Hundreds of thousands of cars with no number plates, no roadworthiness inspection, no oversight: Nobody would dream of allowing that. For AI models, exactly that is the normal state of affairs. Why AI control needs a registration office and how we are building one at SIFAMO.
Imagine hundreds of thousands of cars driving around out there with no number plates, no inspection, no oversight of any kind. Nobody would know which driver did what and when, and in an accident there would be nobody to hold liable. That is exactly what we are seeing with AI tools right now. Models are being thrown out indiscriminately, and nobody is in control anymore. This post is about how we turn that around.
Recorded in Frankfurt am Main, July 2026: AI control explained with the example of vehicle registration.
Why the car is the best example for AI control
You have to register a vehicle at the registration office. It gets a number plate, it has a roadworthiness inspection, it is monitored, and there are rules and guidelines it has to comply with. That is a good thing, because it makes clear who is liable for damage and what happens with the consequences. With AI there is none of that. And yet it makes millions of decisions every day: about loans, about job applications, about insurance, about people. And hardly anyone knows how these decisions come about.
Now imagine it is about your loan decision or your medical records, and the AI accessing them no longer knows any limits with your personal data. Then you have a big problem, and we know exactly where that leads: into complete chaos.
What AI governance means
At its core, AI governance simply means creating guidelines: How can an AI be controlled, and what is it allowed to do in the first place, what is it not allowed to do? That is what we specialise in. What matters to us is that models and tools introduced at banks, insurers or large corporations, for example to comb through case files, are controlled, and that what the AI outputs is actually true.
We all see it with the common GPTs: We type something in, have a presentation generated, get figures back and assume they are correct. But who says they are correct? That is exactly where a control authority has to come in. That this is not a theoretical risk is shown by the case in which OpenAI itself admits that its models invented data in testing (German).
The registration office for AI: our governance layer
Banks, insurers and large corporations have to realise that AI urgently needs to be controlled by a layer that sits above it. We have built that layer and have been expanding it for two years. This is how it works:
Registration. Every AI model registers, like a car at the registration office. No registration, no deployment.
Monitoring. Every step is tracked and documented. You can say precisely how the AI reacts, what comes out and which results it is allowed to produce.
Response. If a model delivers results it should not, you can intervene immediately. Everything stays audit-ready, meaning verifiable for supervisors and internal audit.
The EU AI Act sets the legal framework for this, BaFin supervises, DORA provides the safeguards, MaRisk sets the standards. But rules alone are not enough. It takes technology that implements those rules. What that means for banks in concrete terms, I described in my post on agentic AI governance under BaFin and the EU AI Act (German).
Why we went to Dubai for this
It was no coincidence that we held our AI Summit in Dubai in February last year. We wanted to understand where the journey with AI is heading. Companies like Nvidia, Microsoft and Goldman Sachs were there, and we gained deep insights. The picture that emerged: Banks are introducing AI, insurers are letting models make loan decisions, the human is being taken out, and decisions are made on the basis of old data. And that is only part of it. What about the humanoids that are being built, given an AI and expected to take over human work? Who controls that AI?
That is exactly what we have pursued intensively over the past months, building our own framework for it. We are in talks with banks, large corporations and mid-sized companies about how this layer can be integrated. We keep expanding SIFAMO AI, with a clear goal: AI should not decide about our lives, we decide about AI. Which AI decision in your company could nobody explain today if the supervisor asked? Tell me on LinkedIn or below the video on my YouTube channel.
Frequently asked questions
What is an AI governance layer?
A technical layer that sits between the AI models and the business applications. Models register there, every call is logged, results are checked against rules, and when something looks wrong, you can intervene. The documentation is structured so that it passes an audit.
Which rulebooks apply to AI in banks?
The EU AI Act as the AI-specific law, plus supervision by BaFin, DORA for digital resilience and MaRisk as the minimum requirements for risk management. All of them demand traceability, and that is exactly what a governance layer delivers.
Does this apply to mid-sized companies too?
Yes. A mid-sized company also uses AI to make decisions about customers, applicants or suppliers. Whoever introduces a register and logging today saves themselves the frantic retrofit later, when customers or authorities demand evidence.
Hundreds of thousands of cars with no number plates, no roadworthiness inspection, no oversight: Nobody would dream of allowing that. For AI models, exactly that is the normal state of affairs. Why AI control needs a registration office and how we are building one at SIFAMO.
Imagine hundreds of thousands of cars driving around out there with no number plates, no inspection, no oversight of any kind. Nobody would know which driver did what and when, and in an accident there would be nobody to hold liable. That is exactly what we are seeing with AI tools right now. Models are being thrown out indiscriminately, and nobody is in control anymore. This post is about how we turn that around.
Why the car is the best example for AI control
You have to register a vehicle at the registration office. It gets a number plate, it has a roadworthiness inspection, it is monitored, and there are rules and guidelines it has to comply with. That is a good thing, because it makes clear who is liable for damage and what happens with the consequences. With AI there is none of that. And yet it makes millions of decisions every day: about loans, about job applications, about insurance, about people. And hardly anyone knows how these decisions come about.
Now imagine it is about your loan decision or your medical records, and the AI accessing them no longer knows any limits with your personal data. Then you have a big problem, and we know exactly where that leads: into complete chaos.
What AI governance means
At its core, AI governance simply means creating guidelines: How can an AI be controlled, and what is it allowed to do in the first place, what is it not allowed to do? That is what we specialise in. What matters to us is that models and tools introduced at banks, insurers or large corporations, for example to comb through case files, are controlled, and that what the AI outputs is actually true.
We all see it with the common GPTs: We type something in, have a presentation generated, get figures back and assume they are correct. But who says they are correct? That is exactly where a control authority has to come in. That this is not a theoretical risk is shown by the case in which OpenAI itself admits that its models invented data in testing (German).
The registration office for AI: our governance layer
Banks, insurers and large corporations have to realise that AI urgently needs to be controlled by a layer that sits above it. We have built that layer and have been expanding it for two years. This is how it works:
The EU AI Act sets the legal framework for this, BaFin supervises, DORA provides the safeguards, MaRisk sets the standards. But rules alone are not enough. It takes technology that implements those rules. What that means for banks in concrete terms, I described in my post on agentic AI governance under BaFin and the EU AI Act (German).
Why we went to Dubai for this
It was no coincidence that we held our AI Summit in Dubai in February last year. We wanted to understand where the journey with AI is heading. Companies like Nvidia, Microsoft and Goldman Sachs were there, and we gained deep insights. The picture that emerged: Banks are introducing AI, insurers are letting models make loan decisions, the human is being taken out, and decisions are made on the basis of old data. And that is only part of it. What about the humanoids that are being built, given an AI and expected to take over human work? Who controls that AI?
That is exactly what we have pursued intensively over the past months, building our own framework for it. We are in talks with banks, large corporations and mid-sized companies about how this layer can be integrated. We keep expanding SIFAMO AI, with a clear goal: AI should not decide about our lives, we decide about AI. Which AI decision in your company could nobody explain today if the supervisor asked? Tell me on LinkedIn or below the video on my YouTube channel.
Frequently asked questions
What is an AI governance layer?
A technical layer that sits between the AI models and the business applications. Models register there, every call is logged, results are checked against rules, and when something looks wrong, you can intervene. The documentation is structured so that it passes an audit.
Which rulebooks apply to AI in banks?
The EU AI Act as the AI-specific law, plus supervision by BaFin, DORA for digital resilience and MaRisk as the minimum requirements for risk management. All of them demand traceability, and that is exactly what a governance layer delivers.
Does this apply to mid-sized companies too?
Yes. A mid-sized company also uses AI to make decisions about customers, applicants or suppliers. Whoever introduces a register and logging today saves themselves the frantic retrofit later, when customers or authorities demand evidence.
Related: AI Governance · EU AI Act
Recent Posts
Recent Comments