The Deadline In The Rule

A regulation is not a mood. It is a document with dates attached, and the dates are the part that changes behaviour. Regulation (EU) 2024/1689 entered into force on 1 August 2024. It binds any company that places an AI system on the EU market, puts one into service inside the EU, or whose AI output is used in the EU. Headquarters is not a defence. A team in Ohio running automated screening for an office in Dublin is a deployer inside the scope. The structure is risk based. Four drawers, and every system you own, buy or rent belongs in one of them. Most organisations cannot name the drawer, because most organisations have never listed the systems. Over half of organisations still lack a systematic inventory of the AI they already run. That is the whole problem. You cannot classify what you have never written down.

The first drawer is prohibited. Social scoring, manipulative techniques, emotion recognition in workplaces and schools, real-time remote biometric identification in public spaces outside narrow exceptions, and systems that exploit the vulnerabilities of specific groups. Banned since 2 February 2025, with no grandfathering. You stop, or you are in the top penalty tier. The second drawer is high risk, and it has two halves. Safety components of products covered by Annex I product safety law, where third-party conformity assessment is already required. And standalone systems in the Annex III areas: biometrics, critical infrastructure, education, employment, credit scoring, law enforcement, migration, justice and democratic processes. Annex III applies from 2 August 2026. Annex I applies from 2 August 2027. The third drawer is transparency risk. Chatbots, deepfakes, emotion recognition, biometric categorisation. Those obligations have been active since 2 August 2025. Users get told they are talking to a machine. Synthetic content gets labelled. The fourth drawer is everything else. No specific obligations, voluntary codes of conduct, and the sound of a door closing on everybody who assumed they were standing in this drawer and were not.

Two dates are already behind us. 2 February 2025 banned the prohibited practices and switched on AI literacy duties for providers and deployers. 2 August 2025 brought the general-purpose model obligations, the governance structures, the national competent authorities and the penalty regime. Three dates are ahead. 2 August 2026 for the Annex III high-risk systems, which means Articles 9 to 15 for providers and Article 26 for deployers. 2 August 2027 for the Annex I systems embedded in regulated products: medical devices, vehicles, machinery, toys. 2 August 2030 for legacy public-sector high-risk systems placed on the market before 2026. A buyer reads those dates the way a pilot reads weather. Not as philosophy. As a constraint on what they are allowed to sign. The date is the rule. Everything else is commentary.

Article 99 lays out three tiers. Tier 1, for the prohibited practices, reaches EUR 35,000,000 or 7 per cent of total worldwide annual turnover, whichever is higher. Tier 2, for high-risk non-compliance and systemic-risk failures on general-purpose models, reaches EUR 15,000,000 or 3 per cent. Tier 3, for incorrect, incomplete or misleading information given to a notified body or a national authority, reaches EUR 7,500,000 or 1 per cent. Turnover is measured at group level, not at the level of the subsidiary that did the thing. That single sentence has shaped more compliance programmes than any figure in the table. For SMEs the arithmetic flips. Article 99(6) says the fine is whichever is lower, the percentage or the fixed amount. A company at EUR 2,000,000 of revenue faces a ceiling near EUR 140,000, not EUR 35,000,000. That is the difference between a bad quarter and a closed company. It is not immunity. Article 99(1) requires penalties to weigh the interests of SMEs and their economic viability, and the obligation to stop the practice is identical for everyone. The probability that a competitor writes your AI compliance status into a tender document before August 2026 is not zero. Most boards are pricing it at zero.

A regulation is not a mood. It is a document with dates attached.

For a high-risk system the file is long, and none of it is optional. A risk management process running through the whole lifecycle under Article 9. Data governance under Article 10, with training, validation and testing sets that are relevant, representative and as free of error as the state of the art allows, plus the records that show how the data was prepared. Technical documentation under Article 11 and Annex IV, detailed enough for an authority to assess the system without reverse-engineering it. Automatic logging under Article 12, keeping events, inputs, outputs, decision points and operator identity for at least six months. Transparency and instructions for use under Article 13. Human oversight under Article 14, designed so a named person can intervene, and can interrupt or stop the system. Accuracy, robustness and cybersecurity under Article 15. A quality management system under Article 17. Then the assessment itself. Annex III systems mostly run an internal self-assessment against Annex VI. Annex I systems need a notified body. Each one is registered in the EU database under Article 49 before it goes to market, with its version, its intended purpose and the member states where it will run. Post-market monitoring under Article 72 collects evidence after the launch, not before it. Deployers carry their own list under Article 26: oversight measures with competent people behind them, six months of logs, a fundamental rights impact assessment where the use case demands one, and a duty to inform workers before AI lands inside an employment decision. That last one is a timetable the employer does not control.

Three reporting windows sit inside the Act. A serious incident goes to the national authority and the provider within 15 days. An incident that disrupts critical infrastructure goes within 2 days. An incident involving a death goes within 10 days. Most companies cannot hit those windows today, because an AI incident process does not exist as a process. It exists as a paragraph in a policy nobody has rehearsed. Enforcement is not theoretical. Germany named the Federal Network Agency. France named CNIL. Most member states completed their designations by August 2025. Complaints can come from a customer, a competitor or the authority on its own initiative, and cross-border cases are coordinated through the European AI Office. The pattern is GDPR, and everybody in Europe has already watched that film.

On 19 November 2025 the European Commission proposed a Digital Omnibus on AI. It would move the Annex III obligations to 2 December 2027 and the Annex I obligations to 2 August 2028. It would extend the SME benefits to small mid-caps of up to 499 employees and EUR 100,000,000 of turnover. It would drop the mandatory harmonised post-market monitoring plans, delete database registration for non-high-risk systems used in high-risk areas, move the AI literacy burden to the Commission and member states, and allow special category personal data to be processed for bias detection with safeguards. In March 2026 the European Parliament adopted its position with 569 votes in favour. Trilogue negotiations carried a target for political agreement of 28 April 2026. As of this writing the Omnibus is a proposal. It is not law. Orrick, WilmerHale and DLA Piper each advise planning against 2 August 2026 until the text is formally adopted. If the delay passes before August, the companies that kept working gain a buffer. The ones that stopped preparing, and guessed right, lose nothing. The ones that stopped preparing, and guessed wrong, are standing in the Tier 2 column. A proposal to move a deadline is not a deadline that has moved.

I have built companies in twelve countries and spent sixteen years in the field, from rural Nigeria to Silicon Valley. I have restructured a EUR 75 million industrial group. The pattern in every market is the same. The written rule never arrives in time for anyone to be ready by accident. So here is the rule, in the order the work happens. List every AI system in the business within 30 days, including the ones hiding inside a SaaS contract nobody has read. Classify each one against Annex III and Annex I, and write the reasoning down. Kill anything in the prohibited drawer today. Build the documentation, the logging and the oversight for everything that lands in the high-risk drawer, and start that now rather than in July. The deadline does not negotiate. It decides which contracts you are allowed to sign, which customers will accept your bid, and which of your competitors gets to look conservative and safe while you look exposed. The probability that this stays a problem for somebody else after 2 August 2026 is not zero. It is also not large. The date is already in the clause. The only question left is which side of it carries your name.