In short: The EU AI Act (Regulation EU 2024/1689) has been in force since August 2024 and applies in stages. Following a 2026 amendment (Regulation (EU) 2026/1744), the obligations for high-risk systems under Annex III apply only from 2 December 2027 (see timeline). Public authorities are particularly affected because many typical use cases (administrative decisions, law enforcement, migration, critical infrastructure) classify as high risk. Which AI applications qualify, which duties apply, and what a conformity assessment looks like in practice.
Why authorities are particularly affected
The EU AI Act classifies AI systems by risk. Four categories:
- Prohibited AI (Article 5): social scoring, manipulative AI, real-time remote biometric identification in publicly accessible spaces for law enforcement (with exceptions)
- High-risk AI (Article 6 and Annex III): extensive obligations, conformity assessment
- Limited risk (Article 50): transparency duties (e.g. chatbots must be identifiable)
- Minimal risk: no specific duties
Annex III lists the high-risk areas. Several directly affect authorities:
- Biometrics (e.g. remote biometric identification)
- Education and vocational training (e.g. exam scoring, admissions)
- Employment and HR management (recruiting, performance evaluation)
- Access to essential private and public services (welfare, creditworthiness)
- Law enforcement (risk assessment, evidence analysis)
- Migration, asylum, border control
- Justice and democratic processes
- Critical infrastructure (safety components in management and operation)
Authorities deploying AI in any of these areas are usually subject to the high-risk rules. Exceptions are set out in Article 6 (3), for example for systems that only perform narrow preparatory tasks.
Obligations for high-risk systems
Articles 8 to 15 define the core requirements for the system. They are mainly implemented by the provider. In practice this means:
- Risk management system (Article 9) across the full lifecycle. Identify, assess, minimize, monitor. Documented, regularly updated.
- Data governance (Article 10). Training, validation and test data must be relevant, sufficiently representative and, to the best extent possible, free of errors and complete. Bias testing documented.
- Technical documentation (Article 11 and Annex IV). Extensive: purpose, architecture, datasets, validation results, risks. Detailed enough that a supervisor can assess conformity from documentation alone.
- Logging duty (Article 12). The system must automatically record events over its lifetime. Deployers keep the logs for at least six months (Article 26 (6)), often longer.
- Transparency (Article 13). Clear usage instructions, indication of limitations, output explainability.
- Human oversight (Article 14). A human must review and, if necessary, override the AI result. No fully automated processing for decisions with legal effects.
- Accuracy, robustness, cybersecurity (Article 15). Measurable requirements, regular tests, protection against adversarial attacks.
Conformity assessment
Before a high-risk system is placed on the market or put into service, the provider must carry out a conformity assessment (Article 43). For the high-risk areas in Annex III points 2 to 8, including critical infrastructure, this is generally internal control by the provider under Annex VI. Only for biometric systems (Annex III point 1) may a notified body need to be involved.
Output of the conformity assessment:
- EU declaration of conformity (Article 47)
- CE marking of the system (Article 48)
- Entry in the EU database for high-risk systems (Articles 49 and 71)
Authorities must verify that purchased high-risk systems have undergone this conformity assessment.
Obligations of authorities as deployers
The requirements above are aimed mainly at the provider, i.e. whoever develops the AI system or places it on the market under its own name. Authorities are usually deployers. Their own obligations include:
- Use in accordance with the instructions and human oversight by trained, authorized persons (Article 26 (1) and (2))
- Monitoring of operation, informing the provider and market surveillance of risks or serious incidents (Article 26 (5))
- Retention of logs for at least six months, where they are under the authority's control (Article 26 (6))
- Informing affected persons when a high-risk system makes or assists decisions about natural persons (Article 26 (11))
- Fundamental rights impact assessment before first use, mandatory for bodies governed by public law (Article 27)
- Registration of use in the EU database by authorities as deployers (Article 49 (3))
- AI literacy of staff working with AI systems (Article 4, since 2 February 2025)
If an authority develops a system itself and uses it under its own name, it is also the provider and bears the provider obligations too.
Checklist for authorities
Recommended implementation path:
1. Inventory AI applications
- Which AI systems are in use or planned?
- Per system: purpose, data flow, decision contribution, affected persons
- Output: inventory with risk classification
2. Classify against Annex III
- Per system: does it fall under high-risk?
- For unclear cases: obtain legal assessment
- Output: classification document with reasoning per system
3. Gap analysis for high-risk systems
- Per high-risk system: which provider and deployer obligations are met, which are not?
- Fundamental rights impact assessment planned before use (Article 27)?
- Output: action plan with owners and deadlines
4. Set up data governance
- Classify training data, document bias tests
- Personal data: legal basis under GDPR and compatibility with the EU AI Act
- Output: data catalog with protection levels and responsible owners
5. Build technical documentation
- Only for in-house development (authority as provider): Annex IV defines the structure
- For purchased systems: request and review the provider's instructions for use and documentation
- Living documentation, not one-off
- Output: documented system per authority standard
6. Establish human oversight
- Roles, permissions, escalation paths
- Training for oversight personnel
- Output: rules of procedure for AI oversight
7. Monitoring and reporting
- Secure logs and keep them for at least six months
- Periodic review of model performance
- Incident reporting channels to the supervisor (in Germany, the Federal Network Agency is envisaged as the central market surveillance authority)
- Output: monitoring dashboard, reporting calendar
Procurement: what tender documents must contain
Authorities procuring AI systems must include conformity requirements in tender documents. At minimum:
- Obligation to provide the EU declaration of conformity
- CE marking as eligibility condition
- Access to technical documentation for the operations phase
- Cooperation duties on incidents and conformity checks
- Data-processing guarantees (location, jurisdiction, sub-contractors)
- Audit and inspection rights
Applicability timeline
The AI Act applies in stages. Originally planned were:
- 2 Feb 2025: Prohibited AI practices and AI literacy duty (Articles 4 and 5)
- 2 Aug 2025: Obligations for providers of general-purpose AI models (GPAI)
- 2 Aug 2026: Main obligations, in particular for high-risk systems under Annex III, and transparency duties under Article 50
- 2 Aug 2027: High-risk systems embedded in products under Annex I
Regulation (EU) 2026/1744 of 8 July 2026, the so-called Digital Omnibus on AI, amended these deadlines. It was published in the Official Journal of the EU on 24 July 2026 and entered into force on 27 July 2026. Obligations for high-risk systems under Annex III now apply from 2 December 2027, and for high-risk systems embedded in products under Annex I from 2 August 2028. Most transparency duties under Article 50 still apply from 2 August 2026. Sources: Regulation (EU) 2026/1744 on EUR-Lex, European Commission, 27 July 2026. Authorities should use the extra time for inventory, classification and the fundamental rights impact assessment.
How GermanAI Defense supports
We guide authorities through EU AI Act conformity: inventory, classification, gap analysis, data governance, fundamental rights impact assessment and documentation. On request, we operate the solution in Germany or on your premises.
→ More on our AI Solutions: View AI Solutions