How it works today in most companies
On the IT department's network drive there is a file called “AI-Tools.xlsx”. It was created in spring 2025, when the first articles on the AI Act appeared in the trade press. It has seven rows: ChatGPT, DeepL, the Copilot in the office suite, the translation module in the online shop and three tools nobody remembers adding. There is also a “Risk” column. Five rows say “low”, two say nothing.
Day to day, the list plays no role. Sales has ChatGPT rephrase quotes, using private accounts because the company license is still pending. Marketing is testing an image generator. Accounting uses receipt recognition that the vendor describes as “AI-powered”, and nobody connects it to the list. Whether the company is an operator of these tools, or has already become a provider by adapting a model, nobody has checked.
In fall 2025, management wrote in an all-staff email that AI could be used “only with approved tools”. Which tools were approved was not in the email. Since then, the “privacy@” mailbox occasionally gets questions about whether a particular tool is allowed. Each one is answered individually by email, and none of them makes it into the list.
The AI literacy obligation under Art. 4 of the AI Act was dealt with in February 2025 with a PowerPoint. Twelve slides, sent to all employees, subject line “Important: how we use AI”. Nobody knows who opened the file. New colleagues never receive it because it is not part of onboarding.
Then HR buys a tool that pre-sorts incoming applications and scores CVs for fit. The vendor gave a convincing demo, the contract runs through the HR budget, IT sets up the login. The data protection officer learns about it from a question about the data processing agreement. Nobody mentioned at the time of purchase that recruitment is one of the areas in Annex III of the AI Act to which the obligations for high-risk AI systems are tied.
The moment it becomes apparent
The question usually comes from outside or from the works council. A major customer sends its supplier questionnaire, which now has a section on the AI Act: “Which AI do you use, in which role, and how is it classified?” Or the works council hears about the recruiting tool and wants to know which criteria it scores on and who checks the results.
The Excel list answers neither question. The recruiting tool is missing, the “Risk” column holds an assessment without reasons, and there is nothing on the role. Whoever answers now starts from scratch, with two weeks until the questionnaire is due.
The analysis
Five Gaps in the AI List
They appear even when the list is kept carefully. A list records which tools exist, but not what follows from them.
View the AI Act module in preecoHidden AI is missing from the inventory
The list contains what someone reported. AI features in existing software, such as receipt recognition or text suggestions, are missing because nobody sees them as a separate tool. Every answer to an outside party is incomplete before it is written.
The role is unclear
Whether the company is the operator, provider or importer of a system determines which obligations apply. As long as that is open, nobody can say which obligations fall away either.
Classification without reasons
“Low” in a cell is a claim. Anyone who has to explain to a customer or the works council why a system does not fall under Annex III needs the assessment path behind it and has to find it again later.
Obligations without an owner
Even a correct classification implements nothing. Who fulfills the obligations as operator or writes the transparency notices under Art. 50 is recorded nowhere, and open items stay open until someone asks.
AI literacy without evidence
A slide deck sent by email does not show who read it. New employees slip through, and the only answer to a question about measures under Art. 4 is an email thread.
The target process in six steps
The AI Act does not prescribe a tool. It assumes that you know which AI you use or offer and which obligations follow from it.
1. Record AI systems where your systems already are. Most AI tools are already documented as data processing systems because they process personal data. In preeco | data protection you mark the system as containing AI components. It then appears under AI systems, with protection needs, data location, data processing agreements and processing activities.
2. Ask specifically about hidden AI. Ask departments about features in existing software: text suggestions, automatic sorting, scoring, chatbots in customer service.
3. Create an AI compliance check per system. The check runs through five steps: exemptions from the AI Act, scope of the AI Act, role, risk classification and result. If someone changes an earlier answer, the application asks for confirmation and resets the steps that follow. The check structures the assessment and people decide; the application states that its content is not legal advice.
4. Derive requirements from role and risk class. On saving, the check generates the compliance requirements that match the selected role and risk class, for a high-risk AI system, for example Art. 9 to 15 for providers or Art. 26 for operators (the deployers in the regulation's terms). Each requirement gets a status from not started to implemented or not applicable.
5. Assign and link the work. For each requirement you create tasks and assign them to people. Measures such as access control, logging or documentation of the AI models are recorded as AI technical and organisational measures. The check links the system to the processing activities it is used in; if one of them needs a DPIA, it sits on the same processing activity.
6. Treat AI literacy as an ongoing measure. AI literacy courses run through the training module of preeco | data protection, with progress per person and a certificate of completion.
Which deadlines apply
As of November 2026, the prohibitions under Art. 5 and the AI literacy obligation under Art. 4 have applied since February 2, 2025, the obligations for general-purpose AI models since August 2, 2025, and the transparency obligations under Art. 50 since August 2, 2026. The Digital Omnibus, Regulation (EU) 2026/1744, postponed the high-risk obligations: for stand-alone systems under Annex III, such as recruitment, they apply from December 2, 2027 at the latest, and for systems embedded in regulated products under Annex I from August 2, 2028. In its new wording, Art. 4 requires providers and operators to take measures to support the AI literacy of their staff. It prescribes no specific level and no certificate. In Germany, the KI-MIG governs supervision, with the Federal Network Agency (Bundesnetzagentur) as the central market surveillance authority.
Before and after in direct comparison
| Criterion | Before: AI list in Excel | After: AI systems with checks |
|---|---|---|
| Inventory | Contains what was reported | All systems with AI components from the records |
| Role | Not clarified | Set per system, for example operator or provider |
| Classification | “Low” without reasons | Five-step path to the risk classification |
| Obligations | Unknown | Compliance requirements by role and risk class |
| Implementation | Depends on individuals | Status per requirement, tasks with owners |
| AI literacy | A slide deck sent by email | Courses with progress and certificate |
| Link to data protection | Separate from the records | Linked to processing activities and DPIA |
| Answers for customers and works council | Pieced together for each request | Overview of checks with role, classification and progress |
In practice
This Is What It Looks Like in preeco | data protection
The three building blocks that carry the process described here.
AI Systems and AI Compliance Checks
Systems with AI components appear in a view of their own. Per system, a check in five steps leads to role and risk classification and generates the matching compliance requirements.
Tasks per Requirement
For each compliance requirement you create tasks with an owner and a due date. The status of the requirements shows in the overview how far implementation has progressed for each check.
AI Literacy Training
Online courses for users, decision-makers and data protection, IT and compliance staff. Progress is visible per person, and passing the final test produces a certificate of completion as a PDF.
What the switch means in practice
In mid-sized companies, implementation rarely starts from zero. Most AI tools are already in the records as data processing systems, they are just not marked as AI. The real work lies in the checks. For a translation service the company uses as an operator, the check is done quickly. For a tool in an Annex III area, such as the recruiting tool, it takes conversations with HR, the vendor and the works council. The obligations for high-risk AI systems under Annex III apply from December 2027 at the latest. Checks like this one belong at the start.
Three mistakes that make implementation unnecessarily hard
Asking only about chatbots. AI is more often found in software that is already in use than in new tools. An email asking “Who uses ChatGPT?” will not find the receipt recognition in accounting.
Taking the role for granted. A company that builds a purchased model into its own product, or passes an AI tool on to customers, may no longer be only an operator. Step 3 of the check asks this question for every system, and the answer changes the list of requirements.
Mistaking the check for a legal opinion. The AI compliance check organizes the assessment and records who decided what. A borderline case in an Annex III area calls for a legal assessment, and the decision is made by your company.
How to tell that it is time
A list of AI tools is a good start. It should become a process once one of these points applies:
- A customer or auditor asks about your AI systems and their classification.
- The works council wants to know which AI is used in HR.
- Your list has no column for your company's role.
- A department has introduced a tool for recruitment, education or credit checks.
- You cannot say who received which AI literacy measure.
Start with the system someone asked about most recently.
FAQ
Frequently Asked Questions About Implementing the AI Act
The inventory is the basis for everything else: without it, neither role nor risk class can be determined, and the obligations depend on both. In preeco | data protection the inventory comes from the data processing systems: systems marked as containing AI components are gathered under AI systems.
Following the Digital Omnibus, Regulation (EU) 2026/1744, the obligations for stand-alone high-risk AI systems under Annex III, such as in recruitment, apply from December 2, 2027 at the latest. For AI in regulated products under Annex I, the date is August 2, 2028. Prohibitions, AI literacy and the transparency obligations under Art. 50 already apply.
Providers and operators take measures to support the AI literacy of their staff. No specific level of competence and no certificate are prescribed. Recording who received which measure is a sensible form of evidence. preeco | data protection includes online courses for this, with progress per person and a certificate of completion.
That depends on the individual system. A company that uses a purchased AI system under its own authority is usually the operator, called deployer in the regulation. A company that develops a system, or has one developed, and offers it under its own name is the provider. The AI compliance check asks for the role per system and derives the requirements from it. It does not replace legal advice.
The German act on AI market surveillance and innovation (KI-MIG) has been in force since July 2026. The central market surveillance authority is the Federal Network Agency (Bundesnetzagentur), and BaFin for the financial sector. The KI-MIG does not create a registration requirement of its own.
Classify Your AI Systems With Us
Bring your current list. In 30 minutes we use one of your systems to show how the AI compliance check leads to role, classification and requirements.