What to ask an AI tool before December 2027
Business schools in Europe are starting to put AI tools in front of their participants: a tutor on the MBA capstone, a feedback tool on thesis drafts, an assistant inside a custom programme for a corporate client. A school that lets such a tool read a participant’s work and decide what they do next is, in the language of the EU AI Act, deploying a high-risk system. The obligations that come with that start on 2 December 2027. This piece sets out the dates, what counts, and seven questions to put to any vendor before then, including us. It is not legal advice; your data protection officer and counsel decide what applies to your institution.
The dates, after the Omnibus
The Digital Omnibus on AI, Regulation (EU) 2026/1744, entered into force on 27 July 2026 and moved the high-risk deadlines, as Cooley summarises. The obligations for standalone high-risk systems listed in Annex III, which is where education sits, now apply from 2 December 2027. AI embedded in regulated products follows on 2 August 2028. Public authorities have an extended deadline of 2 August 2030 for their own deployer duties; whether a public university counts as one for this purpose is not settled, and a vendor’s provider duties are unaffected either way.
Three things apply already. The prohibitions in Article 5, including inferring emotions in education institutions, since February 2025. The AI literacy duty in Article 4, since February 2025 and enforced by national authorities since August 2026, as the Commission’s own Q&A explains. And the transparency rules in Article 50, since 2 August 2026. Our earlier post walks the full timeline.
What counts as high-risk in education
Annex III, point 3 lists four uses in education and vocational training: systems that determine access or admission; systems intended to “evaluate learning outcomes, including when those outcomes are used to steer the learning process”; systems that assess the level of education a person will receive; and systems that monitor and detect prohibited behaviour during tests.
Article 6(3) carves out systems that perform a narrow procedural task, improve the result of a completed human activity, detect patterns without replacing human assessment, or do preparatory work for an assessment. But an Annex III system that profiles natural persons is always high-risk. So an AI tutor that reads drafts, judges them against criteria and steers what the participant does next is in the second category by its plain words. A tool that only drafts feedback a human then marks may argue the exception. A vendor must have a position on which it is, and be able to say why.
What a deployer must do from December 2027
Article 26 puts duties on the school as deployer: use the system as the provider instructs; assign human oversight to named people with the competence, training and authority to exercise it; keep the logs the system generates for a period appropriate to its purpose; monitor it, and suspend use and inform the authority if a risk appears; and tell the people subject to it. Public bodies and private entities providing public services must also complete a fundamental rights impact assessment under Article 27 before first use.
Separately, and regardless of the AI Act, a data protection impact assessment under GDPR Article 35 is required where new technology processes personal data in ways likely to carry high risk, which describes an AI reading students’ work. In the Netherlands this runs through SURF’s model processing agreement and its sector DPIAs. SURF’s DPIA of Microsoft 365 Copilot in December 2024 advised institutions not to use it for the time being, on grounds of transparency and incorrect personal data; a 2025 update rated the risk medium. That is the bar a vendor will be measured against, and it is a reasonable one.
Seven questions for any vendor
- Which Annex III category are you in, and why? If the answer is “none”, ask which exception in Article 6(3) they rely on. A vendor without a view has not read the Act.
- Who decides? Name the human step. Can a person override the AI, is that override recorded, and what training do the overseers get? Article 26 makes this your duty, so the tool has to make it possible.
- What is logged, can we keep it, and for how long? The log is yours to retain. If the vendor cannot hand it over in a form a person can read, the duty cannot be met.
- Does it profile, score personality, infer emotion from voice or face, or watch behaviour during tests? Emotion inference in education is banned now. Profiling makes any Annex III system high-risk without exception.
- Where is the data, who processes it, is it used to train models, and how is it erased? This is the DPIA’s core. Ask for the processing agreement, the sub-processors, the hosting region and the erasure path.
- Do participants know they are talking to an AI from the first interaction? Article 50 requires it. Check it on screen, not in the contract.
- Can we see what it teaches from? The Act does not require this. A school should, because it is the only way to supervise what its participants are being told, and to correct it.
Add one internal item: an Article 4 record that the staff who run the tool were trained to use it, which the Commission says can be a simple internal record.
Where Ulern stands
We build Ulern, the applied-learning platform for business schools and executive education, so read this section knowing that, and read our answers as a vendor’s answers.
Our reading of Annex III is that an AI tutor which evaluates learning to steer it falls under point 3(b), and we are preparing as a provider for 2 December 2027: risk management, technical documentation, logging and the design of human oversight. A supervisor decides at every gate, and every decision is recorded. The record of each participant’s work is the log, and it exports. There is no emotion inference, no personality scoring and no proctoring. Data is hosted in the EU, AI providers do not train on it, a data processing agreement comes with every contract, and erasure is on request; the details are on our security page. Every lesson can be followed back to the page of the school’s material it came from, which is question seven, shown in this walkthrough. We have not yet been through a conformity assessment, and we will say so plainly until we have.
This is the last of seven pieces in Assign and hope, a series on what goes wrong while a programme’s project runs, and what a remedy must have. The others: Did they write it, or did the AI? · Activities are not outcomes · Supervision does not scale · The 90 days after the programme · What corporate clients now ask of schools · The thesis in the age of AI