The EU AI Act sets the rules, but who inspects, who takes complaints, who registers which systems and who fines whom in Germany is decided by national law, and that law only arrived in July 2026. The KI-Marktüberwachungs- und Innovationsförderungsgesetz (KI-MIG) of 22 July 2026 (BGBl. 2026 I Nr. 223) entered into force on 29 July 2026 and makes the Bundesnetzagentur (BNetzA) Germany's default AI market surveillance authority, single point of contact and central complaints office. This guide explains the authority map, the independent KI-Marktüberwachungskammer, the national register for certain critical infrastructure AI systems, the German fines on top of Article 99 of the AI Act, the regulatory sandbox and real-world testing rules, and how the EU Digital Omnibus changed the dates. It is for compliance, security and AI governance teams at providers and deployers that place or use AI systems in Germany.
Source note: section content, fines and dates are from the KI-MIG text on gesetze-im-internet.de (BGBl. 2026 I Nr. 223) and the Bundesnetzagentur release of 29 July 2026; AI Omnibus dates are from FPF; English renderings are ours. Confirm against the official text before relying on it.
Key takeaways
- The BNetzA is the default supervisor. Under § 2(1) KI-MIG it is the market surveillance authority wherever no sector authority is named, and under §§ 6 and 8 it is the single point of contact and the central complaints office.
- Sector authorities keep their patch. BaFin supervises AI tied to regulated financial activity, product regulators keep AI in Annex I Section A products, and state authorities supervise state public bodies and media providers.
- A national register applies to critical infrastructure AI. § 20 creates a non-public BNetzA register for Annex III point 2 high-risk systems; providers register before placing on the market, and federal public bodies register as deployers.
- German fines add to EU fines. § 15 sets fines of up to EUR 50,000 for specific duties such as cooperation and explanation duties, while the large penalties still come from Article 99 of the AI Act.
- Dates moved at EU level. The Digital Omnibus on AI, published in the Official Journal on 24 July 2026, moved Annex III high-risk obligations to 2 December 2027 and Annex I obligations to 2 August 2028, so the KI-MIG machinery is in place before most high-risk duties apply.
What the KI-MIG is and its status
The KI-MIG is Article 1 of a German federal act of 22 July 2026. Its official citation is "KI-Marktüberwachungs- und Innovationsförderungs-Gesetz vom 22. Juli 2026 (BGBl. 2026 I Nr. 223)", and under Article 5 of the enacting act it entered into force on 29 July 2026. The BNetzA announced the same day that it was taking up its role immediately. Germany had missed the AI Act deadline of 2 August 2025 for designating national authorities; the government bill (Bundestag printed paper 21/4594) went through committee hearings in March 2026 before adoption.
The law does not repeat the AI Act. § 1 says it implements Regulation (EU) 2024/1689 and supplements it on authorities, innovation measures and fines. Substantive duties, such as risk management for high-risk systems, prohibited practices, transparency under Article 50 and general-purpose AI model duties, come straight from the Regulation. See the EU AI Act Compliance Guide for those obligations.
| Milestone | Date | Source |
|---|---|---|
| AI Act in force | 1 August 2024 | Regulation (EU) 2024/1689 |
| Deadline for national authorities (missed) | 2 August 2025 | AI Act Article 70 |
| Bundestag digital committee hearing | March 2026 | Bundestag hearing statements of 23 March 2026 |
| KI-MIG signed | 22 July 2026 | BGBl. 2026 I Nr. 223 |
| Digital Omnibus on AI published | 24 July 2026 | Official Journal of the EU |
| KI-MIG in force | 29 July 2026 | Article 5 of the enacting act |
| Article 50 transparency applies (50(2) GPAI marking from 2 December 2026) | 2 August 2026 | AI Act as amended by the Omnibus |
| Annex III high-risk obligations apply | 2 December 2027 | AI Act as amended by the Omnibus |
| National sandbox operational | 2 August 2027 | AI Act as amended by the Omnibus |
| Annex I high-risk obligations apply | 2 August 2028 | AI Act as amended by the Omnibus |
| First KI-MIG evaluation due | Within 18 months of entry into force | § 19 KI-MIG |
Dates are as published by the sources listed and checked in November 2026. See the EU AI Act Implementation Timeline Tracker and the EU Digital Omnibus Guide for any later changes.
Who supervises what: the authority map
Germany chose what the BNetzA calls a hybrid model: central coordination and default supervision at the BNetzA, with existing sector regulators retained where product or financial law already has a supervisor.
| AI system or actor | Competent authority | KI-MIG basis |
|---|---|---|
| Default case, including AI used in HR, education and critical infrastructure, and AI in radio equipment | Bundesnetzagentur | § 2(1) |
| AI in products under Annex I Section A Union harmonisation law (machinery, medical devices, toys and similar) | The existing sector market surveillance authority | § 2(2) |
| AI tied to regulated financial activity of supervised institutions | BaFin | § 2(3) |
| Other financial AI not covered by BaFin | Other financial supervisory authorities | § 2(4) |
| Annex III high-risk AI in law enforcement, migration and border control, justice and democratic processes, and Annex III points 6 to 8 | KI-Marktüberwachungskammer at the BNetzA | § 2(5), § 4 |
| AI used by state (Land) public bodies | State authorities (states may use the BNetzA through an Organleihe arrangement) | § 2(6) |
| Media service providers | State authorities; Deutsche Welle under its own law | § 2(8) |
| Notifying authority for Annex III point 1 (biometrics) conformity assessment bodies | The authority designated under Article 52(2) of the Cyber Resilience Act; the BSI until then | § 3(2) |
| Cross-border non-compliance notifications to the Commission | Bundesanstalt für Arbeitsschutz und Arbeitsmedizin (BAuA) | § 7 |
The KI-Marktüberwachungskammer
§ 4 creates an independent chamber inside the BNetzA for the most sensitive Annex III uses. It has three members (the BNetzA President as chair and the two Vice Presidents), decides by simple majority with a casting vote for the chair, takes no instructions, uses a secretariat whose staff answer only to the chamber for that work, and reports to the Bundestag by 31 March each year, starting with the 2026 report. It does not review the formal and substantive requirements for orders authorising high-risk uses under Article 5(2) and (3) and Article 26(10) of the AI Act, which remain with the bodies that issue those orders.
National registration duties
The AI Act already requires providers of Annex III high-risk systems and certain public sector deployers to register in the EU database under Article 49. Article 49 also provides that Annex III point 2 systems, those used as safety components in critical digital infrastructure, road traffic and the supply of water, gas, heating and electricity, are registered at national level instead. § 20 KI-MIG implements that duty in Germany.
- Register holder: the BNetzA runs a register for Annex III point 2 high-risk AI systems. The register is not public.
- Providers: must register before placing the system on the market or putting it into service.
- Deployers: federal public bodies that deploy such systems must register; state public bodies are exempt from the provider registration duty under § 20.
- Access: competent authorities may access relevant register data, and processing of personal data is permitted where needed to run the register.
- Timing: the duty bites when Annex III obligations apply, 2 December 2027 under the amended AI Act, but operators of energy, water and digital infrastructure should inventory candidate systems now.
Registration is a strong reason to have an AI inventory that records the Annex III category, intended purpose, provider and deployer roles and the market placement date. See the EU AI Act High-Risk AI Systems Guide and, for operators covered by KRITIS rules, the Germany KRITIS Rules guide.
Powers, cooperation and the BSI
- Powers (§ 11): authorities have the powers of Articles 14(4) and (5) and 16 of the Market Surveillance Regulation (EU) 2019/1020 plus those in the AI Act. They may use third-party assistants and remote access to inspect systems, and may enter premises outside business hours only to prevent urgent danger. Objections and lawsuits generally have no suspensive effect in the cases listed.
- Cooperation (§ 9): authorities exchange information, including personal data and trade secrets where strictly necessary, and must involve data protection authorities, the cybersecurity authority and the Bundeskartellamt where relevant.
- Cybersecurity of high-risk AI (§ 10): the BNetzA and the authority designated under Article 52(2) of the Cyber Resilience Act set joint guidelines for auditing the cybersecurity requirements of high-risk systems; BaFin sets its own for financial entities. The BSI performs these tasks until the designation is made. Expect Article 15 cybersecurity evidence to be reviewed against BSI style criteria.
- Coordination and competence centre (§ 5): set up at the BNetzA to support authorities on complex cases, coordinate horizontal legal questions, facilitate codes of conduct, involve civil society and industry, and provide training and guidance, including a federal and state AI committee.
- Retention on exit (§ 18): if a provider or authorised representative ceases business, the liquidator must keep the Article 18(1) documentation.
See the EU AI Act Article 15 Guide and the EU Cyber Resilience Act Guide for the cybersecurity requirements the joint guidelines will test.
Fines: what Germany adds
| Source | Conduct | Maximum |
|---|---|---|
| AI Act Article 99 | Prohibited practices, high-risk obligations, incorrect information to authorities | Set in the Regulation (the highest tier is EUR 35 million or 7 percent of worldwide annual turnover) |
| § 15 KI-MIG | Breaches of specified duties under Articles 21 (cooperation with authorities), 27 (fundamental rights impact assessment), 45 (notified body information duties) and failure by certain high-risk deployers to give explanations under Article 86(1) | EUR 50,000 |
§ 16 applies the Administrative Offences Act (OWiG) and criminal procedure rules to proceedings under Article 99(3) to (5), and a prosecutor may discontinue proceedings only with the issuing authority's consent. § 17 names the market surveillance and notifying authorities as the fining authorities, and no fines may be imposed on federal or state public bodies covered by the BDSG. Public sector deployers are therefore exposed to supervisory measures rather than fines under this law.
Innovation measures: sandbox and real-world testing
- Guidance and advice (§ 12): the BNetzA provides general guidance, training, studies and networking, and individual advice to public bodies under an administrative agreement. Its KI-Service-Desk, already operating, offers a first risk-class self assessment tool aimed at SMEs and start-ups.
- KI-Reallabor (§ 13): the BNetzA must set up and run at least one AI regulatory sandbox with the data protection authorities; research institutions and university spin-offs get priority access under conditions. The Federal Ministry for Digital Affairs and State Modernisation may set sandbox rules by ordinance and delegate that power to the BNetzA. Under the amended AI Act a national sandbox must be operational by 2 August 2027.
- Testing in real-world conditions (§ 14): providers that test high-risk systems outside a sandbox must submit a test plan to the competent authority before market placement; approval is deemed granted if the authority does not respond within 30 days.
Where identity fits
Two KI-MIG features touch identity and access teams directly. First, the § 20 register covers AI used as safety components in critical digital infrastructure and utilities, which is where AI agents and automation hold privileged machine credentials to operational systems; the Article 15 cybersecurity evidence that the BNetzA and BSI guidelines will audit includes how those credentials are issued, scoped and revoked. Second, authorities may inspect systems through remote access under § 11, so organisations should plan a controlled, logged and time-limited access route for supervisors rather than improvising shared accounts. See the Germany NIS2 Implementation Guide for the parallel NIS2 duties that apply to many of the same operators.
Implementing: a practical sequence
- Build or refresh the AI inventory: record every AI system you provide or deploy in Germany with its AI Act role, Annex III or Annex I category, sector and intended purpose.
- Map each system to its German authority: use the table above; where two could apply (for example a bank's HR screening tool), document your reasoning and confirm through the BNetzA single point of contact.
- Flag Annex III point 2 systems for the § 20 register: assign an owner and gather the registration data now so the system can be registered before market placement.
- Prepare cooperation duties: designate a contact for authority requests, a process for Article 21 information requests, and a logged remote access route for inspections. These are the duties § 15 fines attach to.
- Plan explanations and FRIAs: deployers in scope should template Article 86(1) explanations and Article 27 fundamental rights impact assessments; see the EU AI Act Fundamental Rights Impact Assessment guide.
- Decide on sandbox or real-world testing: if you will test a high-risk system in live conditions, prepare the § 14 test plan and track the 30-day deemed approval.
- Review annually: watch the chamber's annual report, the § 19 evaluation, BNetzA guidance and any further EU amendments, and update the inventory and mappings.
Decision guidance
| Situation | What to do |
|---|---|
| SME or start-up with a possible high-risk use case | Use the KI-Service-Desk risk-class tool first, then consider the KI-Reallabor before full conformity work |
| Financial institution | Expect BaFin as authority for AI tied to regulated activity and the BNetzA for other uses such as general HR tools; keep both mappings |
| Operator of energy, water or digital infrastructure | Treat § 20 registration as a project now, joined up with NIS2 and KRITIS compliance |
| Federal public body | Register Annex III point 2 deployments; no fines apply, but supervisory measures do |
| Non-EU provider selling into Germany | Appoint the AI Act authorised representative and make sure it can meet Article 21 cooperation duties in German proceedings |
Measuring readiness
- Inventory completeness: percentage of AI systems with an assigned AI Act role, risk class and German authority.
- Registration readiness: number of Annex III point 2 systems with complete § 20 registration data.
- Request response time: time to answer a simulated authority information request under Article 21.
- FRIA and explanation coverage: percentage of in-scope deployments with a completed FRIA and an Article 86(1) explanation template.
- Article 15 evidence: percentage of high-risk systems with current cybersecurity test evidence, including machine credential controls.
Common failure patterns
- Waiting for the deadline: the Omnibus delay is read as permission to pause, when registration, inventory and cooperation processes take longer to build than the remaining time.
- Assuming one authority: groups that operate in finance, media and industry find they answer to BaFin, state authorities and the BNetzA for different systems.
- Ignoring the small fines: EUR 50,000 is modest, but § 15 duties such as cooperation and explanations are the ones most likely to be tested in a first inquiry.
- Confusing EU and national registration: Annex III point 2 systems go to the German register, not the EU database, and the German register is not public.
- Improvised authority access: handing inspectors a shared admin account for remote access instead of a scoped, logged, expiring identity.
Framework mapping
| KI-MIG duty | NIST CSF 2.0 | ISO/IEC 42001 and 27001 theme |
|---|---|---|
| Know your authorities and legal duties | GV.OC-03 | Context of the organisation, legal requirements |
| AI inventory and § 20 registration | GV.OC-03, PR.AA-01 (for system and service identities) | AI system inventory and documentation |
| Cooperation with authorities and incident handling | RS.MA-01 | Contact with authorities |
| Article 15 cybersecurity evidence | PR.AA-01, PR.DS-01 | Access control and protection of information |
Practitioner checklist
- Confirm the KI-MIG citation (BGBl. 2026 I Nr. 223) and the current AI Act dates in your compliance register.
- Map every AI system to its German competent authority using § 2.
- Identify Annex III point 2 systems and prepare § 20 registration data.
- Name a contact and process for authority information and cooperation requests.
- Prepare a scoped, logged remote access route for supervisory inspections.
- Template Article 27 FRIAs and Article 86(1) explanations for in-scope deployments.
- Decide whether to apply to the KI-Reallabor or file a § 14 real-world test plan.
- Track the chamber's annual report, BNetzA guidance and the § 19 evaluation.
Standards and references
- Bundesministerium der Justiz: KI-Marktüberwachungs- und Innovationsförderungsgesetz (KI-MIG)
- Bundesnetzagentur: Bundesnetzagentur übernimmt zentrale Rolle bei der Umsetzung der KI-Verordnung
- EUR-Lex: Regulation (EU) 2024/1689 Artificial Intelligence Act
- Future of Privacy Forum: The AI Act Implementation Timeline: What Changes Under the AI Omnibus?
- activeMind.legal: KI-MIG guide to the German implementation bill
- NIST: Cybersecurity Framework 2.0
Related NHI Mgmt Group resources: EU AI Act Compliance Guide · EU AI Act High-Risk AI Systems Guide · Germany NIS2 Implementation Guide · EU Digital Omnibus Guide · EU AI Act Fundamental Rights Impact Assessment. This guide summarises the German KI-MIG and related EU AI Act provisions and is not legal advice.