Annex III Classification: 6 Critical Fields

The Annex III classification call is the same one it was last month. The evidence expected for it is not. Until 19 May 2026, a deployer running through Article 6 had a sparse interpretive record to lean on. There was the text of Annex III, two early court rulings and cautious national commentary. A reasoned internal memo, signed and dated, was about as defensible as the regulatory baseline allowed.

That baseline has now shifted. The Commission’s draft guidelines opened for consultation on 19 May 2026 and close on 23 June. The draft sets out worked examples across all eight Annex III categories. An Annex III classification call now has an external benchmark. The documentation that records the call is no longer just a reasoned position. It is now evidence of whether the call is consistent with the Commission’s reading.

What the draft adds to Annex III classification

The Commission’s draft is three documents: general principles, the Annex I product-safety route and the Annex III use-case route. The third runs to 148 pages and sets out worked examples for each of the eight areas, from biometrics to the administration of justice. The draft also adds three interpretive tests that any Annex III classification team needs to mirror.

The three new tests

The first is intended purpose versus marketed purpose. The Commission reads documentation, marketing and product positioning together. It refuses to give weight to boilerplate disclaimers buried in terms of service.

The second is the limit on human-in-the-loop carve-outs. A human reviewer at the end of an AI-shaped decision does not, on its own, move a system outside Annex III. The Commission is explicit on this point.

The third is the cross-reference to conformity-assessment routes under Annex VI versus Annex VII of the AI Act. This determines whether an internal or third-party assessment applies.

None of this is binding. The draft says so plainly. But the Commission’s reading is the reading market surveillance authorities will start from. It is the reading practitioner outlets have already adopted as the working baseline. A memo that ignores it has already lost its first argument.

What an Annex III classification memo looks like now

The memo is short. What it needs to do is survive being read by someone hostile to the conclusion. Six fields cover the ground.

The six fields

• System identification. Name, version, provider, deployment context, dates of decision and review.
• Intended-purpose statement. The deployer’s reading of what the system is for, in one paragraph.
• Annex III row tested. Which of the eight categories was considered, including those ruled out and why.
• Draft-guidance example mapped. The closest worked example from the Commission’s draft, with a one-line note on where the deployment differs.
• Exclusion reasoning. If the conclusion is non-Annex-III, the documented basis with reference to trigger language or worked examples.
• Named decision-maker and date. Who signed off, when and against which version of the draft guidance.

One worked example

A mid-market firm runs a screening tool that ranks 200 CVs down to a shortlist of 20. A human recruiter then reviews the shortlist and decides. The lazy reading is that the human takes the deployment outside Annex III. The Commission’s reading is that the screening itself triggers category 4 on employment, regardless of downstream review.

The memo records that reading. It cites the worked example. It names the decision-maker and concludes the deployment is high-risk under Annex III. The memo is not the work; the work is the discipline that produced it. The memo is what travels.

The deployer’s parallel duty under Article 26

The Annex III classification call does not just sit on the provider side. The deployer has a parallel duty under Article 26 of the AI Act. It is activated by the classification itself, not by the documentation. Once a deployment falls inside Annex III, the deployer takes on obligations on input data, monitoring, log retention and human oversight. In some cases a fundamental rights impact assessment under Article 27 also applies.

The documentation is what defends the deployment if the Annex III classification is later challenged. If a market surveillance authority knocks, the question is not “what did you decide.” It is “what was the basis on which you decided.” A memo that maps the deployment to a Commission worked example answers that on first reading. The same logic runs through the 2026 shift in AI governance evidence and through procurement-side evaluation work.

The 23 June consultation as a free regulatory check

There is a second-order use of the consultation that is rarely framed this way. The Commission has invited stakeholder feedback through an official questionnaire on the draft guidelines. The channel is open until 22:00 CET on 23 June 2026.

A practitioner with an Annex III classification edge case can submit a worked example through that channel without naming clients. Responses outside the questionnaire will not feed the final summary; those inside it will. This is the closest thing to a free regulatory diagnostic the AI Act offers in its first interpretive year. A well-framed submission costs an afternoon and lands the edge case on the Commission’s desk before the guidance is final.

What changes this quarter and by August 2026

Three diary entries cover the rest of the year. This quarter, the Annex III classification memo template should be a standing artefact in the governance file for every system in scope. By August 2026, deployer obligations crystallise under the current AI Act timetable, subject to the AI Omnibus that may shift the date to December 2027. In late 2026 or early 2027 the guidance is adopted in final form. The memo template travels across all three.

This closes part of the Article 6 guidance gap we flagged earlier this year. The work that follows is the Annex III classification documentation that proves a deployer noticed.

Open the governance file for one in-scope system this week and write the six-field Annex III classification memo against it. If the file does not yet hold one, that is the gap.