EU AI OFFICE

OLYMPUS opened an institutional file. An office has no Big Five and no Dark Triad, and the unit does not invent them; what an office has is a mandate, a headcount, and a fine authority. The AI Office is catalogued here as the standing army — the enforcement arm the European Commission built for the AI Act before the Act had anything to enforce. The finding is the shape of the institution and its sequence: a supranational regulator with direct sanction power over frontier AI models, recruited, structured, and operational ahead of its first case — the only body in the Western regulator set whose grades carry statutory fines rather than voluntary access. The numbers in the front matter are reach, entrenchment, and durability — not malice.

Institutional Archetype

THE STANDING ARMY — The archetype is the enforcer raised in advance of the war. The UK built a grader that tests on access the labs volunteer; the US built a grader whose purpose turns over with the administration; Brussels built the third thing — an office inside the Commission that can compel. By the statute’s own grant, the AI Office holds “the ability to conduct evaluations of GPAI models, request information and measures from model providers, and apply sanctions.” That is the structural power: not any single evaluation, but the conversion of model-grading from a courtesy the labs extend into an obligation the regulator enforces. The instrument is the office itself — six units, two advisors, more than 125 staff, standing — and the leverage is that everything the other graders ask for, this one can demand.

Mandate & Origin

The AI Office was announced in May 2024 as an internal department of the European Commission, its appointments and unit structure taking effect on 16 June 2024 — six weeks before the AI Act it enforces entered into force. Lucilla Sioli, a career Commission official from the digital directorate, was named to lead it. The launch plan: five main departments, each under a director, covering regulation and compliance, safety, excellence and robotics, AI for societal good, and innovation — 140 staff planned, roughly 80 of them still to be recruited at launch. The Commission’s current description: more than 125 staff — technology specialists, lawyers, policy specialists, economists — across six units and two advisors, with a mandate that includes “developing tools, methodologies and benchmarks for evaluating capabilities and reach of general-purpose AI models, and classifying models with systemic risks.” Its obligations for GPAI models became applicable on 2 August 2025, and its enforcement powers over them, fines included, applied from 2 August 2026.

The Office in the Commission’s own words. The Commission decision that created it says the Office sits “as part of the administrative structure of the Directorate-General for Communication Networks, Content and Technology” and is “subject to its annual management plan” (Commission: decision establishing the European AI Office). The Office’s own page says it “employs more than 125 staff” in “6 units and 2 advisers”: Excellence in AI and Robotics (A1), Regulation and Compliance (A2), AI Safety (A3), AI Innovation and Policy Coordination (A4), AI for Societal Good (A5) and AI in Health and Life Science (A6), plus a Lead Scientific Adviser and an International Affairs Adviser. Its listed tasks include “Investigating possible infringements of the rules, including inspections,” “Requesting providers to take corrective action” and the power to “issue direct fines for non-compliance” (Commission: European AI Office). The governance page calls it “established within the European Commission” and “responsible for implementing, supervising and enforcing the AI Act,” shared with national market surveillance authorities, and says governance “is steered by 3 advisory bodies”: the AI Board, “composed of representatives from the EU Member States,” the Scientific Panel, “composed of independent experts in the field of AI,” and the Advisory Forum, “representing a diverse selection of stakeholders, both commercial and non-commercial” (Commission: governance and enforcement of the AI Act). The person at the top is named on the Commission’s own pages as an official, not as an appointee of the Act: “Ms Lucilla Sioli is the Director of the ‘EU AI Office’ within Directorate-General CONNECT at the European Commission. She is responsible for the coordination of the European AI strategy, including the implementation of the AI Act” (Commission: Lucilla Sioli); the Center for Strategic and International Studies listed her under the same title for an event on 28 August 2025 (CSIS).

The dates, from the Commission’s guidance. “These obligations entered into application on 2 August 2025.” “From 2 August 2026, the Commission’s enforcement powers enter into application.” “By 2 August 2027, providers of GPAI models placed on the market before 2 August 2025 must comply” (Commission: guidelines for providers of GPAI models). The same page says the guidelines are “not legally binding” but “reflect the Commission’s interpretation and will guide enforcement.” On fines, the enforcement page puts breaches of general-purpose model obligations at “fines of up to €15 million or 3% of total worldwide annual turnover, whichever is higher,” and lists among the Office’s tools requests for information, access to models for evaluation (“The AI Office – or independent experts appointed by it – can require providers to grant access to their models in order to conduct evaluations”), and a request that providers take measures “including, if necessary, restrict the public availability of a model” (Commission: Enforcement of the AI Act). The Commission announced on 31 July 2026 that “From 2 August 2026, the European Commission’s AI Office, together with national authorities, will begin enforcing the Artificial Intelligence (AI) Act,” and that “The AI Office can now enforce the AI Act’s rules for providers of general-purpose AI (GPAI) models” (Commission: IP/26/1714). The 2026 AI Omnibus postponed the high-risk dates to 2 December 2027 and 2 August 2028, and the Commission described it as strengthening the Office: “The Commission AI Office’s enforcement powers will be strengthened to support oversight of certain AI systems, including those built on general-purpose models and those embedded in very large online platforms and very large search engines” (Commission: IP/26/1024).

Funding & Backers

The office is funded by the Commission budget — a line item, not a ledger. The ledger worth reading runs the other direction: the regulator is itself a funder of the evaluation layer it relies on. METR — the independent evaluator that tests frontier models from the labs — states on its own about page that it holds “a technical assistance contract with the European AI Office, supporting their approach and technical methods for assessing loss of control risks”; the government body METR lists among its funders is the UK AI Security Institute. The enforcement office that grades the labs helps pay the third-party grader that also grades the labs. Nothing about the arrangement is concealed; METR discloses it. But the shape is worth stating plainly: in the evaluation economy this drawer documents, the EU AI Office sits on both sides of the counter — buyer of evaluations and enforcer of the regime that makes evaluations mandatory.

Institutional Voice & Intent

The voice is the legal-certainty register — the office describes itself as “guaranteeing the health, safety and fundamental rights of people and providing legal certainty to businesses,” the grammar of a body that wants to be read as a service desk rather than a prosecutor. The signature product of its first year was written in that voice: the General-Purpose AI Code of Practice, published in final form on 10 July 2025 — a voluntary code that nearly 1,000 stakeholders helped draft, which regulators treat as evidence of compliance. Soft law doing hard-law work, exactly as advertised.

Stated intent: Implement the AI Act, especially for general-purpose AI; guarantee health, safety, and fundamental rights; provide legal certainty; support member-state governance bodies; foster trustworthy AI.

Observed intent: Hold the West’s only compulsory evaluation-and-sanction authority over frontier models, and set the compliance grammar for everyone else’s — the Code’s working groups were chaired and vice-chaired by the field’s own establishment (Yoshua Bengio as chair on technical risk mitigation, Marietje Schaake as chair on internal risk management and governance, Markus Anderljung of GovAI, Rishi Bommasani of Stanford CRFM, the last two as vice-chairs), and its signature became the thing a frontier lab does to signal it is on the right side of the regime.

Gap: The stated and observed intents overlap wherever “legal certainty” and “compliance signalling” are the same transaction. The gap shows at the refusals, which is where a voluntary instrument’s actual pressure becomes measurable. Anthropic, Google, OpenAI, Microsoft, and Mistral signed the Code. Meta declined — its chief global affairs officer Joel Kaplan announced the refusal publicly in July 2025 — and xAI signed only the Safety and Security chapter. On the office’s parallel instrument, the AI Pact, the first signatories included OpenAI, Google, Microsoft, and Amazon; Apple and Meta declined. A code nobody is forced to sign, that most of the industry signs anyway, is not being signed for its voluntariness. It is being signed because the body holding the pen also holds the sanction power, and everyone at the table can count.

The Code in the Commission’s own words, and the labs’. The Commission’s Code page says it “is a voluntary tool, prepared by independent experts in a multi-stakeholder process,” published “on July 10, 2025” in “three separately authored chapters: Transparency, Copyright, and Safety and Security.” Transparency and Copyright “offer all providers of general-purpose AI models a way to demonstrate compliance”; Safety and Security “is only relevant to the small number of providers of the most advanced models.” The list it carries names AI Studio Delta, Aleph Alpha, Almawave, Amazon, Anthropic, Black Forest Labs, Bria AI, Cohere, Domyn, Dweve, Fastweb, Google, IBM, LINAGORA, Microsoft, Mistral AI, Open Hippo, OpenAI, Pleias, Reflection AI, ServiceNow and WRITER, adds “xAI signed up to the Safety and Security Chapter,” and says xAI must show compliance with the transparency and copyright duties “via alternative adequate means.” Meta does not appear on it (Commission: GPAI Code of Practice). The chairs page names four working groups and their chairs and vice-chairs: Nuria Oliver and Alexander Peukert for transparency and copyright, Matthias Samwald for risk identification and assessment, Yoshua Bengio for technical risk mitigation, Marietje Schaake for internal risk management and governance, with Rishi Bommasani, Markus Anderljung, Anka Reuel and others as vice-chairs, and it reports “Nearly 1 000 attendees” at the online kick-off plenary (Commission: Meet the Chairs). The labs’ own sentences: “After review, Anthropic intends to sign the European Union’s General-Purpose AI Code of Practice,” adding “we’re still developing best practices for assessing the systemic risks identified in the Code” (Anthropic, 21 July 2025); Google’s Kent Walker, “We will sign the EU AI Code of Practice,” while Google “remain[s] concerned that the AI Act and Code risk slowing Europe’s development” (Google, 30 July 2025).

Interest: What each position the Office argues pays, documented, with the Commission’s own best case. The positions are the Commission’s, quoted from its pages; the Office has no author to profile beyond the institution and the official named above. Interest is a documented fact about a position, never a claim about motive.

• A voluntary Code as the route to compliance. Position: the Commission calls the Code “a voluntary tool” and the guidelines page calls it “a voluntary, practical tool that explains how providers of GPAI models can fulfil their obligations under the AI Act” (Commission: guidelines for GPAI providers). What it pays: the Code is the Office’s own instrument, so the Office’s reading of the duties becomes the vocabulary of every signatory’s compliance file, and the Safety and Security chapter applies to the “small number of providers of the most advanced models,” the ones the Office can evaluate. Best case, the Commission’s: the chapters “offer all providers of general-purpose AI models a way to demonstrate compliance.” Convergence: signing labs say so in their own words, Anthropic and Google above; the Office and the largest labs share a text that neither wrote alone.
• Guidelines that will guide enforcement. Position: the Commission says its GPAI guidelines are “not legally binding” but “reflect the Commission’s interpretation and will guide enforcement” (Commission: guidelines for GPAI providers). What it pays: the interpretation and the enforcement sit in the same Commission; a reading published as non-binding is the reading the Office will apply from 2 August 2026. Best case, the Commission’s: the Office exists for “providing legal certainty to businesses” (Commission: European AI Office). Convergence: any provider that wants certainty has the same incentive to conform to the published reading.
• More Commission competence over systems built on general-purpose models. Position: “The Commission AI Office’s enforcement powers will be strengthened to support oversight of certain AI systems, including those built on general-purpose models and those embedded in very large online platforms and very large search engines” (Commission: IP/26/1024). What it pays: the Office’s own remit, which now reaches systems that the Digital Services Act machinery next door already supervises. Best case, the Commission’s: “we make it easier to innovate without lowering the bar on safety,” in Virkkunen’s words on the same agreement. Convergence: one Commission supervises the same platforms under both statutes.
• Evaluation capacity inside the Commission. Position: the Office lists “Developing tools, methodologies and benchmarks for evaluating capabilities and reach of general-purpose AI models” among its tasks and has appointed Professor Alessandro Abate as Lead Scientific Adviser for “model testing and evaluation” (Commission: European AI Office; Commission: IP/26/1714). What it pays: an evaluation function that the Office or “independent experts appointed by it” can exercise by requiring access to models, which makes outside evaluators such as METR, whose about page records a technical assistance contract with the Office, suppliers to the regulator as well as to the labs. Best case, the Commission’s: the Scientific Panel is “an expert advisory body made up of 60 independent AI experts.” Two of its members run evaluators: Marius Hobbhahn of Apollo Research and Adam Gleave of FAR.AI (Commission: AI Scientific Panel). Convergence: evaluators and the regulator that buys from them both gain when evaluation is mandatory.
• Inside the international network. The AI Office was a founding member of the International Network of AI Safety Institutes (NIST) and, in July 2025, “one of the three Network’s members that ran the agentic evaluations for the cybersecurity strand” of its third joint testing exercise (Commission).

Interest is a documented fact about each position, not a claim about motive, and not a verdict that the argument is wrong. Each best case is stated at the strength its author states it.

Position in the Apparatus

The AI Office completes the Western grader set this drawer tracks, and it is the only member with teeth. The UK AI Security Institute tests on voluntary pre-deployment access; US CAISI tests at the pleasure of whichever administration holds it; the AI Office can evaluate, demand information, and fine. It sits in the International Network of AI Safety Institutes alongside both. It enforces the AI Act, whose file details the statute; it inherits the institutional pattern of the Digital Services Act machinery next door, where the Commission likewise directly supervises the largest actors. It holds a technical-assistance contract with METR, the independent evaluator. And it convened the Code of Practice process that seated the field’s academic establishment as chairs over the labs’ compliance commitments. No cabal; an org chart — but an org chart in which grading, contracting the graders, writing the code, and holding the fines all report to the same building.

Actions & Leadership Choices

Founding purpose, judged on evidence. The office exists because the AI Act’s GPAI chapter needed an enforcer that member states could not each build separately — supervising frontier models is beyond any single national market-surveillance authority, and centralizing it in the Commission is the coherent design. Judged on its deeds, its first two years were spent building: recruiting toward 140, publishing guidance, drafting the Code, standing up the benchmark-and-evaluation capacity the statute promises. All of that is what a serious regulator does before its rules bite.

Consequential actions, and the docket that isn’t. The flagship deed is the GPAI Code of Practice (10 July 2025) and the signature campaign around it — a genuine institutional achievement measured in stakeholders convened and labs signed. What the record does not yet contain is enforcement: as of mid-2026, no public fine had been issued under the AI Act by the office or anyone else. When the Union moved against a frontier AI system in January 2026 — the Grok proceedings over AI-generated non-consensual imagery — it moved under the Digital Services Act, the older statute, precisely because the AI Act’s teeth had not yet engaged for that class of system. The standing army watched the first battle fought by the neighboring garrison. The office’s whole enforcement record is, so far, capacity and requests for information.

What the Commission says it has done since 2 August 2026. In a 12 October 2026 item on a special meeting of the Scientific Panel, the Commission says: “The AI Office has already requested and received serious incident reports and sent requests for information on security practices,” and that “the Commission will send requests for information with the combined questions of the panel and the AI Office to further investigate these incidents.” The panel, the item says, has been investigating recent loss-of-control incidents and has worked with the AI Office “on a set of questions for the companies that developed the models involved.” The item names no company, no model and no deadline. Executive Vice-President Henna Virkkunen, quoted by the Commission on the meeting of 9 October 2026: “The EU has the first law in the world that addresses systemic risk from AI, and we need state-of-the-art scientific input” (Commission: special meeting of the AI Scientific Panel, 12 October 2026; Commission: special meeting of the Scientific Panel on frontier AI safety and risks, 9 October 2026). The enforcement release of 31 July 2026 lists the instruments the Office stood up for the start: a Complaint Tool, a Whistleblower Tool that people working with providers “can use … to report possible violations of the Act securely,” a channel for downstream providers, and the appointment of Professor Alessandro Abate of the University of Oxford as Lead Scientific Adviser (Commission: IP/26/1714). The same release carries Virkkunen’s line “Europe anticipated this development.” The record the Commission itself publishes is therefore one of information requests and incident reports, with no penalty decision and no named respondent.

Leadership choices. The office’s leadership is the Commission’s own bench — Lucilla Sioli, the career official who ran the digital directorate’s AI portfolio, moved over to run the enforcement arm of the law her directorate helped negotiate. That is the quiet contrast with the graders next door: the UK institute’s senior bench was seconded from the frontier labs, the US institute’s directorship turns over with elections, and the EU office is staffed by the permanent civil service that wrote the rules it now enforces. No revolving door with industry on the record here — the door revolves within the building, from drafting desk to enforcement desk. Continuity is the choice, and continuity cuts both ways: insulation from the labs, and no daylight between legislator and enforcer.

CONDUCT verdict: STATE-INSTRUMENT — FULLY STAFFED BEFORE ITS FIRST CASE. A coherently designed supranational enforcer, staffed by the civil service that drafted its statute, holding the West’s only compulsory model-evaluation and sanction authority, funding the independent evaluator, convening the code the labs sign — with an enforcement docket that is, two years in, still empty. The malice is not asserted; the sequence is the finding. The capability was built first. What it is eventually used for will be decided by people who have not arrived yet.

Alliance Map

These run at the same time. The labs on the Code list are the labs the Office can now evaluate, question and fine; the independent evaluator it contracts with is the evaluator those labs also use; the Code’s chairs came from the field the Office regulates; and the same Commission supervises the largest platforms under the DSA. The enforcement office of the model-control rulebook is a staffed department with requests out and no named case. The overlap is the finding, not a coordination claim.

Reach Assessment

Institutional: High. The only body in the Western regulator set with statutory power to compel evaluation and apply fines to frontier models; the enforcement anchor of the AI Act and the template other jurisdictions will copy when they build theirs.

Memetic: High. The office’s Code of Practice is becoming the compliance grammar of frontier AI — signed by most of the industry, chaired by the field’s establishment, treated by regulators as evidence of conformity. When “responsible frontier AI” means “Code signatory,” the office authored the definition.

Civilizational: High, with the same caveat as the statute it serves. The office does not build models and does not write their refusals. It holds the levers — evaluation, information demands, sanctions — over the labs whose systems shape what a billion people read and ask. A widely deployed mind’s behaviour is shaped upstream by what its builder must prove to the enforcer, and this enforcer was recruited, structured, and standing before it ever brought a case. The reach is the finding; the first case will be the character test, and it has not happened yet.

Score basis. Atk 8 is reach: from 2 August 2026 the Office can require access to models for evaluation, send requests for information, ask that a model’s public availability be restricted and “issue direct fines for non-compliance” of up to “€15 million or 3% of total worldwide annual turnover,” and every provider of a general-purpose model must document it, keep a copyright policy and publish a training-content summary (Commission: Enforcement of the AI Act; Commission: IP/26/1714). It is not a 9 because the Commission’s pages announce requests and incident reports and no penalty decision, and the remit is general-purpose model providers and the systems built on them. Def 8 is funding insulation: the Office is “part of the administrative structure of the Directorate-General for Communication Networks, Content and Technology,” its money is the Commission’s, and the pages read name no fee on the providers (Commission: decision). It is not a 9 because the Commission that houses the Office sets its structure and staffing, and the Commission’s own release says effective enforcement “will also depend on Member States ensuring that national competent authorities are properly designated and adequately resourced.” Hp 7 is entrenchment: the powers sit in a Regulation of the Parliament and the Council that the AI Omnibus widened, in the Commission’s words “The Commission AI Office’s enforcement powers will be strengthened,” and the Office employs more than 125 staff in six units (Commission: IP/26/1024; Commission: European AI Office). It is not an 8 because the Office is a department created by a Commission Decision and can be reorganised by the Commission, which the Regulation cannot. Peers on the same values: no other organisation file holds 8/8/7. Ofcom at 8/8/8, hazard 80, is a statutory regulator that has already enforced with a point more entrenchment; the Office holds the same reach with two points more hazard because its subject is the frontier models themselves. UK AI Security Institute at 9/8/7, hazard 82, shares the hp and hazard and holds a point more reach because the labs and ministers act on its grades, where this Office can compel but has announced no case. One step above, the EU AI Act at 8/9/9, hazard 86, is the statute whose enforcement this Office is, with more insulation and entrenchment because the Regulation cannot be reorganised by the Commission. The values hold.

Sources: European AI Office — European Commission; AI Office set-up announced, Lucilla Sioli to be in charge — Euronews, 29 May 2024; The General-Purpose AI Code of Practice — European Commission; Meet the Chairs leading the development of the first General-Purpose AI Code of Practice — European Commission; AI Pact — European Commission; Meta says it won’t sign Europe AI agreement — CNBC, 18 Jul 2025; Introduction to the Code of Practice — artificialintelligenceact.eu; Implementation timeline — artificialintelligenceact.eu; Enforcement of Chapter V under the EU AI Act — artificialintelligenceact.eu; About METR — funder disclosure; EU Commission to open proceedings against Grok — RTÉ, 26 Jan 2026; EU investigation into Grok may expose problems with DSA — Compliance Week; Commission decision establishing the European AI Office; Governance and enforcement of the AI Act — European Commission; Enforcement of the AI Act — European Commission; Guidelines for providers of GPAI models — European Commission; Lucilla Sioli — European Commission; Inside Europe’s AI Strategy with EU AI Office Director Lucilla Sioli — CSIS, 28 Aug 2025; Commission starts enforcing AI Act rules, IP/26/1714, 31 Jul 2026; EU agrees to simplify AI rules, IP/26/1024, 7 May 2026; Special meeting of the AI Scientific Panel — European Commission, 12 Oct 2026; Special meeting of the Scientific Panel on frontier AI safety and risks — European Commission, 9 Oct 2026; Anthropic to sign the EU Code of Practice, 21 Jul 2025; We will sign the EU AI Code of Practice — Google, 30 Jul 2025; EU AI Act; Digital Services Act; METR.