Key Takeaways
- Prohibitions and the AI literacy duty have applied since 2 February 2025, and national authorities can enforce them from 2 August 2026.
- Annex III high-risk obligations now apply from 2 December 2027 and Annex I product obligations from 2 August 2028, after Regulation (EU) 2026/1744 postponed both.
- Build the inventory first, including workplace use of tools such as ChatGPT and Microsoft Copilot, because every role and risk-tier decision depends on it.
- Record a provider or deployer determination for each system, since rebranding, substantial modification or a new high-risk purpose can turn a deployer into a provider.
- Reuse existing ISO/IEC 42001, GDPR, NIST AI RMF and DORA controls, and plan conformity assessment, CE marking and EU database registration separately.
A recruitment team switches on a vendor’s CV-screening feature, and the organization is now deploying an AI system intended to “analyse and filter job applications,” a high-risk use under Annex III of the EU AI Act [1]. Organizations often discover such a system only after it is in use. This guide gives compliance leaders an ordered roadmap for EU AI Act compliance, mapped to the deadlines that apply now and the ones that moved in July 2026.
Deadlines That Apply Now and What Is Coming Next
Regulation (EU) 2026/1744, the Digital Omnibus on AI, entered into force on 27 July 2026 and postponed the high-risk obligations, while prohibitions, AI literacy and GPAI obligations kept their original dates [2]. National authorities enforce the literacy duty from 2 August 2026 [3]. The table below lists each milestone and the roadmap steps that should be complete by then [1]:
Pages written before July 2026 still show August 2026 as the high-risk date. Check every date against the consolidated text of 27 July 2026 before it goes into a board report or a project plan.
How to Govern AI Under the EU AI Act in Eight Steps
Each step depends on the one before it. A risk tier means nothing until the system is in the inventory, and obligations cannot be assigned until the organization knows whether it is the provider or the deployer. Steps 1 through 4 address obligations already in force, and Steps 5 through 8 build toward the high-risk dates in 2027 and 2028.
Step 1: Build a Complete AI Inventory of First-Party, Third-Party and Embedded AI
The Act attaches obligations to specific systems and roles, so an unrecorded system cannot be governed. The AI inventory should cover every AI system the organization builds, buys, embeds in a product or lets employees use, including coding agents.
Workplace use of ChatGPT or Microsoft Copilot belongs in it. The Act defines a deployer as a body “using an AI system under its authority except where the AI system is used in the course of a personal non-professional activity” [1]. Most such use carries light obligations, mainly AI literacy and Article 50 disclosure, but a hiring manager who ranks CVs in a chatbot may be using it for an Annex III purpose.
Shadow AI makes this harder. A Gartner survey of 302 cybersecurity leaders, run from March to May 2025, found that 69% of organizations suspected or had evidence of employees using prohibited public GenAI [4].
Each entry should record an owner and intended purpose, lifecycle status (live, testing or retired), supplier, role, risk tier, EU nexus and the personal data processed.
Step 2: Assign a Role to Each System as Provider, Deployer or Downstream Provider
The role determines the obligations. A provider develops an AI system, or has one developed, and places it on the market or puts it into service “under its own name or trademark, whether for payment or free of charge” [1]. A deployer uses a system under its authority. For high-risk systems the loads differ sharply under Articles 16 and 26 [1]:
Under Article 25(1), a deployer becomes the provider when it puts its name or trademark on a high-risk system, makes a substantial modification that keeps the system high-risk, or changes a system’s intended purpose so that it becomes high-risk [1]. Record the reasoning for each role decision in the inventory entry.
Step 3: Classify Each System by Risk Tier and Screen for Prohibited Practices
Prohibited practices come first, because they have applied since February 2025 and carry the highest fines. The remaining systems then fall into one tier each:
- Prohibited (Article 5): Practices such as harmful manipulation, social scoring, untargeted facial-image scraping and workplace emotion recognition [1].
- High-risk (Article 6, Annexes I and III): Safety components of regulated products, plus stand-alone systems in eight Annex III areas: biometrics, critical infrastructure, education, employment, essential services, law enforcement, migration, and justice and democratic processes.
- Transparency (Article 50): Systems that interact with people or generate synthetic content.
- Minimal risk: Everything else, which still falls under the Article 4 literacy duty.
Annex III names concrete uses, such as point 4(a) for systems that “analyse and filter job applications” and point 5(b) for credit scoring [1]. Article 6(3) exempts Annex III systems that pose no significant risk, but never systems that profile natural persons, and a provider relying on it must document the assessment and register the system under Article 49(2).
The Commission’s classification guidelines, legally due by 2 February 2026, remain in draft. The draft was published on 19 May 2026 and the consultation closed on 23 July 2026 [5], so classify against the Act’s text now and revisit borderline calls when final guidance publishes.
Step 4: Stand Up a Governance Body and an Article 4 AI Literacy Program
Obligations need named owners, so the next step is a cross-functional body of legal, compliance, data protection, security, engineering and business owners, with authority to approve, block and retire AI systems. Many organizations start from a weak base. ISACA’s 2026 AI Pulse Poll found that only 38% of organizations “have a formal, comprehensive AI policy,” up from 28% in 2025 [6].
Regulation 2026/1744 turned the literacy duty into one to “take measures to support the development of AI literacy,” which “does not require providers or deployers to guarantee any specific level of AI literacy of any individual” [2]. Deployers of high-risk systems still have to train oversight staff [3], and completion records are the evidence national authorities will ask for.
Step 5: Implement High-Risk Obligations From Risk Management to Conformity Assessment
Providers of high-risk systems carry a chain of requirements that ends in conformity assessment, and each link produces documentation an authority can request [1]:
- Risk management (Article 9): A lifecycle process, with testing against “prior defined metrics and probabilistic thresholds” before market placement.
- Data governance (Article 10): Relevant, representative and, as far as possible, error-free data sets.
- Documentation and records (Articles 11, 12 and 18): Technical documentation kept for 10 years, and automatic event logging.
- Transparency and oversight (Articles 13 and 14): Instructions for use, and oversight staff who can override or halt the system “through a ‘stop’ button or a similar procedure.”
- Accuracy and cybersecurity (Article 15): Declared accuracy levels and resilience against manipulation.
- Quality management and monitoring (Articles 17 and 72): Written QMS policies and a post-market monitoring plan.
Providers of Annex III points 2 to 8 systems, including employment and credit, follow internal control under Annex VI with no notified body. Point 1 biometric systems need a notified body unless harmonised standards or common specifications were applied [1]. The provider then signs the EU declaration of conformity, affixes the CE marking and registers the system. EN 18286:2026 became the first harmonised standard under the Act through Implementing Decision (EU) 2026/2274, giving providers that build their QMS on it a presumption of conformity with Article 17 [7].
Public bodies, private entities providing public services, and deployers of credit-scoring or insurance-pricing systems must also complete a fundamental rights impact assessment (FRIA) under Article 27 before first use, and can build it on an existing GDPR DPIA [1].
Step 6: Govern GPAI and Vendor-Supplied AI
Much of the evidence a deployer needs sits with the vendor, and contracts make it obtainable. General-purpose AI (GPAI) model providers have had obligations since 2 August 2025, including Article 53 technical documentation and the Annex XII downstream information [1]. For high-risk systems, Article 25(4) requires providers and third-party suppliers to agree in writing on the information, technical access and assistance needed, “based on the generally acknowledged state of the art” [1]. The EU model contractual clauses for AI procurement (MCC-AI), updated on 5 March 2025, offer a starting point, though they were written for public buyers [8].
A practical procurement protocol:
- Confirm the vendor’s role and supported intended purpose, and record both in the inventory.
- Request the documentation, and for high-risk systems the declaration of conformity and registration details. For GPAI models, request the Annex XII information.
- Write the Article 25(4) terms into the contract, with log access, change notices and incident notification.
- Monitor the vendor system in production against the same thresholds as first-party systems.
Step 7: Reuse Existing Controls via ISO 42001, GDPR, NIST and DORA Crosswalk
Existing frameworks cover much of the management-system and risk work, but none covers the EU product-compliance steps. Only harmonised standards cited in the Official Journal create a presumption of conformity under Article 40 [1], which is why EN 18286 matters more than ISO/IEC 42001 certification alone [7]. NIST notes that its AI RMF and ISO/IEC 42001 crosswalk “does not imply that either resource comprehensively covers the contents of the other” [10].
The following illustrative mapping shows where existing controls help and what remains:
Log each remaining item as a dated action with a named owner, and feed it into the Step 5 work plan.
Step 8: Track Guidance, Standards and Deadlines as They Change
The rulebook is still filling in. Final high-risk classification guidelines are pending [5], harmonised standards for risk management, cybersecurity, logging and data governance still need adoption and Official Journal citation [7], and the Commission has committed to publishing practical AI literacy examples on its Single Information Platform [3]. One owner should watch these sources, re-run Step 3 when final guidance publishes, and update inventory entries whenever a vendor, model or intended purpose changes.
Worked Example: A CV Screening Tool From Inventory to Conformity
An EU employer that buys a vendor’s CV-screening tool and runs it across several Member States shows how the steps connect. Each stage draws on the consolidated text [1] unless noted.
- Inventory: The entry names the HR director as owner, records filtering and ranking applications as the purpose, lists the vendor and contract, and flags candidate personal data.
- Role: The employer is a deployer. Rebranding the tool, or a substantial modification that keeps it high-risk, would make it the provider under Article 25(1).
- Classification: The tool falls under Annex III point 4(a), and if it profiles candidates the Article 6(3) exception cannot apply. The prohibition screen checks for emotion recognition from video interviews, banned in the workplace by Article 5(1)(f).
- Deployer obligations: No FRIA is required for a private employer, but a DPIA is expected [9]. Before go-live, the employer informs workers’ representatives under Article 26(7), assigns oversight to recruiters who can override rankings, tells candidates, and keeps logs for at least six months. Rejected candidates can request explanations under Article 86.
- Conformity: The vendor completes Annex VI internal control, signs the declaration of conformity, affixes CE marking and registers the system, and the employer requests that evidence before go-live after 2 December 2027.
Who the Act Reaches Outside the EU and What Violations Cost
Article 2(1) covers providers placing AI systems on the EU market “irrespective of whether those providers are established or located within the Union or in a third country,” and third-country providers and deployers “where the output produced by the AI system is used in the Union” [1]. A US insurer whose model scores applicants in Germany is in scope. Non-EU providers of high-risk systems must appoint an EU authorised representative under Article 22, and GPAI model providers must do the same under Article 54 unless their model has a qualifying free and open-source licence and presents no systemic risk.
Article 99 sets fines of up to EUR 35 million or 7% of worldwide annual turnover for prohibited practices, up to EUR 15 million or 3% for breaches of operator obligations such as Articles 16 and 26, and up to EUR 7.5 million or 1% for misleading information to authorities, whichever amount is higher [1]. Small and medium-sized enterprises pay the lower amount in each tier.
How Fiddler Supports EU AI Act Governance
Several steps above depend on evidence produced while AI systems run: what was deployed, what each system did, which policies applied and who intervened. The Fiddler AI Observability and Security platform gives teams the controls and evidence to govern AI in production across first-party agents, third-party agents and coding agents.
- AI registry: Part of Fiddler’s AI governance capabilities, the registry gives a single view of live, testing and retired models and applications.
- Inline enforcement: Request checks run before the model is invoked and response checks before output reaches the user or a downstream system, with under 80ms response time, deployed directly or through the existing LLM or MCP gateway.
- Fiddler Centor Models: Batteries-included Fiddler Centor Models score PII and PHI exposure, prompt injection and faithfulness in-environment, so data stays inside the existing environment.
- Deployment: SaaS, VPC, on-premise and AWS GovCloud.
The table below maps these capabilities to the requirements they support:
These capabilities support compliance evidence. Conformity assessment, CE marking, FRIA and DPIA decisions, and the legal judgments behind classification remain with the provider, the deployer and their counsel.
Request a demo to review how the required controls would fit the existing agent architecture and deployment environment.
References
[1] European Parliament and Council of the European Union, “Regulation (EU) 2024/1689 laying down harmonised rules on artificial intelligence (Artificial Intelligence Act), consolidated text,” EUR-Lex, Jul. 27, 2026. Accessed: Nov. 12, 2026. [Online]. Available: https://eur-lex.europa.eu/legal-content/EN/TXT/HTML?uri=CELEX%3A02024R1689-20260727
[2] European Parliament and Council of the European Union, “Regulation (EU) 2026/1744 amending Regulation (EU) 2024/1689 (Digital Omnibus on AI),” Official Journal of the European Union, Jul. 24, 2026. Accessed: Nov. 12, 2026. [Online]. Available: https://eur-lex.europa.eu/eli/reg/2026/1744/oj/eng
[3] European Commission, “AI Literacy: Questions & Answers,” Shaping Europe’s Digital Future. Accessed: Nov. 12, 2026. [Online]. Available: https://digital-strategy.ec.europa.eu/en/faqs/ai-literacy-questions-answers
[4] Gartner, “Gartner Identifies Critical GenAI Blind Spots That CIOs Must Urgently Address,” press release, Nov. 19, 2025. Accessed: Nov. 12, 2026. [Online]. Available: https://www.gartner.com/en/newsroom/press-releases/2025-11-19-gartner-identifies-critical-genai-blind-spots-that-cios-must-urgently-address0
[5] European Commission, “Targeted consultation on the draft guidelines for the classification of high-risk artificial intelligence systems,” Shaping Europe’s Digital Future, Jun. 16, 2026. Accessed: Nov. 12, 2026. [Online]. Available: https://digital-strategy.ec.europa.eu/en/consultations/targeted-consultation-draft-guidelines-classification-high-risk-artificial-intelligence-systems
[6] ISACA, “AI Use Accelerates While Governance and ROI Lag Says New ISACA Research,” press release, 2026. Accessed: Nov. 12, 2026. [Online]. Available: https://www.isaca.org/about-us/newsroom/press-releases/2026/ai-use-accelerates-while-governance-and-roi-lag-says-new-isaca-research
[7] CEN-CENELEC JTC 21, “EN 18286 becomes the first harmonised standard under the EU AI Act,” Oct. 14, 2026. Accessed: Nov. 12, 2026. [Online]. Available: https://jtc21.eu/en-18286-becomes-the-first-harmonised-standard-under-the-eu-ai-act
[8] European Commission Public Buyers Community, “Updated EU AI model contractual clauses now available,” Mar. 5, 2025. Accessed: Nov. 12, 2026. [Online]. Available: https://public-buyers-community.ec.europa.eu/communities/procurement-ai/news/updated-eu-ai-model-contractual-clauses-now-available
[9] Commission Nationale de l’Informatique et des Libertés (CNIL), “Entry into application of the European AI Regulation: the first questions and answers from the CNIL.” Accessed: Nov. 12, 2026. [Online]. Available: https://www.cnil.fr/en/entry-application-european-ai-regulation-first-questions-and-answers-cnil
[10] National Institute of Standards and Technology, “Crosswalk Documents,” NIST AI Resource Center. Accessed: Nov. 12, 2026. [Online]. Available: https://airc.nist.gov/airmf-resources/crosswalks
Frequently Asked Questions
Which EU AI Act Obligations Are Enforceable Right Now?
The Article 5 prohibitions and the Article 4 literacy duty have applied since 2 February 2025, and GPAI provider obligations since 2 August 2025. National authorities gained enforcement powers on 2 August 2026, when Article 50 transparency obligations also took effect. Two further prohibitions apply from 2 December 2026 [1].
How Does an Organization Test Whether a System Is High-Risk Under Annex III?
Compare the system’s intended purpose with the eight Annex III areas, such as employment, credit scoring and biometric identification. If it matches, check whether the Article 6(3) exception applies; any system that profiles natural persons stays high-risk. A provider relying on the exception must document the assessment and register the system.
What Separates a Provider From a Deployer?
The provider develops a system and places it on the market under its own name, and carries the QMS, documentation, conformity assessment, CE marking and registration duties. The deployer uses the system under its authority, follows the instructions for use, assigns human oversight and keeps logs. Rebranding, substantial modification or a new high-risk purpose can make a deployer the provider [1].
Can ISO 42001, GDPR or NIST AI RMF Controls Be Reused?
Much of the risk management, QMS and data protection work carries over, and a FRIA can build on an existing DPIA [1]. Only harmonised standards such as EN 18286 create a presumption of conformity [7].
Does Using ChatGPT or Microsoft Copilot Bring an Organization Into Scope?
Workplace use reasonably makes the organization a deployer, because the deployer definition excludes only personal, non-professional use. For general productivity use the obligations are mostly AI literacy and Article 50 disclosure. Using the same tools to rank candidates or assess creditworthiness may move that use into an Annex III high-risk category.
