estonian schools under the GDPR, and the circular restates them without adding to them. A processor agreement with a vendor processing personal data on the school's behalf, covering sub-processors. Transparency to learners, staff and where relevant parents about which applications are used, for what purpose and on what principles. A processor's duty to notify the school of a personal data breach without undue delay, and the school's own possible duty to notify the inspectorate and, in some cases, the individuals affected.

The AI Act appears once, and modestly. The circular says that in certain cases the AI regulation provides for separate transparency duties, giving the examples of a person interacting with an AI system and the use of artificially generated content, and points readers to the inspectorate's own AI section for detail. It does not attempt to map AI Act obligations onto school use, and neither do we.

The genuinely new material is organisational. The circular asks schools to keep an overview of the applications in use, recording the purpose, the personal data processed, the vendor, the storage location and whether a processor agreement exists; to make sure staff know which applications are approved and what personal data may never be entered into them; and to designate who assesses a new application before adoption and who staff go to with data protection or information security questions.

The vendor training setting, and the question it puts to procurement

One expectation is worth pulling out because it is concrete and easy to overlook at procurement. The circular says to establish whether user inputs are used to train the vendor's AI models or to develop the service, and whether that can be limited or switched off in the settings.

Its reasoning is stated plainly: if a vendor uses inputs to develop or further train models, the data is being used for a purpose unconnected with teaching and may leave the school's control, and the circular says the school bears responsibility for avoiding that situation. Where it is unavoidable, it says the learner is to be informed, along with their rights and who to contact.

Alongside that sit two input points, and the circular puts both as advice, not obligation. It says a school should assess whether personal data needs to go into the AI application at all, preferring anonymised or pseudonymised data and avoiding special categories such as health data. It also says a school should establish what usage and technical data the application itself collects, naming queries, logs and device information. Neither is a duty the circular imposes.

What we did not verify

What we opened: the inspectorate's own page carrying the full circular text, read in Estonian, and the three page Hariduse ringkiri 2026 PDF, retrieved from the inspectorate's 2026-09 document path and read as extracted text. We compared the two and they carry the same substantive text. We also opened the inspectorate's news item announcing the circular, dated 8 September 2026, which is where the webinar date comes from and which carries the authority's own indicative paraphrase of the two AI sentences.

A sourcing limitation we disclose here instead of burying it: the PDF's extracted text layer garbles several Estonian diacritics, rendering o with tilde as o with umlaut in places. Every piece of Estonian we reproduce, the pull quote and the document and section titles alike, is therefore taken from the inspectorate's own web rendering of the circular, where the characters are correct, and not from the PDF extraction.

What we did not open: the inspectorate's separate AI section, its DPIA threshold guidance, the Hariduse tehnoloogiakompass the circular recommends, and the webinar materials for 17 September 2026. We therefore do not state what the inspectorate's DPIA threshold list contains, and we do not describe the AI Act transparency duties beyond what this circular itself says about them.

What we refuse to claim: we do not say the circular requires human oversight, because its own sentence is in the conditional, and we do not say it mandates a DPIA for profiling, because it conditions that on necessity. We do not say it binds schools, because a circular of this kind states an authority's expectations. We do not characterise it as an AI policy for Estonian education generally; it is addressed to the use of applications, and the AI material is one part of it. Our reading of the Estonian conditional is marked in the body as ours.

Informational analysis for working professionals, not legal advice. Confirm how any rule applies to your situation with qualified counsel.

Related briefings

• Estonia's TTJA guidance on AI market surveillance
• Vietnam's education ministry rules on AI in schools
• Ohio's HB96 mandate for school district AI policies
• Europe AI regulation tracker
• AI Regulation News hub