In most businesses, AI has long been in use, just not on the list: a translation service in the browser, an assistant in the word processor, a plug-in in accounting. Without knowing about it, you can neither protect nor steer it. Governance starts with an honest inventory.
How I work: with source · with approval · in your system.
Starting point
Management wants rules, IT wants no further bans, and the departments have already tried tools. Then comes the EU AI Act with duties that apply in stages. Without a shared picture, the result is either a document nobody reads or nothing at all.
What I do
First I record which AI tools are in use, who uses them and which data goes in. That includes the shadow side: browser extensions, free accounts, local models on private machines.
From that comes a usage policy in plain language: what is allowed, what needs approval, what is forbidden. An approval list names the reviewed tools with purpose, boundary and the person who granted the approval. For critical uses, roles, review steps and a reporting path come on top.
Rollout includes AI literacy training under Art. 4 of the AI Act; the duty has applied since 2 February 2025. It is not a product pitch but a briefing on opportunities, limits and duties in each role.
Result
You get a tool list including shadow AI, a usage policy, an approval list and a training record. Plus a review rhythm so the list does not age within a quarter.
Example
The topics come from daily work: access rights for agents, tools with write access to a production database, data leaving through browser extensions. I wrote up two cases because they show the core: access rights for AI agents and an agent that empties a database. A workshop walks through the duties and the typical mistakes.
Experience from a manufacturer engagement adds to this: knowledge as a vault, binding answer rules, a test catalogue of domain questions. The engagement stays anonymous, no client name, no figures.
Limit
Governance is not a one-off document. Without a person maintaining the list and consequences for violations, it stays paper.
This is not legal advice. The legal assessment of a concrete use case, whether a system counts as high-risk and which duties follow, belongs with a law firm. I deliver the technical and organisational basis that such an assessment can build on.
How it runs
- Inventory: tools, users, data, including the unreported ones
- Assess risk and define approval classes
- Write policy and approval list, align with departments
- Run AI literacy training, collect questions
- Set the review date and owners
Common questions
When does the AI Act apply?
The AI literacy duty under Art. 4 has applied since 2 February 2025. Further duties apply in stages, including for certain systems and for transparency. What applies to your use depends on your role and purpose.
What is shadow AI?
AI tools used without IT’s knowledge. They are usually well intentioned and still a risk, because data can leave the building.
How long should training be?
Two to three hours are enough for a first briefing. What matters is that it covers your own processes, not just quotes from the regulation.
Is a policy enough?
No. It needs an approval list, a reporting path and someone maintaining both. Otherwise nobody knows what is allowed after three months.