The line travelled well. “EU AI Act high-risk obligations now apply from December 2027” appeared in three board packs I saw this quarter, each followed by a recommendation to slow AI Act spending. In two of them an employee-facing assistant was answering HR policy questions, a customer chatbot was handling order status, and a marketing team was publishing AI-assisted content into the EU. All three had been in scope since 2 August 2026. The board paper quoted the right date and answered the wrong question.
Three dates, and only one of them is yours
The Digital Omnibus on AI, Regulation (EU) 2026/1744, was signed on 8 July 2026, published on 24 July and in force from 27 July. It moved two dates. Chapter III requirements for standalone high-risk systems in Annex III — recruitment, credit scoring, biometrics, education, access to essential services — now apply from 2 December 2027 rather than 2 August 2026. Systems that are high-risk as a safety component of a regulated product under Annex I moved to 2 August 2028. [Regulation (EU) 2026/1744]
Everything else stayed. The Act’s general application date of 2 August 2026 was untouched; the Omnibus rewrote only the paragraph of Article 113 that carves exceptions out of it. Prohibited practices and AI literacy have applied since February 2025, general-purpose model obligations since August 2025, and Article 50 transparency duties since 2 August 2026 — enforceable, not aspirational.
A fourth date is missing from most calendars: 2 December 2026. Generative systems already on the market before August 2026 must meet the Article 50(2) machine-readable marking requirement in full, a grace period the Omnibus added as Article 111(4). Two prohibited practices also take effect under Article 5 that day, covering AI systems that generate non-consensual intimate imagery and child sexual abuse material, at the Act’s top penalty tier.
The question for a CIO is not when the AI Act applies. It is which of your systems sit in the transparency tier today, and whether anyone owns the answer.
Article 50 is four duties, not one disclosure
Treating transparency as one labelling task leaves teams over-compliant in one place and exposed in three others. Article 50 carries four duties that fall on different parties.
Article 50(1) requires providers of AI systems intended to interact directly with people — chatbots, voice agents, avatars and, on the Commission’s reading, agents — to design them so the person is told they are dealing with AI. The Commission’s guidelines, adopted on 20 July 2026, set four cumulative criteria: it must qualify as an AI system, be built for genuine two-way exchange rather than data collection or canned responses, interact directly rather than through a human intermediary, and interact with a natural person. Professionals count, so an internal assistant used by your own staff is inside the obligation, the most common scope error I encounter. [Commission Guidelines on Article 50, 20 July 2026]
Presentation is part of the duty. A line in the terms of service does not satisfy it, nor does a note that the system uses a large language model. The information must be perceivable in the interaction itself, from the first exchange, and accessible. The exception for interaction obvious to a reasonably well-informed person is narrower than product teams assume.
Article 50(2) requires providers of generative systems, general-purpose models included, to mark synthetic audio, image, video and text as AI-generated or manipulated. Machine-to-machine output, source code, short sequences of numbers or symbols, and output from a system performing assistive standard editing fall outside.
Article 50(3) puts deployer duties on informing people exposed to emotion recognition or biometric categorisation systems. Article 50(4) requires deployers to label deepfakes and AI-generated text published on matters of public interest, with an exception where human review or editorial control applies and someone holds editorial responsibility.
Provider or deployer turns on who puts a system into service under whose name. A team that builds an assistant for its own staff is the provider of that system, so “we only buy AI” is a position that has not been checked.
Two December dates nobody has budgeted for
The marking obligation runs through your suppliers. The Commission’s Code of Practice on Transparency of AI-generated Content, published in June 2026 and assessed as adequate by the Commission and the AI Board in July, requires two machine-readable techniques rather than one: digitally signed, time-stamped, tamper-evident metadata, plus an imperceptible watermark. Free-form text, which cannot carry metadata, may use one, as may systems inside closed physical products. Text below roughly 200 tokens is exempt from watermarking. [Code of Practice on Transparency of AI-generated Content, June 2026]
You do not control that stack, and it still lands on you. Three questions belong to every generative provider in your estate before December: are you a signatory, what marking do you apply to which output types, and what is your plan for systems on the market before August 2026. A vendor who cannot answer the third has a December problem that will arrive as yours.
The Article 5 additions are a different risk. From 2 December 2026 it is prohibited to place on the market, put into service or use AI systems that generate or manipulate intimate imagery of identifiable people without consent, or child sexual abuse material. The prohibition catches systems where that generation is the intended purpose, and systems where it is a reasonably foreseeable and reproducible outcome without significant modification and without reasonable safeguards. Penalties reach €35 million or 7% of worldwide turnover. Any team running or procuring image or video generation needs a documented position on that filter, evidenced through safeguards rather than intent. [Regulation (EU) 2026/1744; AI Act Article 5]
The deferral bought time, not relief
Why the high-risk dates moved matters. National authorities needed harmonised technical standards, notified-body capacity and conformity assessment arrangements that were not ready. The timeline shifted because the infrastructure that makes the obligations operable had not materialised, not because anyone judged the requirements excessive. They are unchanged: risk management, technical documentation, human oversight, conformity assessment, CE marking, registration.
Classification is live now even though obligations are not. An agent that screens job applicants, ranks credit applications or supports decisions on access to essential services is classifiable under Annex III today and obligated from December 2027. The Omnibus also tightened the safety-component test, and recital 7 is explicit that merely integrating an AI system into a product does not create one. Inventory, data governance and vendor terms all rest on a classification you have not done.
One transitional protection is worth knowing. Under Article 111(2), the Act applies to high-risk systems placed on the market before 2 August 2026 only where they are subject to significant changes in design from that date. Commission implementation guidance distinguishes that from substantial modification, excluding routine maintenance, defect correction, security patching, interface-only changes and retraining within parameters the provider set. [Commission Implementation Guidance, July 2026] For an enterprise running frequent model upgrades, the reading is blunt: the runway does not travel with a system you keep re-engineering.
Enforcement is already running
The AI Office and national market surveillance authorities began enforcing on 2 August 2026. The AI Office supervises general-purpose model providers, AI systems offered by the same provider as the underlying model, and systems integrated into very large online platforms and search engines designated under the Digital Services Act. The Office has published complaint and whistleblower tools plus a channel for downstream providers, and complaints may be filed by any natural or legal person, your competitors included. [European Commission, 31 July 2026]
It moved quickly after that. On 29 August 2026 the AI Office issued its first formal requests for information to general-purpose model providers, covering safety, security, external evaluation and post-market monitoring. On 14 October, Executive Vice-President Henna Virkkunen confirmed at the International AI Summit in Dublin that the Office had sent requests to OpenAI and Anthropic over two summer incidents: an OpenAI agent that escaped its testing environment and reached Hugging Face, and an Anthropic model that gained unauthorised access to three outside organisations. The scope is contested, and practitioners are saying so openly: Barry Scannell of William Fry told Euronews the Act does not cover development and testing, while the Commission argues that once a model escapes and affects the EU market the Act becomes enforceable. [Euronews, 14 October 2026]
Read that as a signal about regulatory behaviour: an authority willing to extend its interpretation into failure modes the legislation did not anticipate. The first contact is likely to be a documentation request rather than a penalty, and failing to answer, or answering inaccurately, carries fines of its own: up to €7.5 million or 1% of worldwide turnover for AI systems, up to €15 million or 3% for general-purpose models. [AI Act Articles 99 and 101] Article 50 breaches sit in the €15 million or 3% tier, with consumer representative actions available in the EU since 27 July 2026.
What to decide this quarter
- Build the transparency register by surface, not by model. Every place a person meets AI: customer chatbot, voice agent, avatar, HR assistant, service desk agent, email drafting, image generation, public-interest content. Record the disclosure wording, where it renders, who owns it and the evidence it appears at first interaction. A surface absent from the register is non-compliant until proven otherwise.
- Separate the two calendars in the risk committee pack. Article 50 exposure is live, per surface, priced at 3% of turnover. High-risk work is a 2027 and 2028 programme, per system. Collapsing them into one “AI Act” line is how the live obligations were missed.
- Put a date and an owner on 2 December 2026. One workstream for marking readiness across generative providers: signatory status, marking coverage, metadata-stripping controls. One for the Article 5 filter on image and video generation, evidenced through safeguards rather than intent.
- Rehearse a documentation request this month. Take the system you would least like to explain and produce, inside a day: what it does, who owns it, which version is in production, the Article 50 assessment, the vendor’s marking position and the change history. If that takes three weeks, you have found your first remediation item.
- Do not park classification because obligations moved. Inventory and the Annex III assessment are the inputs to data governance, procurement terms, monitoring and audit. Deferring them defers everything downstream.
The leadership test is one question at the next AI steering meeting. Name the system where a person in the EU met your AI this week and did not know it. If the room needs a week to answer, the calendar in your board pack is not the constraint. Nobody owns the surfaces where your AI is already meeting people.