Insight · July 2026
On 15 July 2026, Ramirez BV submitted a contribution to the European AI Office's stakeholder consultation on the draft guidelines for the classification of high-risk AI systems under Article 6 of the AI Act. The consultation ran from 19 May to 23 July 2026 and feeds into the final guidelines the Commission will adopt.
This article summarises what we put forward and why. The full contribution, as submitted, is available below.
Why we took part
Section IV of the draft guidelines covers the education use cases of Annex III, including systems that evaluate learning outcomes. Where the line falls between summative assessment, which counts towards a final decision, and formative feedback, which supports learning, determines how an entire category of systems is classified.
That concerns us directly. Our engine has been in production since March 2026 at Esmond, our evaluation platform for higher education, active in three markets: students use it at their own initiative to get critical feedback on a text they wrote themselves. The system writes nothing, produces no grade or score, and the examiner never sees its output. Precisely that category, student-initiated evaluative feedback, is not yet visible enough in the draft text.
Our input
We did not file objections; we proposed clarifications, each referencing the relevant paragraph of the draft text.
Formative feedback on a text the institution later assesses itself, a thesis draft for example, should not fall under the high-risk provision merely because the work eventually receives a grade. What should be decisive is whether the institution uses the system's output in that assessment.
We asked the guidelines to name the two decisive criteria: who initiates the use, the student or the institution, and who decides the outcome, the examiner independently or the system. Any other reading would make even a spellchecker high-risk once the text is eventually graded. That cannot be the intention and would be disproportionate.
The profiling provision should operate only after a system falls within scope. A system that stays outside on the guidelines' own criteria should not be pulled back in through a side door.
We also proposed one concrete example for the guidelines to include: a system a student chooses to use, at their own initiative and expense, for critical feedback on their own written work, with no grade or score, invisible to the assessors and leading to no credential. Such a system strengthens the student's own capacity before the assessment takes place.
Beyond education
One observation in our contribution reaches beyond education. The draft text rightly warns that a limitation asserted only in the terms of service counts for little when the system's broader presentation permits risky use. We support that strict line, and asked for the converse to be recognised as well: a limitation enforced in the system's own design, for example a system architecturally incapable of producing a grade or score, is a stronger and more verifiable signal than a contractual clause.
That is not legal hair-splitting; it is how we build. Our systems assign no scores, not because a terms page forbids it, but because they are not built to. Providers who anchor limitations technically, rather than write them away, deserve to have that difference recognised.
Status
The guidelines are a draft and the consultation document is a working document of the AI Office; the final text may differ. What we publish here is our position and reasoning, not a judgement on the classification of any system.
The contribution was submitted on behalf of Ramirez BV, with consent for publication by the Commission.