Organization A treats August 2, 2026 as the EU AI Act high-risk compliance deadline. Its compliance team classifies every AI system against Annex III, builds a risk management system under Article 9, drafts technical documentation per Annex IV, and schedules conformity assessment in Q1 2026. Organization B reads about the Digital Omnibus, assumes a 16-month extension to December 2027, and puts the project on next year’s budget.
The Digital Omnibus provisional agreement of May 7, 2026 must still be formally adopted before it becomes law. Until then, August 2, 2026 remains the legally operative deadline for Annex III high-risk AI systems under current law. EUR 15 million or 3% of global turnover: the penalty for non-compliance with high-risk obligations under Articles 16, 22, and related provisions [EU AI Act Art. 99(4)]. Violations of Article 5 prohibited practices carry a higher ceiling of EUR 35 million or 7% [EU AI Act Art. 99(3)].
Organization B is betting its compliance program on a legislative outcome it does not control. If formal adoption is delayed, August 2, 2026 remains binding. No scenario exists where starting early produces a worse outcome.
Fourteen requirements span Articles 8 through 49. The critical sequencing insight: classification (Phase 1) determines which of the remaining 13 obligations apply, and getting classification wrong means building controls for a risk tier that does not match your system. The organizations completing Phase 1 now will have Phase 4 done before either deadline arrives.
EU AI Act high-risk compliance requires providers to complete 14 obligations by August 2, 2026 under current law: risk management, data governance, technical documentation, logging, transparency, human oversight, accuracy safeguards, quality management, conformity assessment, CE marking, and EU database registration across Articles 9-49 [EU AI Act]. Deployers carry independent obligations under Article 26.
How Does the EU AI Act Classify High-Risk AI Systems?
EU AI Act high-risk compliance begins with classification: Article 6 of the EU AI Act regulation defines two pathways into the high-risk category, four exemptions out, and one override blocking all exemptions [EU AI Act Art. 6]. The classification decision determines whether the remaining thirteen requirements apply to your organization at all. Get this step wrong and every downstream investment targets the wrong risk tier.
Pathway 1: Safety Components Under Annex I Product Legislation
AI systems acting as safety components of products covered by 30+ EU harmonisation directives fall into Pathway 1 [EU AI Act Art. 6(1), Annex I]. Machinery, medical devices, vehicles, civil aviation, radio equipment, lifts, pressure equipment, and marine equipment: if the product already requires third-party conformity assessment, the AI component inherits the high-risk classification under the EU AI Act. The existing notified body relationship carries over.
Pathway 2: The Eight Annex III Use-Case Areas
Pathway 2 covers AI systems deployed in eight areas where decisions affect natural persons [EU AI Act Art. 6(2), Annex III].
| Annex III Area | Scope | Example |
|---|---|---|
| 1. Biometrics | Remote identification, emotion recognition, categorization | Facial recognition for building access |
| 2. Critical Infrastructure | Safety components of digital infrastructure, water, gas, heating, electricity | AI managing power grid load balancing |
| 3. Education | Admissions, outcome assessment, learning monitoring | Automated exam grading systems |
| 4. Employment | Recruitment, selection, contract decisions, performance monitoring | Resume screening algorithms |
| 5. Essential Services | Credit scoring, insurance pricing, emergency dispatch | Creditworthiness assessment models |
| 6. Law Enforcement | Risk assessment, polygraphs, evidence evaluation, profiling | Predictive policing systems |
| 7. Migration | Asylum application assessment, border surveillance, risk detection | Automated visa processing tools |
| 8. Justice | Case research, law interpretation, dispute resolution support | Sentencing recommendation systems |
The Four Exemptions and the Profiling Override
Article 6(3) provides four exemptions from high-risk classification: narrow procedural tasks, improving a completed human activity, detecting decision patterns without replacing human judgment, and preparatory tasks for assessment [EU AI Act Art. 6(3)]. Each exemption requires documented rationale demonstrating the system does not pose a significant risk to health, safety, or fundamental rights.
The override is the provision competitors miss. AI systems performing profiling of natural persons are always high-risk. No exemptions apply [EU AI Act Art. 6(3) final paragraph].
A recruitment tool scoring candidates against behavioral profiles is high-risk regardless of whether it falls within a listed exemption category. The European Commission missed its February 2, 2026 deadline to publish practical classification guidance [IAPP 2026]. Organizations must self-classify without official examples.
The audit fix. (1) Inventory every AI system in the organization. For each system, document: intended purpose, Annex III area (if applicable), profiling status, and Article 6(3) exemption applicability with written rationale. (2) Flag every system performing profiling of natural persons as high-risk. No exemptions.
(3) Store the classification register alongside AI risk management documentation. Update it when new systems deploy or existing systems change purpose.
Articles 9 through 15 define the requirements classified systems must meet.
Core Technical Requirements for EU AI Act High-Risk Compliance (Articles 9-15)
Seven articles define what every high-risk AI system must deliver, and each article prescribes specific deliverables that become audit artifacts during conformity assessment [EU AI Act Arts. 9-15]. These are not aspirational principles. Risk management (Art. 9) feeds every downstream requirement: data governance decisions depend on identified risks, documentation reflects mitigations, logging captures risk-relevant events, and accuracy metrics tie to risk thresholds.
Risk Management and Data Governance (Articles 9-10)
Article 9 demands a continuous, iterative, documented risk management process running the entire system lifecycle [EU AI Act Art. 9]. Four statutory components form the backbone per Article 9(2): (a) identification and analysis of known and foreseeable risks, (b) estimation and evaluation of those risks, (c) adoption of risk management measures, and (d) post-market evaluation procedures. Article 9(8) adds a testing requirement against predefined performance thresholds. Article 9(9) adds a separate obligation to give explicit consideration to risks for vulnerable groups, including persons under 18. The system is not a one-time assessment. It updates with every system modification, data change, and deployment context shift.
Article 10 mandates data quality criteria: training, validation, and testing datasets must be relevant, representative, and as error-free as possible [EU AI Act Art. 10]. Bias examination is required before and during deployment. The critical nuance competitors overlook: Article 10(5) permits processing of special category personal data (race, ethnicity, political opinions, health status) ONLY for bias detection and correction, under strict safeguards including pseudonymization and data minimization [EU AI Act Art. 10(5)]. This is not a general GDPR exemption. It is a narrowly scoped provision for fairness validation.
Documentation, Logging, and Transparency (Articles 11-13)
Technical documentation per Annex IV must be completed before market placement [EU AI Act Art. 11, Annex IV]. Nine mandatory sections cover: general system description, development process, training methodology, testing procedures, risk management outcomes, and post-market monitoring plans. Article 12 requires automatic logging capability for traceability throughout the system’s lifetime [EU AI Act Art. 12]. For biometric identification systems: logs must capture date, time, reference database queried, input data, and identification results for every use.
Article 13 requires instructions for use covering system capabilities, limitations, declared accuracy levels, maintenance requirements, and foreseeable misuse scenarios [EU AI Act Art. 13]. The transparency obligation extends to deployers: they must understand what the system does, what it does not do, and under what conditions it fails.
Human Oversight and Technical Safeguards (Articles 14-15)
Article 14 requires human-machine interface tools enabling the overseer to understand capabilities, monitor operation, interpret output, override or reverse decisions, and stop the system [EU AI Act Art. 14]. These five oversight capabilities are specified in Article 14(4). For biometric identification systems listed in Annex III point 1(a): Article 14(5) mandates dual-person verification. Two natural persons must independently verify biometric identification results before any action is taken [EU AI Act Art. 14(5)]. An exception applies where Union or national law deems the two-person requirement disproportionate for law enforcement, migration, border control, or asylum applications. This provision appears in almost no competitor checklist.
Article 15 requires declared accuracy metrics with statistical confidence intervals, resilience against data poisoning, model poisoning, adversarial examples, and confidentiality breaches [EU AI Act Art. 15]. Continuously learning systems must address feedback loop risks preventing output degradation over time.
- Build risk management system covering four statutory components (Art. 9(2)) plus testing thresholds (Art. 9(8)) and vulnerable-groups consideration (Art. 9(9))
- Document data governance criteria: relevance, representativeness, error management (Art. 10)
- Complete nine-section Annex IV technical documentation before market placement (Art. 11)
- Implement automatic logging with traceability retention (Art. 12)
- Draft instructions for use covering capabilities, limitations, and foreseeable misuse (Art. 13)
- Build human oversight interfaces with five Article 14(4) capabilities: understand, monitor, recognize bias, interpret, override/stop
- Implement dual-person biometric verification for Annex III point 1(a) systems per Article 14(5)
- Declare accuracy metrics and implement resilience against adversarial attacks (Art. 15)
The audit fix. (1) Build the risk management system first (Art. 9). It feeds every downstream requirement. Start by listing every reasonably foreseeable risk to health, safety, or fundamental rights. (2) Map each identified risk to a mitigation control. Test each control against predefined performance thresholds.
(3) For biometric systems, implement dual-person verification per Article 14(5) and per-use logging (Art. 12). (4) Document everything in Annex IV format before engaging a conformity assessor. Link technical documentation to deployer obligations under Article 26 for the operational side of these requirements.
Organizational requirements assign responsibility for building and operating the system.
Provider and Deployer Obligations for EU AI Act High-Risk Compliance
The EU AI Act splits compliance duties between providers who build the system and deployers who operate it, and each role carries distinct obligations, distinct penalties, and distinct documentation requirements [EU AI Act Art. 16, Art. 26]. Dual-role status is common: organizations customizing third-party AI for their own use become both provider and deployer, stacking obligations from both checklists.
Provider Obligations: QMS, Post-Market Monitoring, Incident Reporting
Article 16 lists 12 provider obligations spanning design through decommissioning [EU AI Act Art. 16]. Article 17 requires a documented quality management system (QMS) with 13 specific elements: regulatory compliance strategy, design and development procedures, quality control and assurance procedures, examination and validation protocols, technical specifications, data management, risk management integration per Article 9, post-market monitoring per Article 72, incident reporting per Article 73, communication procedures, record-keeping systems, resource management, and an accountability framework [EU AI Act Art. 17].
The financial institution exception is a provision competitors miss entirely. Article 17(3) provides that organizations subject to EU financial services internal governance rules satisfy QMS requirements automatically, except for three specific elements: risk management (Art. 9), post-market monitoring (Art. 72), and serious incident reporting (Art. 73) [EU AI Act Art. 17(3)]. Post-market monitoring requires active, systematic collection of performance data throughout the system’s lifetime [EU AI Act Art. 72]. Serious incident reporting carries a 15-day deadline from the date of awareness, not from occurrence [EU AI Act Art. 73(2)].
The harmonised standard for QMS (prEN 18286) entered public enquiry on October 30, 2025, but will not finalize until Q4 2026 at earliest [CMS LawNow 2025]. No harmonised standard for Article 17 exists as of mid-2026.
Deployer Obligations (Article 26)
Article 26 sets out deployer obligations [EU AI Act Art. 26]. Key requirements: use the system according to provider instructions, assign qualified human overseers with documented competence and authority, retain automatically generated logs for a minimum of six months, inform affected workers and their representatives before deploying AI in the workplace, and suspend use immediately upon detecting a risk. Deployers controlling input data must confirm data relevance and representativeness.
When Deployers Must Conduct a Fundamental Rights Impact Assessment
Four categories of deployers must complete a Fundamental Rights Impact Assessment (FRIA) before first use [EU AI Act Art. 27]. Public bodies, private entities providing public services, deployers of credit-scoring AI, and deployers of life or health insurance risk-assessment AI: each must document affected populations, specific harm risks, human oversight measures, and complaint mechanisms. FRIA results must be shared with the market surveillance authority before the system goes live.
| Obligation | Provider Duty | Deployer Duty |
|---|---|---|
| Risk Management | Build and maintain system (Art. 9) | Monitor for risks during use (Art. 26) |
| Human Oversight | Design interface tools per Art. 14(4); dual-person verification for biometrics per Art. 14(5) | Assign qualified overseers (Art. 26) |
| Logging | Implement automatic logging (Art. 12) | Retain logs minimum 6 months (Art. 26) |
| Incident Reporting | Report within 15 days of awareness (Art. 73(2)) | Report risks to provider (Art. 26) |
| Documentation | Complete Annex IV (Art. 11); retain 10 years (Art. 18) | FRIA if applicable (Art. 27) |
The audit fix. (1) Determine whether your organization is a provider, deployer, or both. Dual-role is common with customized AI. (2) For providers: start the Article 17 QMS build now. Use ISO 42001 as a structural foundation (40-50% coverage overlap with EU AI Act requirements).
(3) For deployers: assign human oversight personnel with documented competence, training, and authority. Set up a six-month log retention pipeline per Article 26. (4) If your organization falls into one of the four FRIA trigger categories, schedule the assessment before the system goes live.
Market access requires proving compliance through conformity assessment, CE marking, and EU database registration.
How Do Conformity Assessment, CE Marking, and EU Database Registration Work?
No high-risk AI system enters the EU market without passing conformity assessment, affixing CE marking, and registering in the EU database [EU AI Act Art. 43, Art. 48, Art. 49]. The absence of harmonised standards as of mid-2026 is the single biggest practical obstacle to August 2026 readiness. The medical device regulation (MDR) precedent is instructive: a large share of manufacturers failed to complete certification by the MDR deadline, and notified body queue times exceeded 18 months.
Two Conformity Assessment Pathways (Article 43)
Path A (Annex VI, internal control): the provider self-verifies compliance when harmonised standards are fully applied [EU AI Act Art. 43, Annex VI]. This pathway is available for Annex III categories 2-8 and, when harmonised standards are fully applied, for Category 1 biometric systems as well. Path B (Annex VII, notified body): a third-party auditor examines the QMS and technical documentation. For biometric systems (Category 1), Path B is required when harmonised standards have not been applied, when the provider deviates from available standards, or when common specifications exist but were not followed. Because no harmonised AI Act standards have been published as of mid-2026, biometric providers currently face Path B by default until harmonised standards become available and are applied. For law enforcement biometrics, the market surveillance authority acts as the notified body rather than a private conformity assessment body.
Conformity certificates are valid for four years, renewable for another four after re-assessment [EU AI Act Art. 44]. Substantial modifications to the system trigger a new assessment cycle. The four-year validity creates a planning horizon: budget for re-assessment starting in year three.
CE Marking and EU Database Registration (Articles 48-49)
CE marking must be visible, legible, and indelible on the product or its packaging [EU AI Act Art. 48]. For digital-only systems: a machine-readable code is acceptable.
EU database registration is required before market placement [EU AI Act Art. 49]. Providers register themselves and each system. Public authority deployers register their specific use case. Areas 1 (biometrics for law enforcement), 6 (law enforcement), and 7 (migration) enter a restricted, non-public database section [EU AI Act Art. 71].
The Standards Gap: No Harmonised Standards Until Q4 2026
CEN and CENELEC missed their 2025 deadline for harmonised technical standards. The draft QMS standard prEN 18286 entered public enquiry on October 30, 2025, with a publication target of Q4 2026 at earliest [CMS LawNow 2025]. Until harmonised standards are adopted and published in the Official Journal of the EU, providers have no presumption of conformity [EU AI Act Art. 40]. This is the gap the Digital Omnibus “stop-the-clock” mechanism attempts to address.
Without harmonised standards, providers cannot claim presumption of conformity and notified bodies have nothing to audit against. Organizations building QMS on ISO 42001 now create the closest available proxy for compliance evidence.
No harmonised standard means no presumption of conformity. The conformity assessor has no official benchmark, and the provider has no safe harbor. ISO 42001 fills that gap today: organizations that build their QMS on it now will have a structural foundation ready to align when prEN 18286 finalizes.
The audit fix. (1) Contact notified bodies now. If your system requires third-party assessment (biometrics where harmonised standards are unavailable, or any system where standards have not been applied), begin the engagement process immediately. Capacity is limited. (2) Build your QMS using ISO 42001 + prEN 18286 draft structure as a scaffold.
(3) When the harmonised standard finalizes, map existing documentation to official requirements. (4) Budget for the 4-year conformity certificate renewal cycle. Begin re-assessment planning in year three. Link to EU AI Act penalties and enforcement for the consequences of missing the deadline.
The standards gap is the operational reality. The Digital Omnibus proposes to address it. The question is whether formal adoption occurs before August 2, 2026.
Will the August 2026 Deadline Move? The Digital Omnibus Explained
The European Commission proposed the Digital Omnibus on November 19, 2025, and its AI provisions include a “stop-the-clock” mechanism linking enforcement timing to the availability of compliance support measures [OneTrust 2025]. On May 7, 2026, the Council of the EU and the European Parliament reached a provisional agreement on the package. If formally adopted, Annex III high-risk obligations shift from August 2, 2026 to a backstop of December 2, 2027. No scenario exists where starting early produces a worse outcome.
How the Stop-the-Clock Mechanism Works
High-risk obligations do not apply until the Commission confirms adequate compliance support exists: harmonised standards, guidelines, and common specifications [Morrison Foerster 2025]. Once confirmed, six months for Annex III systems, twelve months for Annex I systems. Backstop dates apply regardless of confirmation: Annex III obligations take effect December 2, 2027. Annex I obligations take effect August 2, 2028. Both backstop dates apply only once the Omnibus is formally adopted.
Additional provisions favor smaller organizations. If the Digital Omnibus provisional agreement is formally adopted, Small Mid-Caps (up to 750 employees or EUR 150 million turnover) would receive simplified documentation requirements [IAPP 2025]. Legacy systems already on the market before rules apply would remain without new certification if no significant design changes are made. Both provisions depend on formal Omnibus adoption.
Legislative Status as of May 2026
The provisional agreement of May 7, 2026 must be formally endorsed and adopted by the co-legislators before it becomes law. Until formal adoption, August 2, 2026 remains the legally enforceable deadline for Annex III high-risk AI systems. Organizations should monitor the adoption process and proceed against August 2026.
The audit fix. (1) Build your compliance program on the August 2, 2026 timeline. If formal Omnibus adoption confirms the extension, treat the extra months as buffer for testing and refinement, not as permission to delay. (2) Map your project to five phases: inventory and classify by April 2026, technical requirements build through June 2026, QMS documentation by July 2026, conformity assessment engagement by August 2026, CE marking and registration upon completion. (3) If the Omnibus grants Small Mid-Cap status to your organization, evaluate the simplified documentation provisions and apply them to reduce documentation burden.
One question remains unresolved: does the August 2026 deadline hold before formal Omnibus adoption?
The EU AI Act high-risk compliance program has 14 distinct requirements, no harmonised standards, and a deadline the Digital Omnibus has provisionally but not yet formally moved. Organizations treating August 2, 2026 as a planning assumption rather than a hard deadline will face enforcement actions, not the organizations who started too early. Start with Article 6 classification. Build the Article 9 risk management system. Everything else follows from those two foundations.
Frequently Asked Questions
What does EU AI Act high-risk compliance require?
EU AI Act high-risk compliance requires providers to implement 14 obligations across Articles 8-49: risk management (Art. 9), data governance (Art. 10), technical documentation (Annex IV), automatic logging (Art. 12), transparency (Art. 13), human oversight (Art. 14), accuracy and cybersecurity safeguards (Art. 15), a quality management system (Art. 17), conformity assessment (Art. 43), CE marking (Art. 48), and EU database registration (Art. 49) [EU AI Act]. Deployers carry independent obligations under Article 26, including six-month log retention and human oversight assignment.
When is the EU AI Act high-risk deadline?
August 2, 2026 is the legally binding deadline for Annex III high-risk AI system obligations under current law [EU AI Act Art. 113]. The Digital Omnibus provisional agreement of May 7, 2026 would extend this to a backstop of December 2, 2027 once formally adopted. Compliance planning should target August 2026.
How do you determine if an AI system is high-risk under the EU AI Act?
Article 6 defines two pathways: the system acts as a safety component under Annex I product legislation requiring third-party assessment, or the system falls within one of eight Annex III use-case areas covering biometrics, critical infrastructure, education, employment, essential services, law enforcement, migration, and justice [EU AI Act Art. 6]. Systems performing profiling of natural persons are always high-risk with no exemptions.
What are the penalties for EU AI Act non-compliance?
Penalties follow a three-tier structure: EUR 35 million or 7% of global turnover for Article 5 prohibited practices [Art. 99(3)], EUR 15 million or 3% for violations of high-risk obligations under Articles 16, 22, and related provisions [Art. 99(4)], and EUR 7.5 million or 1% for providing incorrect, incomplete, or misleading information to authorities [Art. 99(5)]. For cooperative violations where the entity makes substantial efforts to comply, national authorities may take cooperation and mitigating factors into account within the bounds of their enforcement discretion [Art. 99(6)]. SMEs and startups receive the lower of the fixed amount or the percentage, not the higher [EU AI Act Art. 99].
Does ISO 42001 satisfy EU AI Act requirements?
ISO 42001 covers approximately 40-50% of EU AI Act high-risk requirements, providing strong alignment for risk management (Art. 9), data governance (Art. 10), documentation (Art. 11), transparency (Art. 13), and QMS structure (Art. 17) [ISO 42001:2023]. It does not cover conformity assessment (Art. 43), CE marking (Art. 48), EU database registration (Art. 49), incident reporting (Art. 73), or the Fundamental Rights Impact Assessment (Art. 27).
What is the Digital Omnibus stop-the-clock provision?
The stop-the-clock mechanism delays high-risk obligations until the European Commission confirms adequate compliance support exists, including harmonised standards, guidelines, and common specifications [Morrison Foerster 2025]. Once confirmed, obligations apply after six months for Annex III systems. A backstop date of December 2, 2027 applies regardless of Commission confirmation, once the Omnibus is formally adopted. The May 7, 2026 provisional agreement must still complete formal adoption before these dates become law.
How does conformity assessment work for high-risk AI systems?
Two pathways exist under Article 43: internal control (Annex VI) allows provider self-verification when harmonised standards are fully applied, and third-party assessment by a notified body (Annex VII) is required when standards are unavailable, not applied, or only partially applied [EU AI Act Art. 43]. For biometric systems (Category 1), internal control is available when standards are fully applied; because no harmonised AI Act standards exist as of mid-2026, biometric providers currently require notified body assessment. Certificates are valid for four years, renewable after re-assessment. Substantial modifications require a new cycle.
What is the Fundamental Rights Impact Assessment under the EU AI Act?
Public bodies, private entities providing public services, deployers of credit-scoring AI, and deployers of life or health insurance AI must complete a Fundamental Rights Impact Assessment before first use of a high-risk AI system [EU AI Act Art. 27]. The assessment documents affected populations, specific harm risks, human oversight measures, and complaint mechanisms. Results must be shared with the market surveillance authority.
Subscribe to The Authority Brief for next week’s analysis.
