- High-risk AI systems need risk management, data governance, technical documentation, automatic logging, information for deployers, human oversight, and accuracy, robustness and cybersecurity
- Conformity assessment, the EU declaration of conformity, CE marking and, in most cases, registration in the EU database are obligations of the provider
- Deployers must follow the instructions for use, assign competent human oversight, monitor operation, inform workers and keep logs under their control for at least six months
- Do not assume your vendor is handling compliance: ask now for the EU declaration of conformity, the instructions for use and a plan to reach December 2027
This page covers what to put in place for high-risk AI: the requirements, the split between provider and deployer duties, and a four-week plan. For the enforcement timeline, what Annex III covers and the fine structure, see our EU AI Act enforcement timeline.
What high-risk AI systems must have in place by December 2027
High-risk AI systems must operate within a documented, auditable framework. Articles 9 to 15 set the technical requirements, and the Act adds obligations that sit with the provider, such as conformity assessment and registration.
Providers, meaning organisations that develop and market the systems, bear primary responsibility for building those requirements in. Deployers must operate within the constraints providers establish and meet their own obligations around deployment and monitoring.
The main requirements and obligations are:
- Risk management system. A continuous, documented process for identifying and mitigating risks across the AI system's lifecycle. Providers must maintain and update it as the system evolves, not treat it as a one-time exercise.
- Data governance. Documentation of the training, validation, and testing data used to build the system. The data must be relevant, sufficiently representative and, to the best extent possible, free of errors and complete, and the provider must examine it for possible biases.
- Technical documentation. A detailed technical file describing the system's design, development, and capabilities. National supervisory authorities can request this file at any time.
- Automatic event logging. The system must allow events to be recorded automatically over its lifetime, with traceability appropriate to its purpose, so risks can be identified and its operation monitored.
- Transparency and information to deployers. Providers must give deployers documentation that enables them to understand the system, use it correctly, and meet their own obligations under the Act.
- Human oversight. High-risk AI systems must allow human review and override. A person must be able to intervene, correct, or halt the system. Fully automated decisions without human oversight are non-compliant across most high-risk categories.
- Accuracy, robustness, and cybersecurity. Systems must meet defined accuracy thresholds, resist adversarial attempts to manipulate outputs, and protect the data they process.
- Conformity assessment, EU declaration and CE marking. These are provider obligations, separate from the technical requirements: the provider must complete the conformity assessment, issue the EU declaration of conformity, affix the CE marking and, in most cases, register the system in the EU database before placing it on the market or putting it into service.
Most of the EU database for high-risk AI systems is public, with exceptions: critical infrastructure systems are registered at national level, and biometric, law enforcement and migration systems sit in a non-public section. Deployers that are public authorities must also register their use. Ask your vendors to confirm registration and supply valid conformity declarations before December 2027.
Your duties as a deployer versus what your AI vendor must handle
The Act draws a clear line between providers and deployers. Understanding where that line sits prevents both duplicated effort and gaps in your compliance position.
Providers are responsible for: building risk management and data governance into the system, creating and maintaining technical documentation, ensuring the system generates automatic event logs, keeping documentation for 10 years, issuing the EU declaration of conformity, registering the system in the EU AI database, and giving deployers clear instructions for use.
Deployers are responsible for: using the system according to the provider's instructions for use, assigning human oversight to people with the competence and authority to exercise it, making sure input data under their control is relevant, monitoring operation and reporting risks or serious incidents to the provider and authorities, informing workers before using high-risk AI in the workplace, carrying out a fundamental rights impact assessment where Article 27 requires one (public bodies, providers of public services, and organisations using AI for credit scoring or life and health insurance pricing), and keeping the automatically generated logs under their control for at least six months.
Many AI tools were not built with EU AI Act compliance in mind. If your vendor does not yet have an EU declaration of conformity or the required instructions for use, ask for a remediation plan with owners and dates. Organisations that find gaps early have time to fix them; those that wait until mid-2027 will not.
Do not assume your AI vendor is handling compliance. Not every vendor, inside or outside the EU, will have its documentation ready in time. Ask for evidence now. "We are working on it" is not a defensible position once enforcement begins.
Building your AI compliance foundation
With December 2027 as the Annex III deadline, organisations have time to build a proper compliance programme. The Digital Omnibus delay is not a reason to defer the work. Harmonised standards are now emerging and the December 2027 date is firm. Organisations starting now can inventory, classify, engage vendors, and document oversight without the pressure that comes with a looming deadline.
Week 1: Build your AI inventory
Ask every business unit the same question: what software do you use that makes or recommends decisions about people? Include HR platforms, recruitment tools, CRM systems with lead or credit scoring, workforce management software, fraud detection tools, and customer service routing systems. For each tool, record the vendor, the use case, the likely Annex III category, the categories of personal data processed, whether the system is used in the EU or produces output used in the EU, and the internal owner.
Week 2: Classify and prioritise
For each AI system in your inventory, determine whether it meets the Annex III definition of high-risk. A chatbot that routes customer queries does not qualify. A system that ranks job applicants or scores customers for loan eligibility does. Record the reasoning behind each classification. If a vendor relies on the Article 6(3) exception, ask for its documented assessment. Prioritise high-risk systems for immediate action. Bring in your legal or compliance team to confirm classifications where doubt remains.
Week 3: Engage vendors and review documentation
Contact each vendor of a high-risk AI system and request the EU declaration of conformity (or the timetable for issuing it), the instructions for use required by Article 13, confirmation of registration in the EU database where it applies, and the vendor's plan for reaching December 2027. Log every vendor communication in writing. If a vendor cannot produce these documents, escalate internally and decide whether continued use of that tool fits your risk appetite.
Week 4: Document human oversight, logs and incident reporting
For each high-risk AI system in use, write down the human oversight process: who reviews AI outputs, how they can override the system, and how those override decisions get recorded. Tell affected employees in writing where any system monitors or evaluates their performance. Decide how system-generated logs will be kept for at least six months and how serious incidents will be reported to the provider and the authority. Assign a named owner for each system's ongoing compliance.
Where to focus now
The December 2027 deadline is firm and the obligations are not optional. The Digital Omnibus delay reflected unfinished technical standards, not a softening of requirements. Compliance means operating AI tools with documentation, oversight, and accountability, not abandoning them.
Inaction carries two distinct risks. Regulators can fine your organisation once enforcement begins. If an AI-supported decision harms an individual and you cannot demonstrate appropriate safeguards, you face civil liability on top of regulatory penalties.
December 2027 gives most organisations time to build a mature programme. Start with the foundation: an AI inventory classified against Annex III, vendor conformity documentation confirmed, human oversight procedures written down, and logs retained. That foundation supports good governance regardless of when authorities act.
- Complete an AI inventory across all business functions
- Classify each AI use against Annex III and confirm scope with legal or compliance counsel
- Request EU conformity documentation from every vendor of a high-risk AI system in writing
- Document human oversight procedures for every high-risk AI tool currently in use
- Notify employees in writing where AI systems assess or monitor their performance
- Keep logs generated by high-risk systems for at least six months and set out how serious incidents will be reported
- Check which Article 50 duties fall on you: chatbot notices and synthetic content marking (providers) and deepfake disclosure (deployers); the Article applies now
- Assign a named owner for each high-risk AI system and schedule a review date for ongoing compliance