The Compliance Bill Was Already Due

The Regulation is (EU) 2024/1689. It entered into force on 1 August 2024. The first enforcement date landed on 2 February 2025 - the prohibited practices list. The second landed on 2 August 2025 - the transparency obligations, the GPAI obligations, the governance bodies, the penalty regime. The third lands on 2 August 2026 - the Annex III high-risk obligations. The fourth lands on 2 August 2027 - the Annex I high-risk obligations. The fifth lands on 2 August 2030 - the legacy public-sector high-risk systems. Most boards are operating against the wrong one. They are planning against a calendar that runs from 2 August 2026 forward. The calendar that should drive the work runs from 2 February 2025 backwards. The prohibited list has been enforceable for seventeen months at the time of writing. That is a compliance gap most companies have not measured. The probability that a company shipping an AI product into the EU since February 2025 has done the underlying compliance work sits below fifteen per cent. The structural reason is the gap between the regulation date and the marketing date. Lawyers draft against the next milestone. Engineers ship against the next release. Nobody owns the bill that is already due.

The prohibited list lives in Article 5 of the regulation. Social scoring of individuals is banned. Manipulative techniques that distort behaviour and cause significant harm are banned. Exploiting vulnerabilities of specific groups - age, disability, social or economic situation - is banned. Emotion recognition in workplaces and in schools is banned. Real-time remote biometric identification in public spaces for law enforcement is banned, with narrow exceptions. Biometric categorisation that infers sensitive attributes - race, political opinion, sexual orientation - is banned. The list has been enforceable since 2 February 2025. The penalty for any violation is Tier 1 - thirty-five million euros or seven per cent of worldwide turnover, whichever is higher. For an SME, the rule is the LOWER of the two, not the higher. A two-million-euro-revenue company faces a maximum fine of one hundred forty thousand euros. A ten-million-euro-revenue company faces seven hundred thousand. None of these are existential on a spreadsheet. They are existential on a runway. The probability of a Tier 1 enforcement action landing on a European SME inside the next twelve months sits below five per cent, because enforcement is still ramping. The probability of an enforcement action inside the next thirty-six months sits above sixty per cent.

The 2 August 2025 list covers transparency obligations, GPAI model obligations, governance structures, and the penalty regime. Chatbots must disclose that the user is interacting with AI. Deepfakes must be labelled. Emotion recognition systems and biometric categorisation must notify affected persons. GPAI providers must produce technical documentation, comply with copyright law, and publish a sufficiently detailed summary of training data. National competent authorities must be operational. Penalty enforcement must be active. Each of these has been on the clock since 2 August 2025. The structural reason most companies missed the date is the absence of a single accountable owner. Engineering did not see it as an engineering task. Legal did not see it as a legal task. Marketing saw it as a marketing task and produced a one-line disclosure nobody updated. The probability of a Tier 2 enforcement action under the transparency or GPAI rules inside the next twelve months sits below ten per cent. The probability inside thirty-six months sits above fifty per cent. The Tier 2 ceiling is fifteen million euros or three per cent of worldwide turnover, whichever is higher - again, LOWER for SMEs. A ten-million-euro company faces three hundred thousand. A hundred-million-euro company faces three million. The fines are large enough to be material. The probability of being caught today is low. The probability of being caught within the life of a venture-cycle funding round is not.

Annex III carries the standalone high-risk obligations. Biometrics. Critical infrastructure. Education and vocational training. Employment, worker management, self-employment. Access to essential private and public services - credit scoring, insurance pricing, emergency services, public benefit eligibility. Law enforcement. Migration, asylum, border control. Administration of justice and democratic processes. Each of those areas is in scope from 2 August 2026. The Digital Omnibus proposal from November 2025 would push the date to December 2027. The proposal is in trilogue. The European Commission, the European Parliament, and the Council of the EU are negotiating. The target political agreement date is 28 April 2026. Until the regulation formally amends, the original date stands. Law firms including Orrick, WilmerHale, and DLA Piper advise treating 2 August 2026 as binding. The compliance work to clear Annex III is the most expensive item on the calendar. Risk management system under Article 9. Data governance under Article 10. Technical documentation under Article 11 and Annex IV. Automatic logging under Article 12 - six-month minimum retention. Transparency under Article 13. Human oversight under Article 14. Accuracy, robustness, cybersecurity under Article 15. Quality management system under Article 17. Conformity assessment under Article 43. EU AI database registration under Article 49. Post-market monitoring under Article 72. Eleven deliverables. Each one is documented. The probability of the Digital Omnibus being adopted on time to push the date sits below thirty per cent. The probability of some delay being agreed sits above seventy per cent. The structural reason is the number of competent authorities that have not finished building the supervisory apparatus required to receive the Annex III filings.

The compliance bill was already due before most companies started writing it.

The compliance work for an Annex III system is not a one-line disclosure. The risk management system under Article 9 has to identify and mitigate risks to health, safety, and fundamental rights across the full lifecycle. The data governance regime under Article 10 has to cover training, validation, and testing datasets - bias detection and mitigation records, data preparation logs, version control, and provenance linking datasets to model versions. The technical documentation under Article 11 plus Annex IV has to cover system architecture, design choices, development process, intended purpose, performance metrics, limitations, and instructions for use. The automatic logging under Article 12 has to capture timestamps, inputs, outputs, decision points, user or operator identity, and event descriptions for a six-month minimum. The human-oversight regime under Article 14 has to define who can intervene, how they intervene, and what the fallback is when they do not. The conformity assessment under Article 43 is internal self-assessment per Annex VI for most Annex III systems. The EU AI database registration under Article 49 has to be filed before the system goes on the market. The post-market monitoring under Article 72 has to be active and systematic after deployment. A realistic SME budget for the full Annex III package sits between three hundred thousand euros and one point five million euros in year one, depending on the system classification and the existing documentation depth. The probability of an SME clearing Annex III inside the eight-week window without external support sits below ten per cent. The probability of clearing it inside twelve months with external support sits above seventy per cent. The structural reason is the volume of deliverables, not the difficulty of any single one.

Tier 1 fines apply to the prohibited practices - thirty-five million euros or seven per cent of worldwide turnover, whichever is higher. Tier 2 fines apply to high-risk non-compliance - fifteen million euros or three per cent of worldwide turnover, whichever is higher. Tier 3 fines apply to incorrect, incomplete, or misleading information to notified bodies or national authorities - seven point five million euros or one per cent of worldwide turnover, whichever is higher. The trap is the Tier 3 rule. A company that files an incomplete technical documentation package to the EU AI database does not face the seven-per-cent penalty. It faces the one-per-cent penalty - the Tier 3 ceiling. That ceiling still costs seven point five million euros at the high end. For an SME, the rule is again the LOWER of the amount or the percentage. A ten-million-euro company faces one hundred thousand on Tier 3. The trap is that the Tier 3 fine is easier to trigger than the Tier 2 fine, because the evidence standard for misleading filing is lower than the evidence standard for non-compliance. The probability of a Tier 3 enforcement action on a documented-but-incomplete filing sits above the probability of a Tier 2 action on a fully-documented-but-non-compliant system. The structural reason is the asymmetry between what is provable from paperwork and what is provable from behaviour.

Article 99(6) of the regulation explicitly addresses SMEs. The fine for an SME is the percentage or the amount, whichever is LOWER. A two-million-euro-revenue company facing a Tier 1 charge does not face thirty-five million euros. It faces the seven per cent - one hundred forty thousand euros. The ceiling flips the burden from company-scale to percentage-scale. The discount is real. It is also insufficient. One hundred forty thousand euros is existential for a seed-stage company. The ceiling does not change the obligation to cease the practice. It does not change the reputational damage of a public enforcement action. It does not change the cost of the legal defence. The probability of an SME using the Article 99(6) discount to negotiate a lower fine inside the next five years sits above fifty per cent. The probability of the discount saving the company from going under sits below fifteen per cent. The structural reason is that fines are never the only cost. The compliance work itself - the eleven deliverables for Annex III - is the same for an SME as for a multinational. The Article 99(6) discount does not buy a discount on the work. It buys a discount on the consequence.

Three conditions decide whether a company is materially exposed. First, the AI-system classification has to be honest. If the system is high-risk under Annex III and the company treats it as limited-risk, the gap is the entire Tier 2 exposure. Second, the compliance work has to be live. Risk management under Article 9, data governance under Article 10, technical documentation under Article 11, automatic logging under Article 12, human oversight under Article 14, conformity assessment under Article 43, EU AI database registration under Article 49 - each one has to be documented before the system goes into service. Third, the EU-list status of any counterparty structures has to be checked. Cross-border payment rails into high-risk-counterparty structures add friction. The honest read: the compliance bill for any AI system that touches the EU market was already due the day the model went into production. The 2 August 2026 date is the next milestone, not the first. The probability of an AI system shipping into the EU since 2 February 2025 that meets every obligation in the regulation sits below ten per cent. The probability of the gap closing inside twelve months sits below twenty per cent. The work is real. The cost is real. The deadline is closer than the marketing calendar says.