AI oversight in Germany: the Bundesnetzagentur is in charge
Since late July 2026 Germany has an AI supervisor. What the Federal Network Agency really offers, what its Compliance Compass does and what it cannot replace.
In short
Since 29 July 2026, Germany’s Bundesnetzagentur has been the market surveillance authority for the EU AI Act under Section 2(1) of the KI-MIG. It runs an AI Service Desk and the AI Compliance Compass, a free online tool for initial orientation. Both are expressly non-binding. Orientation helps, but it is not evidence that holds up in an audit.
Auf Deutsch lesen: deutsche Fassung
Update of 20 August 2026. When I wrote this text, the Digital Omnibus had been adopted but not yet published. It is now law in force: Regulation (EU) 2026/1744 of 8 July 2026, in the Official Journal of 24 July 2026, in force since 27 July 2026. That changes nothing about the substance, but it does change how binding it is. I am leaving the text as it stands and putting the reference above it.
For two years, the most frequent follow-up question in my conversations about the EU AI Act was the same: and who actually checks that here? For a long time the honest answer was: nobody yet, it has not been settled. Since 29 July 2026 that answer is no longer true. Germany has an AI supervisor, it sits at the Federal Network Agency (Bundesnetzagentur), and it has brought a free tool with it.
I looked into this because I had to know, not because it is a pleasant topic. And I am writing about it because two stories are currently doing the rounds in my circles, and neither is true. One says the great wave of bureaucracy is now coming. The other says there is now a tool from the authority, so the matter is dealt with. Both are wrong, and the second mistake is the more expensive one.
What actually happened on 29 July
On that day the German AI Market Surveillance and Innovation Promotion Act came into force, KI-MIG for short, promulgated as Federal Law Gazette (BGBl.) 2026 I No. 223. It is the German implementing act for the EU AI Act, and it answers exactly the question that was open before: who is responsible.
The answer is in Section 2(1): the Bundesnetzagentur, unless the act provides otherwise. On top of that come four roles that you should keep apart:
- Market surveillance authority, meaning the body that can actually inspect and intervene.
- Coordination and competence centre, known internally as KoKIVO, which supports other authorities.
- Single point of contact for questions about the EU AI Act.
- Central complaints office, which affected persons can turn to in the event of infringements.
In addition, the act sets up an independent AI market surveillance chamber at the Bundesnetzagentur, responsible for the areas that are sensitive in terms of fundamental rights: law enforcement, migration and border control, justice and democratic processes, and biometric systems where they are used in these areas. And the sectoral supervisors stay where they are. For the regulated financial sector that is BaFin, the Federal Financial Supervisory Authority. For medical devices it is the federal and state market surveillance authorities already responsible for them. On cybersecurity questions the BSI, the Federal Office for Information Security, is involved for the time being. For data protection it is the data protection supervisory authorities. The Bundesnetzagentur has the residual competence. This is not a new authority built on a greenfield site, but an assignment to existing ones.
What I find notable is the tone the authority itself strikes. At the launch its president said it sees itself “as an enabler and point of contact for the entire AI ecosystem”. You do not have to believe that. But you should take note of it before you use the word bureaucracy.
What the authority offers you free of charge
The AI Service Desk has been online since 3 July 2025 and is expressly aimed at small and medium-sized enterprises and start-ups. There is an FAQ section, a contact form for everything that is not covered there, and factsheets to download, one of them specifically for SMEs.
The really interesting piece is the AI Compliance Compass. It is not a PDF but a tool in the browser, and it is meant to help you with the question of which risk level your AI system falls into. Free of charge, without registration, without anyone trying to sell you anything.
I say this clearly because it is rarely said: this is a good offer. Anyone who works with it for half an hour understands the structure of the regulation better than after the third webinar. Anyone who dismisses it as bureaucracy has not used it.
A free tool from your own supervisory authority is not a bureaucracy problem. Not using it is one.
And what it expressly is not
Now comes the part that interests me as an auditor. The authority itself writes that the Compass offers “an initial orientation”, and that the information is provided “for orientation and is non-binding”. That is not a modesty formula. It is a precise statement about how far the tool reaches.
Orientation is not evidence. The difference is the same as between a glance at the machine and an inspection record. In concrete terms, after using the Compass you lack everything that would hold up in an audit or before an authority:
- The reasoning. Why is your system not high-risk? A result without the argument behind it is worthless when in doubt. The question asked is not about the result but about how you got there.
- The state of things. Which description of your system is the classification based on, in which version, as of which date? Systems change, and a classification without a reference point ages unnoticed.
- The responsibility. Who made the classification, who checked it, who stands behind it? A result without a name is an opinion.
- The retention. Where will the result be in a year when someone asks for it? A browser window is not a filing location.
This is not a criticism of the authority. It says itself what its tool does. The mistake arises on the other side, namely when someone prints out the result and considers the matter closed. I see the same pattern with screenshots submitted as evidence, and with verbal information that is supposed to count as proof. A pointer is turned into proof because it is convenient.
Do it the other way round: use the Compass as a starting point and then write down what it does not deliver. Then within two hours you have something that holds up, instead of a screen view that does not.
Regulatory sandboxes: this is where most people get it wrong
Almost every report on this topic mentions AI regulatory sandboxes, and it regularly sounds as if you could sign up tomorrow. That is not true, and I checked it twice because I had it wrong in my own head.
No AI regulatory sandbox is currently operational in Germany. There was a simulation project by the Bundesnetzagentur together with the Hessian Ministry for Digitalisation and the Federal Commissioner for Data Protection, which ended in March 2026 and was meant to test the procedures. Alongside that, a European pilot programme is running. A real national sandbox will come later.
And the deadline is a different date from the one many people have in mind. Originally the AI Act said 2 August 2026. The amending regulation from the Digital Omnibus package has turned that into 2 August 2027. Anyone who writes today that Germany has missed the sandbox deadline is calculating with a date that no longer applies.
I had the sandbox date wrong in my own head. That is exactly why I check figures at the source and not in the summary.
The deadlines, so that no alarm arises here
Because this topic is often sold with urgency, here is the sober position. The obligations for high-risk systems under Annex III apply from 2 December 2027. For AI in products that are already subject to other EU legislation, that is Annex I, from 2 August 2028. Both dates were pushed back by the amending regulation.
That does not mean nothing applies until then. The transparency obligations under Article 50 are already running, and the supervisory structure is now in place. It means: you have time that you can use properly, instead of building the wrong thing in a rush.
One more sentence on fines, because people often mix up who imposes what. The well-known amounts in the millions are in Article 99 of the EU AI Act itself. The German act essentially assigns the competence and, alongside that, contains its own, considerably smaller framework for certain cooperation and documentation obligations. Anyone who comes to you with the big numbers should be able to say, when asked, which obligation they mean and from when it applies to you.
This is my assessment as an auditor and managing director, not legal advice. For a binding assessment of your case you need someone who is liable for it.
What I would do this week in your place
Not panic, not nothing. The effort for the first sensible step is one to two hours, and it depends on a question that has nothing to do with law: which AI are you actually using?
In the companies I see, the answer is almost never complete. The pre-selection of job applications runs through a tool that was bought in HR, and management does not know about it. A vision inspection sorts parts by surface defects and is regarded as a machine, not as AI. In the ERP, a function suggests order quantities, and nobody calls that AI because it has always been there. All three can fall under the regulation, and for the pre-selection of applications that is even likely.
So start with the list, not with the classification. A list that nobody has is the real finding. After that, take the Compass, go through your systems and write down for each result what it is based on and who is responsible for it.
And then leave it and look at it again in three months. If your list is still correct then, you have a process. If not, you know what you need to work on. Either is worth more than any certificate you could buy today.
Be honest: could you present a complete list of your AI systems this week?
Primary sources
A first assessment of your AI systems, in six steps
List before classification
First write down which AI is actually running in your organisation. Ask expressly about purchased functions in HR, sales, production and ERP. Otherwise anything nobody calls AI slips through.
Clarify the role per system
Record whether you provide the system yourself or only deploy it. The obligations differ considerably, and most mid-sized companies are deployers, not providers.
Go through the Compass
Work each system once through the AI Compliance Compass of the Bundesnetzagentur. It costs nothing and puts the structure in order. Take the result as a starting point, not as the final state.
Add the reasoning
For each result, note which system description it is based on, in which version and as of which date, and why the classification turns out as it does. Without these four details the result cannot be verified.
Name the people responsible
For each classification, write down who made it and who checked it. A classification without a name is an opinion and holds up in no conversation with an authority.
Set a review date
Set a date in three to six months and check whether the list is still correct. Systems change, providers add functions, and a classification without a review date goes out of date without anyone noticing.
Frequently asked questions
Who is responsible for the EU AI Act in Germany?+
Under Section 2(1) of the KI-MIG, in force since 29 July 2026, the Bundesnetzagentur is the competent market surveillance authority, unless the act provides otherwise. Sectoral competences remain in place, for example BaFin for the regulated financial sector and the market surveillance authorities for medical devices. The Bundesnetzagentur has the residual competence.
Is the AI Compliance Compass sufficient as evidence for an audit?+
No. The Bundesnetzagentur itself describes it as an initial orientation and expressly labels the information as non-binding. For robust evidence you lack the reasoning behind the classification, the reference to a specific system version with a date, a named responsible person and a filing location where the result can still be found in a year.
Can I already sign up for an AI regulatory sandbox?+
Not in Germany at present. There was a simulation project by the Bundesnetzagentur with the Hessian Ministry for Digitalisation and the Federal Commissioner for Data Protection, which ended in March 2026, and a European pilot programme is running. The deadline by which at least one national sandbox must be operational is 2 August 2027 under the amending regulation.
From when do the obligations for high-risk AI apply?+
For systems under Annex III from 2 December 2027, for AI in products under Annex I from 2 August 2028. Both dates were pushed back by the amending regulation from the Digital Omnibus package. The transparency obligations under Article 50 already apply, regardless of that.
Author & expert review: Lars Zimmermann · ISO/IEC 42001 Senior Lead Auditor & Senior Lead Implementer · ISO/IEC 27001 Lead Auditor & Lead Implementer (PECB)
Last updated: 10 November 2026. Researched and reviewed to the best of our knowledge; not a substitute for individual legal advice.
Questions about your own case?
In a free 15-minute intro call we assess where you stand on ISO 42001, ISO 27001 and the EU AI Act, honestly and without a sales pitch.