EU AI Act: High-Risk Obligations Deferred to December 2027 - What Still Starts in August 2026

Set up AI infrastructure with documented data paths and responsibilities? WZ-IT integrates AI Cube at your own site or designs dedicated GPU infrastructure on controlled owned or European infrastructure. Access, data paths, logging and operations are defined and documented for the specific use case. Explore AI Cube · Plan GPU infrastructure · Schedule an initial consultation

Updated 12 October 2026. This post was written in May, when the deferral was still a proposal. It has since been adopted: Regulation (EU) 2026/1744 (Digital Omnibus) was published in the Official Journal on 24 July 2026 and entered into force on 27 July 2026.

The EU AI Act's high-risk obligations do not start on 2 August 2026. Stand-alone high-risk systems under Annex III now have until 2 December 2027, high-risk AI in regulated products under Annex I until 2 August 2028.

What has applied since 2 August 2026 are the transparency obligations under Article 50: providers must design AI systems that people interact with directly so that those people recognise it (Art. 50(1)). Synthetic audio, image, video and text content must be marked in a machine-readable way (Art. 50(2)). Only for systems placed on the market before 2 August 2026 did the Omnibus introduce a transition period for this marking until 2 December 2026 (new Art. 111(4)).

This guide explains the concrete requirements and provides a checklist. The core requirements for high-risk systems stayed the same, but besides the deadlines the Omnibus also changed content, for example on AI literacy, registration, SME relief and a new prohibition.

Table of Contents

• Timeline: What applies when
• The delay debate
• What is a high-risk AI system?
• The seven obligations for high-risk AI
• Penalties and enforcement
• Practical checklist for enterprises
• On-premise AI as compliance advantage
• How WZ-IT handles this
• Related guides

Timeline: What applies when

The EU AI Act entered into force on August 1, 2024. Implementation is phased:

The high-risk obligations affect enterprises that use AI in the Annex III areas, for example for applicant screening and performance evaluation, creditworthiness assessment of natural persons, risk assessment and pricing in life and health insurance, or remote biometric identification. Annex III does not list general customer scoring, and pure biometric verification (such as access by matching a face against the authorised person) is expressly excluded (Annex III No. 1(a)). These obligations now apply from 2 December 2027.

The delay debate

The path was short: in Parliament the Internal Market Committee (IMCO) and the Civil Liberties Committee (LIBE) led the file. On 7 May 2026 Council and Parliament reached a trilogue agreement, on 16 June 2026 the plenary approved it, on 29 June 2026 the Council, and on 8 July 2026 the regulation was signed. Regulation (EU) 2026/1744 was published on 24 July and has been in force since 27 July 2026.

The reasoning stayed the same throughout: many enterprises - particularly SMEs - are not ready, the necessary harmonised standards are not yet fully available, and enforcement without clear standards creates legal uncertainty.

What this means:

• Deferred are the obligations for stand-alone Annex III high-risk systems to 2 December 2027 and for Annex I products to 2 August 2028
• Not deferred are the transparency obligations under Art. 50, which have applied since 2 August 2026. The Omnibus only created a transition period until 2 December 2026 for Art. 50(2) marking for systems placed on the market before 2 August 2026 (Art. 111(4))
• Prohibited practices have applied since February 2025, GPAI obligations since August 2025. New from 2 December 2026 is the prohibition of AI for generating non-consensual intimate content and child sexual abuse material (Art. 5(1)(ba) and (bb))
• AI literacy (Art. 4) rewritten: providers and deployers take measures to support the AI literacy of their staff. They no longer have to ensure a specific "sufficient level"
• SME relief extended: the lower fine amount under Art. 99(6) now also applies to small mid-caps with fewer than 750 staff and up to 150 million EUR turnover
• Registration reduced: the registration duty under Art. 49(2) for systems that providers themselves classify as not high-risk was narrowed
• Supervision bundled: the AI Office is competent for AI systems based on a GPAI model of the same provider
• Unlike the Commission draft of November 2025, the new dates are fixed calendar dates rather than tied to the availability of harmonised standards. A further deferral would need a new legislative procedure

For enterprises this means: you have gained time, not an all-clear. If you operate an Annex III system you now have a reliable date instead of a contested one - and just over a year to build the documentation that previously had to be in place within weeks.

What is a high-risk AI system?

Annex III of the EU AI Act defines eight categories of high-risk AI systems:

1. Biometric identification and categorization - Facial recognition, emotion recognition (where not prohibited)
2. Critical infrastructure - AI control of electricity, water, gas, and transport networks
3. Education and vocational training - Exam grading, access control for educational institutions
4. Employment and workforce management - Applicant screening, AI-assisted performance evaluation, termination decisions
5. Access to essential services - Creditworthiness assessment, insurance evaluation, social benefit decisions
6. Law enforcement - Risk assessment, lie detection, predictive analytics
7. Migration, asylum, and border control - Entry risk assessment, document verification
8. Administration of justice and democratic processes - AI-assisted court decisions, election influence

Important: It does not matter whether the system is marketed as "AI." What matters is whether it qualifies as an AI system under the regulation and is deployed in one of these areas. A machine learning model that pre-screens job applications is high-risk - regardless of whether the vendor calls it "intelligent filtering."

Commission guidelines on classification under Article 6 so far exist only as a draft of 19 May 2026; the final version is announced for the end of 2026.

Special case: Safety components in products

AI systems used as safety components in regulated products (medical devices under MDR, machinery under the Machinery Regulation, vehicles) also fall under high-risk classification. These are subject to the longer deadline of 2 August 2028, as they are tied to existing EU product legislation.

The seven obligations for high-risk AI

Providers placing a high-risk AI system on the market must meet seven core requirements. Deployers have their own obligations under Article 26, for example on human oversight and log retention:

1. Risk management system (Article 9)

Continuous, documented risk management throughout the AI system's lifecycle:

• Identification and analysis of known and foreseeable risks
• Risk assessment under intended use and reasonably foreseeable misuse
• Risk mitigation measures
• Regular review and updates

2. Data governance (Article 10)

Training, validation, and testing must be based on datasets that:

• Ensure relevance, representativeness, and accuracy
• Have undergone bias checks
• Are documented and traceable
• Contain special categories of personal data only under strict conditions

3. Technical documentation (Article 11)

Comprehensive technical documentation must be prepared before placing on the market:

• General description of the system and its purpose
• Detailed description of the development methodology
• Information on training data and its preparation
• Performance metrics and measurement methodology
• Description of risk management measures

4. Record-keeping / Logging (Article 12)

High-risk AI systems must generate automatic records (logs):

• Start and end time of each use
• Reference database against which input was checked
• Input data that led to a match
• Identification of persons involved in human oversight

Logs must be retained for a period appropriate to the intended purpose - at minimum six months.

5. Transparency (Article 13)

Users must be adequately informed:

• About the system's capabilities and limitations
• About the degree of accuracy and known error sources
• About circumstances under which risks may arise
• About human oversight measures

6. Human oversight (Article 14)

High-risk AI systems must be designed to allow effective human oversight:

• The human must be able to understand the system's outputs
• The human must be able to interrupt or stop the system
• The human must be able to disregard or override the system's results
• The human must be able to recognize automation bias

7. Quality management system (Article 17)

Providers of high-risk AI must implement a QMS:

• Strategy for regulatory compliance
• Design and control techniques
• Risk management procedures
• Post-market monitoring
• Cybersecurity measures

Penalties and enforcement

Fines are tiered by severity of violation:

In general the higher of the two amounts applies (Art. 99(3) to (5) of Regulation (EU) 2024/1689). For SMEs including start-ups and, since the Omnibus, small mid-caps (fewer than 750 staff, up to 150 million EUR turnover) the lower amount applies (Art. 99(6)).

Enforcement happens at the national level. In Germany the AI Market Surveillance Act (KI-MIG) of 22 July 2026 (BGBl. 2026 I No. 223) governs supervision; it has been in force since 29 July 2026. The central market surveillance authority is the Federal Network Agency (Bundesnetzagentur) (Section 2(1) KI-MIG). BaFin is competent for financial activities, the Länder for public bodies. The KI-MIG adds national fines of up to 50,000 EUR (Sections 15, 16).

Practical checklist for enterprises

Step 1: Create an AI inventory

Capture all AI systems in the enterprise - not just the obvious ones:

• Which models are in use? (In-house, API-based, embedded in SaaS)
• In what context? (HR, customer service, financial decisions, production)
• Who is the provider, who is the deployer?
• What data flows into the system?

Step 2: Risk categorization

For each identified system, check:

• Does it fall under Annex III? (High-risk standalone)
• Is it a safety component of a regulated product?
• Does it fall under transparency obligations (chatbots, deepfakes)?
• Or is it minimal risk (no specific obligations)?

Step 3: Gap analysis

For each high-risk system, determine the gap between current state and target state:

• Does technical documentation exist?
• Is there a risk management system?
• Are logs being stored?
• Is human oversight implemented?
• Has a bias check been performed?

Step 4: Prepare conformity assessment

Most Annex III high-risk systems can be placed on the market through a self-assessment (internal conformity assessment). For biometric systems under Annex III No. 1 a notified body is generally involved. If the provider fully applies harmonised standards or common specifications, internal control is possible here as well (Art. 43(1)).

For self-assessment you need:

• Complete technical documentation
• Evidence of a QMS
• EU declaration of conformity
• CE marking
• Registration in the EU database

Step 5: Registration and CE marking

Before placing on the market or putting into service, the system must be registered in the EU database for high-risk AI systems. The database is operated by the EU Commission and is partially publicly accessible. For systems the provider classifies as not high-risk under Art. 6(3), the Omnibus reduced the registration under Art. 49(2).

On-premise AI as compliance advantage

The EU AI Act's requirements sound complex - and they are. But one factor significantly simplifies compliance: control over the infrastructure.

Why local inference helps

Concrete example: Applicant screening

An enterprise uses an LLM for pre-screening job applications. This is a high-risk use case (Category 4: Employment).

With Cloud API (e.g., GPT-4):

• Application data (name, CV, photo) sent to US servers
• No insight into model logic (Article 13 transparency problematic)
• Logs only available via API dashboard (Article 12 logging limited)
• Bias checks only on output (Article 10 data governance restricted)
• Under CLOUD Act request: access to applicant data by US authorities possible

With on-premise AI (e.g., Llama 3.3 70B on own GPU server):

• Application data never leaves the enterprise
• Model, prompt, and decision logic fully documentable
• All logs stored locally and auditable
• Bias checks on input data, prompt design, and output possible
• No third-country access, no CLOUD Act issue

Both approaches can be operated in compliance with the EU AI Act. But the on-premise approach makes compliance practically achievable, while the cloud approach relies on provider assurances.

How WZ-IT handles this

We help enterprises build AI infrastructure that is EU AI Act-compliant - not with consulting slides, but with concrete technology:

1. Assessment: Which AI systems are in use? Which fall under high-risk? Which under transparency obligations?

2. Infrastructure design: High-risk applications may use dedicated GPU servers on controlled infrastructure or AI Cube at the organisation's own site. Data location, models, access and logs are defined in the target architecture.

3. Logging and audit trail: We configure complete request-response logging with timestamps, user IDs, and model versions - exportable for audits and regulatory inquiries.

4. Managed Operations: Ongoing operations including monitoring, updates, and compliance documentation - so the technical documentation doesn't go stale.

5. GDPR + AI Act: We address both regulatory frameworks simultaneously. Local inference solves GDPR transfer issues and simplifies AI Act documentation.

Whether GPU servers on owned or European infrastructure or AI Cube as the local entry point, WZ-IT designs data location, access paths and operational responsibility around the specific use case.

Related guides

• AI Sovereignty: Why German Companies Should Not Send Data to US AI Services
• GDPR-Compliant AI Inference with GPU Servers
• Ollama vs. vLLM: Comparison for Self-Hosted LLMs
• LLM Observability with Langfuse: AI Logging for the EU AI Act
• Local AI Inference with the AI Cube
• Managed Operations: Compliance
• GPU Server at WZ-IT
• AI Cube

As of October 2026, after Regulation (EU) 2026/1744 and the KI-MIG entered into force. Commission guidelines, for example on high-risk classification, are still being added. Enterprises should follow the official channels of the EU Commission and the Federal Network Agency (Bundesnetzagentur).