AI Supervision Started Before the Deadline

On Monday the EU softened the AI Act’s literacy duty. On Wednesday Germany’s financial regulator announced it would supervise it.

Two days. That gap tells you more about how AI supervision will actually work than the Omnibus headlines did, because national regulators are not waiting for Brussels to finish simplifying. BaFin has powers now, a sampling method, and a fine ceiling of €35 million.

AI supervision started before the high-risk deadline

Germany’s KI-Marktüberwachungsgesetz, the national AI market surveillance act, builds the AI supervision architecture the AI Act requires every Member State to put in place. It makes the Bundesnetzagentur the central AI authority from August 2026, and it hands financial services to BaFin.

BaFin confirmed on 29 July that it has begun. It will monitor transparency obligations, prohibited AI practices and measures promoting employee AI competence across banks and insurers. Named use cases include customer chatbots, creditworthiness checks, credit scoring, and pricing in life and health insurance.

Now read that against the Omnibus. It pushed high-risk obligations for standalone Annex III systems out to 2 December 2027, so the sequencing looks odd. Credit scoring is an Annex III high-risk use. Its documentation duties are eighteen months away. Yet AI supervision of the transparency and prohibited-practice rules covering the same systems is live.

Nothing contradictory is happening here. The deferral moved one chapter of the AI Act, not all of it. But an organisation reading only the delay coverage would conclude it had two years. In Germany’s financial sector it does not.

Two supervisors, one system

The more useful detail is structural. BaFin’s remit covers AI used in regulated activities. AI used in functions that are not regulated financial activities, human resources being the example BaFin itself gives, falls to the Bundesnetzagentur instead.

So the supervisor is determined by what the system does, not by what the system is.

Where AI supervision splits by use, not by tool

Take one large language model licensed once and deployed twice. Pointed at customers, answering questions about a loan product, it sits inside BaFin’s perimeter. Pointed at applicants, ranking CVs for the same bank’s graduate intake, it sits inside the Bundesnetzagentur’s. Same contract, same vendor, same model. Two regulators, with different priorities and different inspection habits.

That has a practical consequence for anyone keeping an AI inventory. AI supervision asks what a system is used for and by whom. An inventory organised by system cannot answer that. One organised by use case can.

It also revives a point from earlier this week. Once AI work spreads into HR, legal and operations, the question is not only who is competent to oversee a system. It is which authority comes asking, and whether anyone in that function knows.

The literacy duty nobody expected to be supervised

Here is the sharpest part. Among the things BaFin says it will monitor is whether firms take measures to promote AI competence among employees.

That is Article 4. The Digital Omnibus, in force since 27 July, softened it. How far remains contested: the Commission’s own summary describes the company-level requirement as replaced by non-binding encouragement, while legal commentators read the amended text as a weaker duty that survives, obliging organisations to support literacy rather than guarantee a level.

Either way, the direction was towards less pressure from Brussels. BaFin went the other way inside forty-eight hours.

The lesson is not that the Omnibus was meaningless. It is that a softened duty at EU level can still be a supervised expectation under national AI supervision. That is especially true in a sector whose regulator already inspects governance, training and control functions by routine. Sectoral supervisors hold their own instruments. They do not need the AI Act to be maximally strict before asking a question.

What AI supervision looks like in practice

BaFin has been unusually clear about method. That is worth reading closely, because it sets expectations for how AI supervision may work in other Member States.

Jens Obermöller, the regulator’s AI specialist, described the approach as pragmatic and risk-based, adding that BaFin will “review a sample of the AI applications that are used by many financial entities”. So this is not a systems audit. AI supervision here means market monitoring, targeting the tools with the widest footprint across the sector.

There is a strategic implication in that. Widely adopted vendor tools attract AI supervision that bespoke internal systems do not. Reviewing one popular product tells the regulator something about many firms at once. Concentration in the vendor market becomes a concentration of regulatory interest.

The €35 million number

On enforcement, Obermöller confirmed that fines can reach €35 million or 7% of annual turnover in extreme cases. He added that early engagement should keep such penalties exceptional.

That ceiling is the AI Act’s prohibited-practices tier, and it is the highest in the regulation. It applies from now rather than from December 2027.

What to have ready

None of this requires a new programme. It requires that four things be answerable on request.

• Which of your AI systems support a regulated activity, and which support functions like HR, because that decides which AI supervision applies
• Whether your AI inventory is organised by use case rather than by system or licence
• What evidence exists that staff operating these systems have been given competence measures, whatever the Omnibus did to the binding standard
• Which of your AI tools are widely used across your sector, since those are the ones a risk-based sample is most likely to reach

The pattern worth noticing is that AI supervision is arriving through existing supervisors rather than through a brand new AI regulator in each country. Financial firms already know how BaFin asks questions. Only the subject matter is new.

Simplification at the centre does not mean quiet at the edges. It means the pressure moves, and it usually moves closer to whoever already had your file open.

If you want to see where your own evidence gaps sit, the Future Prep free assessments and checklists are built for that kind of check.