Find the authority before interpreting its powers
Germany’s authority map depends on use and sector; legal powers are not evidence of a specific inspection.
AI-assisted learning visualization · schematic article explanation
The same model can serve different sectors and purposes. Self-hosting alone does not identify the supervisory authority.
Germany's KI-Marktüberwachungs-und-Innovationsförderungs-Gesetz (KI-MIG) entered into force on 29 July 2026. Section 2(1) names the Bundesnetzagentur (BNetzA) as the default market-surveillance authority for the EU AI Act. Sections 2(2)–(8) allocate particular cases to other authorities.12 Self-hosting a model does not by itself identify which authority has jurisdiction or which AI Act duties apply. The use case, operator role and regulated sector matter.
The authority map
| Case | Competent authority under KI-MIG § 2 |
|---|---|
| General rule, unless an exception applies | BNetzA, § 2(1) |
| Systems linked to products covered by Annex I Section A harmonisation law | Existing product-law market-surveillance authorities, § 2(2) |
| Systems directly linked to regulated financial activity of covered institutions | BaFin or another competent financial supervisor as specified in § 2(3)–(4) |
| Certain law-enforcement, migration, border-management and justice uses | Independent AI Market Surveillance Chamber at the BNetzA, § 2(5) |
| AI systems of Länder public bodies | Authorities designated by Land law, § 2(6) |
| Certain federal tax cases | BNetzA enforcement subject to the consent mechanism in § 2(7) |
| Certain media-service uses | Authorities designated by Land law; Deutsche Welle has a special rule, § 2(8) |
The BNetzA's announcement also names employment, critical infrastructure and education among its areas of responsibility, and describes the arrangement as a hybrid system. It does not make BNetzA the authority for every AI system.
What § 11 permits
KI-MIG § 11 gives the competent market-surveillance authorities the powers in Articles 14(4)–(5) and 16 of Regulation (EU) 2019/1020 and the AI Act. It permits technical assistance by third parties and exercise of specified powers through APIs or other remote-access means. Limits and procedural conditions still apply; for example, § 11(3) restricts the timing of entry under Article 14(4)(e), and § 11(7) specifies when appeals have no suspensive effect.3 The authority may request relevant information, but an information request does not make a postponed high-risk duty immediately applicable.
The general AI Act application date was 2 August 2026. Regulation (EU) 2026/1744 postponed Chapter III, Sections 1–3 for Annex III high-risk systems to 2 December 2027, and for Annex I high-risk systems to 2 August 2028.4 Article 4 AI literacy, Article 5 prohibitions and Article 50 transparency follow their own application dates. Refer to the Omnibus timeline guide for the role and date distinctions.
What to document
A practical file can record the system's intended use; whether you are its provider or deployer; the relevant Annex III or Annex I category, if any; how the current Article 4, 5 and 50 duties are met; and the GDPR basis if personal data is processed. For a potentially high-risk system, plan technical documentation, logs, oversight and conformity work for the applicable deadline. These are preparatory records, not a claim that every Chapter III duty is already enforceable.
Article 50(1) places direct-interaction notice duties on providers of qualifying systems. Article 50(2) concerns provider-side machine-readable marking of synthetic outputs. Article 50(4) concerns disclosure by deployers of certain deepfakes and public-interest text, with stated exceptions.5 A back-office hiring model does not acquire an Article 50(1) direct-interaction duty solely because it handles applications. Article 26(7) workplace information is part of the postponed Chapter III Section 3 duties for affected high-risk systems; it should not be listed as an already applicable AI Act duty in September 2026.4
KI-MIG § 20 provides a non-public national register for high-risk systems under Annex III point 2 (critical infrastructure). Its text sets registration before placing on the market or putting into service for providers, and before use for covered federal public-body deployers.6 It is not a general registry for all self-hosted AI. Whether a particular registration is due requires checking the system's classification and the interaction with the amended AI Act application dates; the existence of the register alone does not settle that question.
What the evidence does not show
The earlier version of this article described a coordinated September 2026 inspection wave, identified agencies and claimed an EU AI Office confirmation on 10 September. No primary authority publication was located that supports those particulars. They have been removed. The BNetzA's legal powers and its July announcement are documented; they are not proof of a particular inspection docket.
Legal status checked on 29 September 2026. This is general information, not individual legal advice.
Footnotes
-
KI-MIG full text, including § 1 and commencement. ↩
-
KI-MIG § 2, paragraphs 1–8. ↩
-
KI-MIG § 11; Regulation (EU) 2019/1020, Article 14. ↩
-
Regulation (EU) 2026/1744, Article 1(40)(b). ↩ ↩2
-
AI Act Article 50, paragraphs 1, 2 and 4. ↩
-
KI-MIG § 20, paragraphs 1–3; AI Act Article 49(5). ↩