European AI Office and the AI Act governance architecture — who enforces EU AI law, and the finding that agrifood has no seat anywhere in it

Europe (EU-27 plus EEA-EFTA observers; one third-country member listed)

Content

The July 2026 regulatory scan described the four AI Act governance bodies and left the question of agrifood representation as a gap (G-219, G-220). This cycle answered it by extracting the Commission’s Expert Group register record for the Advisory Forum and screening all entries. The answer is unambiguous: there is no agrifood, farmer, forestry, fisheries, food-processing or agritech organisation anywhere in the AI Act’s governance architecture. That absence is the corpus’s sharpest EU finding this cycle.

Headline. The EU built a four-body AI governance architecture — AI Office, AI Board, Scientific Panel of Independent Experts, Advisory Forum — seated only in June 2026, enforcing from 2 August 2026, with more than 125 staff, a 174-member Advisory Forum selected from over 700 applications, and zero agricultural seats across all of it.

European AI Office

Established within the Commission (DG CONNECT) by Commission Decision, it “enforces the rules for GPAI models and supports the governance bodies in Member States in their tasks”, with powers to “conduct evaluations of GPAI models, request information and measures from model providers, and apply sanctions”. Enforcement over GPAI and prohibited practices began 2 August 2026; penalties run to €15m or 3% of worldwide annual turnover for GPAI breaches and €35m or 7% for prohibited practices.

Organisation: six units plus two advisers — A1 Excellence in AI and Robotics; A2 Regulation and Compliance; A3 AI Safety; A4 AI Innovation and Policy Coordination; A5 AI for Societal Good; A6 AI in Health and Life Science; plus a Lead Scientific Adviser and an International Affairs Adviser. Note what is absent: there is no agrifood unit. The only agrifood reference on the AI Office’s own page is the GenAI4EU initiative, whose illustration names “AgriFood and Sciences” as one of fourteen industrial sectors; the July 2026 Cybersecurity and AI action plan has no agrifood line either.

Headcount and budget: “more than 125 staff” per the Commission page; a Commission official put it at “145 staff across six teams” in July 2026, with 34 in regulation and compliance and 38 in AI safety; an external policy lead described enforcement staffing as “only about 35 people”; the Commission has proposed 38 additional employees, subject to the 2027 budget. No published budget line for the AI Office was found (G-430). A separate hiring round advertised 40 new posts dedicated to enforcement, with an interest deadline of 8 September 2026.

Leadership: Director Lucilla Sioli; head of unit for AI Safety Matthieu Delescluse (January 2026); Lead Scientific Adviser Alessandro Abate (Oxford, announced 11 August 2026).

Enforcement instruments: requests for information (simple or by decision), model evaluations and requests for access, orders to take measures including restricting public availability, interviews and inspections; plus three intake channels — an AI Act Complaint Tool, an AI Act Whistleblower Tool, and a complaints channel for downstream providers using third-party GPAI models. No formal investigation, fine or model-restriction decision had been announced as of 14 September 2026; the office’s own FAQ describes “technical compliance dialogues” as its preferred first instrument.

Division of labour: the AI Office takes GPAI models, systems built on a GPAI model by the same provider or business group (widened by the Omnibus), and systems inside DSA-designated very large online platforms; national market surveillance authorities take other AI systems; the EDPS takes EU institutions. Where a national authority cannot finalise a high-risk investigation for lack of GPAI information, the AI Office must supply it (Article 75(3)).

European AI Board

Composed of member-state representatives, supported by the AI Office as secretariat, and — distinctively — “chaired by one of the EU Member States”: at its eighth meeting (11 June 2026) it met “under the chairmanship of the Cypriot Presidency of the EU Council”, which also announced Moldova’s observer status. Ireland holds the Council Presidency from 1 July to 31 December 2026, so chairing passes to Ireland’s delegation for this semester — inferred from the presidency calendar, not confirmed by a Board-specific Commission statement.

Work items: eight meetings from 10 September 2024 to 11 June 2026; sub-groups on Standards, GPAI, and AI Act / MDR-IVDR interplay. Board-endorsed documents include the GPAI Code adequacy assessment (1 August 2025), the Code of Practice on Transparency of AI-Generated Content adequacy assessment (9 July 2026), an MDR/IVDR-AI Act interplay FAQ, and three standards sub-group reports. No agrifood-specific sub-group exists.

Named national authorities (Single Points of Contact): AESIA (Spain), DGCCRF (France — not CNIL), Bundesnetzagentur (Germany), ACN (Italy, cybersecurity agency), AI Office of Ireland, RDI (Netherlands), Traficom (Finland), MDIA (Malta), AKOS (Slovenia), Post- och telestyrelsen (Sweden), ANACOM (Portugal), KRiBSI (Poland), the Danish Agency for Digital Government, Lithuania’s Communications Regulatory Authority, ANCOM (Romania). Many designations are not finally adopted — including Spain, France, Germany, Italy’s MSA row, Netherlands, Sweden, Poland, Czechia, Estonia, Romania, Luxembourg, Slovenia, Cyprus, Greece’s sectoral set and Portugal. CNIL states it “is already designated as the authority to monitor prohibited uses and should soon be designated as the market surveillance authority for certain high-risk AI systems”; Germany’s Bundesnetzagentur designation still requires the KI-Marktüberwachungs- und Innovationsförderungsgesetz to pass.

Scientific Panel of Independent Experts

Established by Article 68 with Implementing Regulation (EU) 2025/454. It “advises the AI Office and national authorities on the implementation the AI Act and assessment of the impacts and risks of General-Purpose AI (GPAI) models”, with responsibilities to “alert the AI Office to systemic risks, advis[e] on GPAI classification and evaluation methodologies and support market surveillance”. Powers: qualified alerts (Article 90, simple-majority approval, can trigger Commission information requests, evaluations and systemic-risk designation), information requests (Article 91(3)), and model evaluations (Article 92).

60 independent experts appointed 1 June 2026, on 24-month renewable terms in personal capacity, at least one per member state and EFTA/EEA country. Named members include Yoshua Bengio (Canada), Miles Brundage (US), Adam Gleave (UK), Marius Hobbhahn and Judith Simon (Germany), Michèle Finck (Luxembourg), Maarten de Rijke and Nandi Schoots (Netherlands), Seán Ó hÉigeartaigh (Ireland), Céline Castets-Renard, Raja Chatila and Tom David (France). No agrifood specialist is evident from the published biographies, no chair or vice-chair has been published, and no opinion, alert or evaluation output exists as of 14 September 2026 (G-429). The Commission expects a model-evaluation capacity call to make the Panel “operational by 2027”.

Advisory Forum — the 174 members, screened

Mandate: “a general advisory body to the European Commission and the AI Board, supporting the implementation and enforcement of the AI Act”; it may prepare opinions, recommendations and written contributions, create sub-groups, and is a mandatory consultee for standardisation requests (Article 40) and common specifications (Article 41). Members serve two-year renewable terms (maximum four years); the Forum meets at least twice a year, publishes an annual report, and elects two co-chairs.

174 members selected from more than 700 applications, announced 1 June 2026, plus five permanent members (FRA, ENISA, CEN, CENELEC, ETSI). The kick-off meeting was held 19 June 2026. The Commission’s Expert Group register record (E04009, last updated 13 August 2026) lists 178 entries (173 appointed organisations plus the five permanent members).

The screening result: the roster contains no agrifood, farmer, forestry, fisheries or agritech organisation. Copa-Cogeca is not a member. CEMA is not a member. What is there: industry and tech (DIGITALEUROPE, Amazon Europe Core, Anthropic, Google Ireland, Microsoft, Mistral AI, OpenAI, Cohere, Hugging Face, IBM, Qualcomm, Siemens Energy, Thales, RELX, IKEA/Ingka, Credo AI and others); civil society and rights (Access Now Europe, EDRi, ECNL, Ada Lovelace Institute, CDT Europe, Panoptykon, Irish Council for Civil Liberties, Avaaz, Mozilla Foundation, SaferAI, AlgorithmWatch, BEUC, ANEC, ENAR, EDF, ERGO Network); labour and commerce (ETUC, EPSU, EUROCHAMBRES, EuroCommerce, MEDEF, BDI, DIHK, Bitkom, Confindustria, Orgalim, ACEA, CER, FEM, ELA); and research institutes (Inria, CEA, BSC, IMEC, Fraunhofer-adjacent appliedAI, Eurecat, TECNALIA, GovAI and many universities). The register also lists the National Research Council of Canada as a third-country participant.

ECNL records that 36 of the 174 members represent non-profit organisations, that selection concluded “9 months after the call for expressions of interest closed”, and that as of 7 July 2026 “key procedural aspects — such as rules of procedure or the selection of co-chairs — are yet to be resolved”. No Advisory Forum opinion or annual report is published (G-428).

Where agrifood’s only EU AI touchpoints sit: Copa-Cogeca appears as the agriculture-sector interviewee for the OECD review of the EU Coordinated Plan on AI — a research interview, not a governance seat. CEMA lobbies on adapting AI guidelines to agricultural machinery — a lobby position, not a seat. The Commission’s own DG AGRI + DG CNECT agriculture workshop (24 June 2026) is a dialogue, not a governance body.

What this unit is doing in the taxonomy

Anchors the EU AI governance-architecture cell — who enforces, who advises, and who is absent. Distinguishes from:

Why it matters for talks

Critical context