The Auditor Is Coming: What the First EU AI Act Audits Now Mean for Your Business

Four numbers are enough to understand the current situation: August 2, 2026, December 2, 2027, August 2, 2028 – and somewhere in between, an auditor knocking on your door. The Digital Omnibus simplification package has pushed back the deadlines for most of the EU AI Act's high-risk obligations. Anyone who concludes from this that the topic can wait is reading the news wrong. Because while the big deadline moves further out, other obligations have long been in force – and the first market surveillance authorities are already starting to look more closely. What that concretely means for your business is what we break down in this article.

The Delay Is Real – and Why Audits Are Already Happening Anyway

To set the context, let's start with the facts themselves. Since August 2, 2026, most of the high-risk obligations from Annex III of the EU AI Act have in principle applied. The Digital Omnibus simplification package, which according to current knowledge has been in force as Regulation (EU) 2026/1744 since July 27, 2026, pushes these obligations back for two large groups of systems, however. Standalone Annex III high-risk systems – for example in biometrics, critical infrastructure, education, employment and recruiting, law enforcement, and migration and border control – get time until December 2, 2027. High-risk systems embedded in products under Annex I, for instance in machinery or medical devices, don't need to be compliant until August 2, 2028.

That's genuine relief, no question. Many mid-sized businesses that had just started mapping their Annex III systems gain valuable months as a result. Only: the AI Act doesn't consist solely of Annex III obligations. Several central obligations remain completely untouched by the delay, because they were never tied to the deadline that's now been pushed back. These include the transparency and labeling obligations under Article 50, the obligations for providers of general-purpose AI models, which have already applied since August 2025, the ban on certain AI practices under Article 5, which has been in force since February 2025 and will be extended to two further prohibited practices from December 2, 2026, as well as the AI literacy obligation under Article 4, which has likewise applied since February 2025 and was tightened in July 2026. Anyone who defers their entire compliance budget to 2027 is overlooking part of the bill that's already due.

This is exactly where the current development comes in that gives the topic its urgency. At the European level, the AI Office received expanded sanctioning powers in September 2026 – reports are dated around September 1 and September 14–15 – and sent formal requests for information to several providers of general-purpose AI models. According to the available reports, the focus was mainly on three topics: securing the models, independent external testing, and post-market monitoring. The same sources also report that no violations have been identified so far. Important for context: these specific details come from German-language business and trade portals, not from a directly locatable official EU press release. Before citing these figures in your own documents, it's worth checking the primary source at the European AI Office.

Regardless of the exact number of requests, though, the episode shows something fundamental: the responsible authorities are no longer in observation mode. They're asking specifically for documents, test evidence, monitoring processes. And Spain already provides a practical example that goes beyond mere announcements. With AESIA, the Agencia Española de Supervisión de la Inteligencia Artificial based in A Coruña, the EU has had its first standalone national AI supervisory authority since 2024. According to available reports, by mid-2026 it had already opened preliminary investigations into AI systems at Spanish organizations and published around 16 compliance guides. That's not an isolated case at the edge of the continent, but a signal of the direction in which supervisory practice is developing across the entire EU.

The auditor doesn't ask whether your company is already fully compliant. They ask whether you're able to show, within a reasonable time, what you're working on – and who at your company is responsible for it.

The Central Basis for Audits: Article 11 and Technical Documentation Under Annex IV

When an auditor actually shows up, they typically don't ask first about marketing claims or statements of intent. They ask for the technical documentation. Article 11 of the EU AI Act obliges providers of high-risk systems to create this documentation before market launch and to keep it up to date throughout the system's entire lifecycle. Its content is set out in Annex IV across nine sections in total: a general description of the system, information on development and design, information on monitoring and control of the system, concrete performance metrics, the risk management system used, tracking of relevant changes across the lifecycle, the standards applied, the EU declaration of conformity, and a post-market monitoring plan.

This documentation is exactly what market surveillance authorities and notified bodies review during an inspection. It's therefore the actual basis for any audit – not a general self-declaration, but a structured, traceable dossier. And this is the practical core of our article: even though the formal deadline for standalone Annex III systems doesn't kick in until 2027, authorities generally expect readiness to provide evidence when they make an inquiry, not a finished binder only by the deadline. Documentation you assemble under time pressure because a request is already sitting in your inbox tends, in our experience, to be more expensive, more incomplete, and more stressful than documentation that has grown alongside your operations over months.

Who Actually Comes Knocking in Germany: the Bundesnetzagentur, Not a Patchwork

One of the most common uncertainties in the Mittelstand concerns a simple question: who's actually responsible once things get concrete? In Germany, the Act on Market Surveillance and Institutionalization in the Field of Artificial Intelligence, known as KI-MIG for short, has designated the Bundesnetzagentur (Federal Network Agency) as the central AI market surveillance authority since July 2026. The approach is hybrid: existing sector-specific supervisors, for example in the financial sector or healthcare, remain responsible for their respective sectors. The Bundesnetzagentur takes over the areas for which there has so far been no suitable sectoral authority, and additionally acts as a coordination and competence center as well as the central point of contact for European institutions, in particular the AI Office.

In very practical terms, that means: before a request arrives, it's worth clarifying which authority would be responsible for your specific system, should it come to that. If, for example, you operate an AI-supported credit scoring model within a financing division, a sectoral supervisor may be responsible. If, on the other hand, you run a recruiting tool with no direct link to an already-regulated industry, responsibility tends to fall to the Bundesnetzagentur. This upfront clarification takes little effort, but if it comes to that, it gives you a decisive head start, because you know who to approach and what that body typically expects from documentation.

Also important here is the role the Bundesnetzagentur plays as a coordination and competence center. It's not just an executing oversight body, but also the interface to the AI Office at EU level. Questions that start at the European level – for instance in connection with the GPAI requests for information – can therefore, sooner or later, have national repercussions if a system operated in Germany is built on an affected base model. Anyone who knows their own AI model supply chain and documents which provider supplies which model is therefore prepared not only for a direct inquiry from the Bundesnetzagentur, but also for the case where a request lands on your desk via the detour of an upstream model provider.

The Costly Gray Zones in the Mittelstand: Recruiting and Creditworthiness Checks

Two Annex III categories affect the German Mittelstand especially often, without many decision-makers having this on their radar. The first is recruiting and applicant screening. AI is considered high-risk-relevant if it targets job postings, analyzes or filters applications, evaluates candidates, or supports decisions on promotion, termination, task allocation, and employee monitoring. One point that's crucial here and regularly underestimated in practice: according to the current legal interpretation, it makes no difference whether the AI makes the final decision or merely pre-sorts and ranks for a human reviewer. Even a pure pre-sorting tool, where a human always makes the final decision, counts as high-risk-relevant as soon as it evaluates or prioritizes candidates.

The second category is the creditworthiness assessment of natural persons under Annex III, point 5(b), for example in the form of scoring systems in retail banking. An exception applies here only to systems used exclusively for fraud detection. This, too, is easy to overlook: even upstream scoring models that merely provide a recommendation to a caseworker fall under the high-risk obligations according to expert interpretation, even though a human makes the final decision here as well.

Both examples show a pattern that points beyond the specific categories: the legislator ties the classification not to the final decision-making authority, but to the influence an AI system has on a decision about a natural person. Anyone who reassures themselves with the argument "a human makes the final call anyway" is thereby systematically underestimating how many of the tools already in use actually fall within scope.

In practice, this means: both HR and any financing or leasing division you may have should each keep their own small inventory, independently of one another. In HR, a simple look at the recruiting software in use and the question of whether it filters, ranks, or evaluates applications is often enough. In the financing division, it's worth examining every scoring or credit model that feeds into granting loans, leasing contracts, or installment plans to private individuals – even when the model comes from an external provider and is only integrated into your own software as an upstream recommendation.

What a Fine Really Costs – and What the SME Rule Actually Means

That brings us to the question that ultimately drives most conversations: what's at stake? Article 99 of the EU AI Act provides for a three-tier fine system. For prohibited AI practices under Article 5, fines of up to €35 million or 7 percent of global annual revenue are possible. For violations of the high-risk obligations, which include Article 11 documentation, as well as for violations by GPAI providers, it's up to €15 million or 3 percent of annual revenue. For misleading or false information provided to authorities, the range is up to €7.5 million or 1 percent of annual revenue.

For large corporations, the higher of the two values applies at each of these tiers. For small and medium-sized enterprises within the meaning of EU Recommendation 2003/361/EC – up to 250 employees, up to €50 million in revenue, or up to €43 million in balance sheet total – the lower value applies instead. A sample calculation from consulting practice makes the difference tangible: with annual revenue of €10 million, the maximum possible sanction at the middle tier would work out to €300,000 rather than €15 million, because here the percentage, not the fixed amount, sets the ceiling.

That's noticeable relief compared to the "€35 million fine" headline. But it's explicitly not a blanket exemption from the substantive obligations. The SME rule changes the ceiling on the sanction, not the existence of the obligation itself. And documented compliance efforts, an early voluntary disclosure, and cooperation with the responsible authority, according to advisory sources, have a noticeably mitigating effect on the specific fine assessment. Anyone who can show, if it comes to that, that they had already engaged with the topic in a structured way before the inquiry is noticeably better off than someone who only becomes active after the letter arrives.

The Roadmap for the Coming Weeks

From all of this, a realistic roadmap emerges that can be implemented with manageable effort. First: create a complete inventory of all AI tools used in your business and check each one against the Annex III categories – not just the obvious cases, but also pre-sorting and recommendation tools. Second: for every potentially high-risk-relevant tool, request Annex IV-compliant technical documentation from the provider, or produce it yourself – with a system description, risk management, the origin of training data, the human oversight process, and logging. Third: clarify the question of jurisdiction between the Bundesnetzagentur and any possible sectoral supervisor before a request arrives. Fourth: document your AI literacy measures under Article 4 in a verifiable way, because this obligation already applies now, regardless of the high-risk deferral. And fifth: check where you use content subject to labeling requirements under Article 50, and make sure that labeling actually happens.

None of these five steps requires you to be already fully compliant by December 2, 2027. They only require that you can start today to show that you've taken a structured approach. In our experience, this is exactly what distinguishes companies that can calmly respond to a first inquiry from an authority from companies that panic the moment the first letter lands in their inbox.

Based on our current research, we are not aware of any publicly confirmed, concluded audit cases at German mid-sized companies – the first structured audits are only just getting underway, already visible in Spain, at the EU level via the AI Office's GPAI requests, and in Germany via the newly established framework of the Bundesnetzagentur. Anyone starting now to build up their Annex IV documentation is therefore not acting hastily, but simply in good time.