Essential Annex III Reading for AIGP

The AIGP exam has always asked one question about high-risk AI: is this Annex III, yes or no? Since 19 May the Commission has added a quieter second question, and AIGP Annex III scenarios will start testing it. The first question is "what is the answer?" The second is "what reasoning would defend the answer?" That second question separates a confident exam read from a guess dressed up in regulation. Candidates revising over the next eight to ten weeks should treat AIGP Annex III readings as a two-part skill.

What the Commission published

On 19 May 2026 the Commission released draft guidelines on the classification of high-risk AI systems under Article 6 of the AI Act. The text runs to roughly 148 pages. It splits into general principles plus separate Annex I and Annex III strands. The Commission has opened it for stakeholder consultation until 23 June 2026. The IAPP's own coverage frames it as the most detailed interpretative material on Article 6 to date.

For the AIGP exam, none of this changes the Body of Knowledge. The BoK is the IAPP's domain map that defines what the exam tests. The risk classification framework continues to sit inside the domain on AI-specific laws. What changes is the evidence layer the exam can fairly expect a candidate to recognise inside a scenario stem. Draft guidance is indicative, not binding; graders write to current Commission thinking. Treat it as part of the live reading.

Three AIGP Annex III categories the exam keeps returning to

The exam loves Annex III categories where the trigger feels obvious but the carve-outs are sharp. Employment is the textbook case. A CV-screening tool that ranks 200 applications into a shortlist of 20 is in scope. A tool that schedules interview rooms is not. The Commission reads "intended purpose" across documentation, marketing and product positioning. A vendor cannot disclaim its way out by writing one sentence in the terms of service.

Education access is the second classic stem. AI that grades exams and feeds the result into admissions sits squarely in scope. A proctoring tool that flags suspected cheating sits in scope too, even where a human reviews every flag. The presence of a reviewer is not a carve-out. It is a high-risk compliance requirement. Candidates who treat human-in-the-loop as an off switch will lose this question.

Law enforcement is the third stem where the reasoning earns marks. Predictive-policing tools that score recidivism for parole decisions are in scope. An AI that transcribes witness statements is not, unless the same tool scores witness credibility. At that point it moves back in. Some adjacent uses, including risk assessments built solely on personality profiling, are not high-risk at all; they are prohibited under Article 5. The exam tests whether the candidate notices the difference.

The four exam moves a single AIGP Annex III scenario can trigger

A well-written scenario rarely stops at "classify this." It can quietly trigger four moves in sequence. First, classify the system against Annex III. Second, perform or review an impact assessment on the system as designed and built; that is the development-side reasoning. Third, perform or review an impact assessment on the system as selected and deployed; that sits at a different point in the lifecycle. Fourth, identify the third-party integration risks where a general-purpose model has been wrapped into the system. Get the order wrong and the scenario unravels.

The reason matters for exam strategy. The Commission's draft makes clear that general-purpose AI models do not classify into Annex III in isolation. The integration does. The integrator can step into provider obligations under Article 25. Answers that miss this point will read as confident but wrong.

Common traps the Commission has made harder to defend

Three misreadings still cost candidates marks. The first is the human-in-the-loop misread, treated above. The second is the marketed-purpose conflation. A system is labelled "decision support" in marketing, but the documentation, demos and contract describe high-risk decision-making. The Commission reads the whole picture, not the label.

The third is the general-purpose-AI confusion. A foundation model itself is not the Annex III system. The hiring tool or credit-scoring engine built on top of it is. The integrator may inherit provider duties.

Candidates who want to see how vendor-side governance reads under that integration pressure have two practitioner shorthands. Future Prep's coverage of AI evaluation-report reading and vendor jurisdiction risk both walk the third exam move. The recent analysis of operational AI governance lays out the inventory step that sits behind any honest classification call.

How to revise this in the next eight weeks

Revise to the original AI Act timeline. The political agreement around the Omnibus may shift application dates; the BoK does not chase political timing. The missing high-risk guidance has been a known gap since February. The broader 2026 roadmap analysis helps candidates separate exam content from political weather. The Commission's news announcement gives the official framing.

Read the eight Annex III categories. Lock in the carve-outs. Rehearse the four-move scenario walk. That is the cleanest path through any AIGP Annex III question this cycle.

Want structured revision aligned to the 2026 BoK? Start with the AIGP exam prep suite at 22academy.com/study.