Share article
See how other mid-sized companies govern AI
In the Mittelstands AI Village at d:u27, companies show how they’ve built approval processes, AI inventories and training. April 13–14, 2027, Münster.

Why do companies need AI governance?
Most executives ask this for an understandable reason. AI use in their own company looks manageable: a few licenses, a pilot project, some early experience in marketing. So why a whole set of rules?
The answer lies in the gap between usage and rules. According to a recent survey by Bitkom (in German), Germany's digital industry association, 57 percent of companies with 20 or more employees now use AI. For the first time, that's a majority, up from 36 percent a year earlier. On the rules side, Bitkom's survey on shadow AI (in German) from summer 2025 painted a much thinner picture. At the time, only 23 percent of companies had set rules for using AI tools.
At the same time, a quarter of companies knew that employees were using private AI tools at work, and another 17 percent assumed they were. In other words, the usage is already there. In many companies, the framework isn't.
That gap creates three very real problems. First, data ends up in systems nobody has a contract for. Second, nobody can say where AI is actually being used in the company, and therefore which obligations under the AI Act apply. Third, good ideas stall because no one knows who is allowed to approve them.
The third point is the one most often underestimated. In our experience, AI governance is less of a brake than an accelerator. When it's clear who decides and based on which criteria, new use cases make it into day-to-day operations faster. How companies have taken this step is something practitioners share on the Mittelstands Stage of the data:unplugged Festival on April 13 and 14 in Münster.
What is AI governance, and how is it different from data governance?
AI governance is the framework a company uses to steer its use of AI. It defines which AI systems may be used, who decides on new applications, which data may go where, how outputs are checked and how all of this is documented.
AI governance and data governance get mixed up all the time. The two belong together, but they cover different ground. Data governance deals with the data itself: quality, ownership, access rights, definitions. It applies to every report and every system, with or without AI.
AI governance sits one level above that. It asks what an AI system is allowed to do with that data, what it is used for and who is accountable for its results. An example from HR makes the difference tangible. Making sure application documents are accurate, complete and accessible only to HR is a matter of data governance. Whether an AI system may pre-screen those applications, and who is ultimately responsible for the selection, is a question of AI governance.
That's why AI governance without solid data governance stands on shaky ground. Conversely, good data management alone isn't enough once systems start making suggestions, generating text or preparing decisions.
What the EU AI Act requires from companies
For many companies, the EU AI Act is what puts the topic on the agenda, and the number of those affected keeps growing. In the latest Bitkom survey, 37 percent of companies expect to be affected by the AI Act as deployers, up from 23 percent the year before. Of those who are affected or still checking, 89 percent expect a high implementation effort.
One distinction takes a lot of the anxiety out of it. Most mid-sized companies are not providers of AI systems but deployers. They use tools that others have built. The most extensive obligations, such as technical documentation and conformity assessments, lie with the providers.
For deployers, four points matter most. Since February 2025, certain practices have been banned, including emotion recognition in the workplace. The AI literacy requirement has also applied since February 2025. Since the Digital Omnibus came into force on July 27, 2026, it has been worded more softly. Companies must take measures to support the development of their staff's AI literacy, but they are explicitly not required to guarantee a specific level of competence. The obligation to act remains.
Transparency obligations have also applied since August 2, 2026. Anyone who publishes AI-generated images or videos that look real must, in many cases, label them as such. The fourth point concerns high-risk applications, for example AI in recruitment or in evaluating employee performance. Following the postponement, the related deployer obligations, such as human oversight, logging and informing the workforce, apply from December 2, 2027.
We've summarized the timeline and risk levels in detail in our article on AI, data protection and the GDPR. For AI governance compliance, the key point is that obligations depend on the use case. A company has to know where it uses AI, and that is exactly what a governance framework delivers. Since the end of July 2026, Germany's supervisory authority has been the Federal Network Agency (Bundesnetzagentur). Its AI Service Desk offers a self-assessment tool, the AI Compliance Compass, which works well for an initial assessment.
Who decides what AI can do? AI governance roles and responsibilities
This question is at the heart of any AI governance, and mid-sized companies can answer it more leanly than large corporations. You don't need a dedicated committee with its own rules of procedure. You need three clearly assigned roles.
Management
Responsibility for the use of AI lies with the company's leadership and cannot be delegated. Management decides which risks the company is willing to take and rules on cases that go beyond the everyday. That doesn't mean every new application has to land on its desk.
A designated AI lead
Unlike the GDPR with its data protection officers, the AI Act does not require an AI officer. We still recommend clearly naming one person or a small team. This role collects requests, reviews new use cases against fixed criteria, keeps the overview of systems in use and is the go-to contact for all questions. In smaller companies, this often falls to someone from IT or data protection, ideally with dedicated time rather than as a side job.
The business units
Whoever uses a tool knows best what it's good for and where it gets things wrong. The business units bring in use cases and are responsible for checking the results in their area. Where there is a works council, it should be involved early. Many AI tools can track employee performance or behavior, which makes them subject to the works council's co-determination rights. Clarify this at the end instead of the beginning, and you'll lose months.
Implementing an AI governance framework in four building blocks
To start with, a workable framework fits on just a few pages. A good starting point is the guide from the Mittelstand-Digital Zentrum Berlin (in German). It outlines an AI policy in seven steps and points to free templates, such as a sample policy from IHK Schwaben, the regional chamber of commerce. In our view, four building blocks are what matter.
Approvals
New use cases need a fixed process: Who submits the request, who reviews it, and by when is there an answer? A three-tier classification has proven effective. Low-risk applications, such as drafting texts without confidential data, run without individual review. Applications involving internal data need approval from the AI lead. Anything that touches HR decisions or has an external impact also goes to management.
Approved tools
A list of approved tools creates more clarity than a catalog of bans. Each tool should come with a note on which classes of data may go into it, which makes the list the practical core of any AI governance policy. How the tool question ties in with unofficial use, and why bans rarely work there, is something we cover in detail in our article on shadow AI in companies.
Documentation
A simple spreadsheet is enough to begin with. It records which system is used for what, who is responsible, which data is processed and how the risk has been classified. This AI inventory is the basis for mapping the obligations under the AI Act in the first place. It also saves you a frantic search when customers, auditors or regulators come asking.
Training
Rules only work if people know and understand them. That's why training is part of governance, not a separate project. It should differ by role, because someone who approves tools needs different knowledge than someone who writes with them every day. What this looks like in practice for mid-sized companies is the subject of our article on AI training in companies.
Once it's set up, the framework needs regular review, about every six months. Tools change quickly, and the questions change with them. How other companies build their approval processes and AI inventories, and what data foundation they rest on, is something you'll experience first-hand at d:u27 in Münster, across seven different stages and in 80 masterclasses where you work directly on your own framework.
Case study Delivery Hero: governance that doesn't slow you down
For all the structure, one obvious concern remains: Doesn't a framework like this slow teams down? On the Data Stage at d:u26, Juan Carlos Vasquez of Delivery Hero showed why governance doesn't have to be a brake. He was also candid about how the company initially went the other way.
Delivery Hero operates in more than 70 countries, and many of its local brands joined through acquisitions, each with its own tech stack. At first, local teams were largely free to make their own decisions, and headquarters simply provided the tools. The result: personal data often wasn't labeled, each brand had its own idea of what a customer was, and three key dashboards showed three different customer counts. In our experience, many mid-sized companies know this pattern, just on a smaller scale.
Delivery Hero drew a clear conclusion: providing tools isn't enough. You need a single, consistent path where the rules are built in from the start. How that change of course succeeded comes down to three learnings that, in our view, translate well to mid-sized companies.
Automation beats policies
Delivery Hero already had rules for labeling sensitive data. Many people just didn't follow them, because skipping them was faster and the risk only becomes visible once something goes wrong. Today, a central backend automatically checks the rules every time something is published and blocks critical data products if anything is missing, via code for engineers and via a graphical interface for business teams. For mid-sized companies, we think this can look much simpler, for example by only making approved AI tools available through company accounts.
Governance that matches the risk
Delivery Hero sorts data products into tiers. Business-critical ones require data contracts and SLAs, meaning binding agreements on content, quality and availability, while analyses for personal use get by without these controls. To stop anyone from taking the easy route for something important, critical data products may only draw on other verified data products. In our view, the same principle is behind the tiered approval process described above.
Compliant by default
"The easiest path has to be the right one, then everyone will follow it," is how Vasquez sums up the approach. Because the platform delivers the rules automatically, users no longer have to take care of them themselves. Instead of around 300, roughly 3,000 people at Delivery Hero now build data products on their own.
What comes next is especially interesting for the AI question. Because governance is automated, Delivery Hero has all the metadata it needs to let people query data conversationally in the future, in natural language. The tiers apply here too: data products for personal use are excluded so that the AI doesn't work with unverified data. It shows just how closely data governance and AI governance are linked.
More use cases and experiences like this are shared by data and AI experts and practitioners from their own projects on the stages of data:unplugged.
How important is governance in AI projects?
More important than project plans usually suggest. Many AI projects never make it past the pilot stage. Often, what holds them back isn't technical problems but open questions that only surface during the move into operations: Who is liable for errors? Can this data be processed on an ongoing basis? Who checks the results once the project team moves on?
If you settle these questions before the pilot, you avoid the loop afterwards. That's why governance belongs in project planning, as a fixed checkpoint before the decision to roll a project out more widely.
There's a second effect that doesn't appear in any project plan: trust. When it's clear what's allowed, employees use AI openly. They talk about their experiences, share good prompts and report problems instead of hiding them. That way, the company learns where AI actually creates value. Governance that mainly prohibits achieves the opposite.
Conclusion
Clear responsibilities, not a thick rulebook
At its core, AI governance answers the question of who decides what AI can do. For mid-sized companies, that doesn't mean bureaucracy. It means three clear roles and a framework built on approvals, approved tools, documentation and training. The AI Act sets the minimum requirements. The real value comes from clarity in everyday work.
Our recommendation for getting started: Begin by taking stock of which AI tools are already in use today, including unofficially. Then name a responsible person and set the first rules. After six months, refine them. A framework that people actually live by is worth more than a perfect document nobody knows about.
How companies take this path in practice, which approval processes have proven themselves and how they've involved their works councils is at the center of d:u27 on April 13 and 14, 2027 in Münster. Around 17,000 participants come together across seven stages and in 80 masterclasses, with a dedicated Mittelstands Stage and the Data Stage for everyone who thinks about data and AI from the ground up.
AI governance is a team effort between management, IT, data protection and the business units. That's why it pays to come together. Get your tickets for d:u27 now!





















































