Two dates circulate in European boardrooms, and they tell opposite stories. On 2 August 2026, part of the AI Act became enforceable. On 2 December 2027, another part will, after the Digital Omnibus package pushed it back. Many teams remembered the delay and filed the whole subject away with it.
The AI Act (Regulation EU 2024/1689) is the world's first binding framework for artificial intelligence. It sorts AI systems by risk level and attaches a distinct set of obligations to each. The AI strand of the Digital Omnibus package, Regulation (EU) 2026/1744, in force since 27 July 2026, reshuffles that calendar and changes several substantive provisions.
Telling what the Omnibus moved from what it left in place takes ten minutes of board time, and it avoids two symmetrical mistakes: mobilising teams against a deadline that has slipped, or ignoring a duty that already bites.
What applies since 2 August 2026?
Since 2 August 2026, Article 50 of the AI Act imposes four transparency duties: tell users they are interacting with an AI system, mark synthetic content in a machine-readable format, disclose manipulated content, and inform people subjected to emotion recognition.
Who carries each transparency duty
The detail matters, because the burden does not always sit with the same actor in the chain:
- Direct interaction with an AI system. The provider must ensure the person knows they are talking to a machine, unless it is obvious to a reasonably well-informed user. In practice: any conversational agent facing a customer, a candidate or an employee.
- Synthetic content. The provider must mark audio, image, video and text outputs in a machine-readable format, with a technical solution that is effective and interoperable. Systems placed on the market before 2 August 2026 have until 2 December 2026 to comply.
- Manipulated content and text on matters of public interest. Here the deploying company carries the duty to disclose that content was artificially generated or altered. One exception rests on review: text that went through human review or editorial control, with a person holding editorial responsibility for its publication, falls outside the scope. Content generated before 2 August 2026 does not need to be labelled retroactively.
- Emotion recognition and biometric categorisation. The deployer informs the people exposed, and GDPR obligations continue to apply on top.
The regulation requires the disclosure to be clear and distinguishable, and given no later than the first interaction. A line buried in terms and conditions does not meet that bar, and neither does a banner that appears after three exchanges with the agent.
How the Omnibus amended Article 50
The Omnibus left the substance of these duties intact and amended Article 50 in two places. Paragraph 7 is rewritten: the Commission, after hearing the AI Board, assesses whether signing up to codes of practice is enough for marking (paragraph 2) and labelling (paragraph 4), and may impose common rules if it finds a code inadequate. A paragraph 4 added to Article 111 creates the 2 December 2026 deadline, reserved for providers of generative systems already on the market. The rest of Article 50 has applied since 2 August 2026, with no grace period.
What does a customer-facing chatbot have to show?
The question arises as soon as a company puts an agent in front of its customers. A conversational agent in customer service has to state what it is on the first screen, in wording the reader can parse.
The Commission's code of practice and guidelines
The regulation itself prescribes no standard formula. On 10 June 2026 the Commission published the final version of a code of practice on marking and labelling AI-generated content, followed by transparency guidelines on 20 July 2026. Signing up to the code stays voluntary, and the adequacy opinion of 9 July makes it the recognised route for paragraphs 2, 4 and 5 of Article 50 (marking, labelling, and how the information is given), with a caveat from the Commission: adherence does not constitute conclusive evidence of compliance. For the disclosure a chatbot has to make, the guidelines do most of the work of clarifying who owes what.
The code splits into two parts. For providers: multi-layered marking, preservation of marks already carried by input data, detectability of outputs. For deployers: labelling of deepfakes and of certain texts, with optional EU icons which, when used, have to be visible from first exposure and embedded in the content. A company that picks another route keeps the right to do so, and carries the burden of showing its own measures reach the same result.
What a compliant disclosure has to guarantee
What separates a disclosure that holds from a cosmetic one comes down mostly to visibility: the message has to be read before the first exchange, the latest point Article 50 allows. According to the Commission guidelines of 20 July 2026 (point 40), a single, prominent notification is likely to suffice in most cases, but periodic reminders are likely to be needed in riskier contexts such as vulnerable users, sustained or immersive interactions, financial, legal or health advice, and complaints handling. A customer service chatbot that handles complaints should therefore restate what it is as the conversation goes on.
In practice, the product team settles the wording up front, where it sits (at the top of the conversation window rather than in a footer), whether it reappears when a conversation resumes, and how a voice channel handles it, with the announcement at the start of the call. The information also has to meet the applicable accessibility requirements.
Then comes proof, which a supervisory authority can test: being able to demonstrate, six months after an incident, what was displayed on a given date and for which deployed version of the system. That means versioning your disclosure wording the same way you version code.
Internal assistants are in scope too
The same reasoning covers internal use. An assistant that chats with candidates or answers employee HR questions sits inside the scope, with no public exposure at all. A tool that pre-screens CVs for recruiters falls under recruitment in Annex III, and therefore under the high-risk regime from December 2027.
Were the high-risk obligations really postponed?
Yes, and that is where the confusion comes from. The Digital Omnibus package moves the obligations for stand-alone Annex III high-risk systems, covering recruitment, credit scoring, education, biometrics and border control, from 2 August 2026 to 2 December 2027. AI embedded in already-regulated products under Annex I (medical devices, toys, lifts, radio equipment) moves from 2 August 2027 to 2 August 2028. The text also narrows the high-risk qualification for certain assistance functions unrelated to safety.
What the Omnibus adds or reinforces
The supervisory architecture comes out of the package reinforced. The EU AI Office can now request information, run inspections, find non-compliance and impose fines, for AI systems built on a general-purpose model by the same provider and for those integrated into very large online platforms. The Omnibus also adds two prohibited practices to Article 5, covering AI-generated intimate imagery of an identifiable person without consent and child sexual abuse material, applicable on 2 December 2026. It pushes back to 2 August 2027 the duty for each member state to run at least one operational regulatory sandbox.
Part of the text lightens the load. The AI literacy duty in Article 4, in force since February 2025, now asks providers and deployers to support their staff's AI literacy, without requiring them to guarantee any given level. The simplified technical documentation form, already provided for SMEs, is extended to small mid-cap enterprises (SMCs).
The fine ceilings set by the AI Act
The penalty regime is tiered, and the Omnibus kept its amounts. Prohibited practices under Article 5 face fines of up to 35 million euros or 7 % of worldwide annual turnover. A breach of Article 50 sits one tier down: 15 million euros or 3 %. Supplying incorrect information to authorities caps out at 7.5 million euros or 1 %. For SMCs, the two lower tiers now apply whichever ceiling is lower, a rule SMEs already had.
Who enforces: national authorities, France included
Which leaves the question of who imposes those fines. Enforcement rests with the authorities each member state designates under Article 70, and the designation is still unfinished in several of them. The Commission's list of single points of contact, updated on 7 September 2026, flags Germany, France, Spain and the Netherlands among the member states whose designation awaits final adoption.
France shows how long the process can take. The digital section of the DDADUE bill, the vehicle carrying the designation, was adopted at first reading in the Senate on 18 February 2026 and, as of mid-September 2026, was still at committee stage in the National Assembly. The scheme put forward by the DGCCRF and the DGE, and endorsed at first reading, spreads supervision across some fifteen sectoral authorities, with the DGCCRF holding the single point of contact required by Article 70 and the CNIL acting as a pivot on data processing and biometrics.
What should companies do with the reprieve on high risk?
The temptation is familiar: shelve the file, reopen it in autumn 2027. That plan tends to backfire, and the research published on AI projects in production shows why.
What published studies say about keeping AI systems under control
According to Deloitte (State of AI in the Enterprise, 2026 edition), 21 % of companies have mature governance of their agents. MIT (2025) finds that 95 % of generative AI pilots have no measurable effect on the P&L. S&P Global (2025) finds that 46 % of AI proofs of concept are abandoned before production. Gartner (2025) expects over 40 % of agentic AI projects to be cancelled by the end of 2027. The figures point to the same blind spot: organisations know how to launch systems, far less how to keep them running under control.
The building blocks of high-risk compliance
High-risk compliance does not get built in a quarter. It assumes a system inventory, a classification you can defend to a third party, technical documentation, human oversight, usable logging and periodic review. Every one of those bricks has to be laid on systems already in production, by teams with other commitments on their roadmap. The reprieve therefore shifts the workload onto a tighter calendar, at the point where those same teams will be absorbing the next wave of deployments.
Each brick maps to an article of the regulation: risk management (Article 9), technical documentation (Article 11), logging (Article 12), human oversight (Article 14) and, for the company using the system, the deployer obligations in Article 26. Take an HR department as an illustration: a CV-screening tool classified as high-risk means being able to retrieve, for every rejected application, the model version, the criteria applied and the person who approved the recommendation. A management standard such as ISO 42001 gives that evidence a durable frame, whereas the regulation sets requirements without prescribing a format.
What is left to do is handle transparency now, since it is enforceable, and spend the months gained mapping what already exists.
How do you know whether your company is in scope?
A first board discussion can start from the following questions, none of which calls for legal expertise:
- How many AI systems are running here today, and who keeps the list? The count can be surprising, because adoption came through business units as much as through IT.
- Which ones are exposed to someone outside the team that runs them? Those are the ones Article 50 targets first, and their duties bite today.
- Do we publish AI-generated content? If so, on which channels, and with what human review actually recorded.
- Which ones will fall under Annex III in December 2027? Recruitment and creditworthiness assessment are on the list, two uses to check first.
- Who answers if a regulator asks for the documentation in six months? Naming that owner in the inventory keeps the documentation from going stale.
If these answers come out of a single meeting, the inventory already exists. If they take an internal survey, that inventory is where the work starts. Our AI maturity assessment covers that ground in a few minutes and places the level of governance reached, agentic dimension included.
Who should own AI Act compliance internally?
The default answer points at legal or the DPO. It produces clean files on paper and exposed systems in production, because the decisions that create the risk are taken elsewhere: in the choice of a model, in the scope of action granted to an agent, in the trade-off between full automation and human validation.
The functions that need to be involved
At minimum, the subject pulls in a legal or compliance function that qualifies the systems, an IT function that documents, logs and can replay a decision, and a business owner who stands behind the use cases and the trade-offs attached to them. Governance means writing down who decides what, what an agent may do on its own, what goes back to a human, and how that decision can be found in the logs six months later. That is the principle behind the LOOP™ methodology, and it is also, phrased differently, the documentation the AI Act ends up demanding.
In LOOP™, that split is settled action by action. In the green zone the agent acts alone and everything is logged. In the orange zone a designated person validates before execution. In the red zone the agent stops, documents the case and alerts the owner, who has to decide within four hours. In the black zone the action is blocked and the CISO is alerted.
For a mid-sized company without dedicated AI staff, that steering role can be handed to an outsourced Chief AI Officer, with LOOP™ governance built in.
For the full calendar by risk level and the obligations attached to high-risk systems, our AI Act compliance page sets out the complete set of deadlines.
What the delay does not cover
The Digital Omnibus moved the high-risk calendar, narrowed the qualification of certain assistance functions, added two prohibitions to Article 5 and widened the powers of the AI Office. On transparency, it stopped at a four-month transition for marking by generative systems already on the market. A customer service chatbot that does not declare itself as AI has been in breach since 2 August 2026, on one of the few parts of the regulation a customer can check without expertise. The rest of the calendar leaves room, provided you start by knowing what is running.
The blog's other analyses of the AI Act, ISO 42001 and agent governance are gathered in the AI governance and compliance section.