In short. An AI inventory is a list of every AI system in your company, with purpose, data, owners and risk class. The EU AI Act does not require it in so many words. Without it, though, you cannot show that you meet the obligations that already apply: the prohibitions, AI literacy and labelling. Most AI is not found in the IT department but in purchased software and in private accounts. According to Bitkom, 42% of companies know or suspect that employees use private AI at work. To start, a table with twelve fields, six steps and one rule are enough: no new AI without its own row.
What an AI inventory is and why you need one
An AI inventory is a stocktake, not a legal opinion. Every AI system gets one row: what it does, who uses it, which data goes in and who is accountable for it. Some call the list an AI register or AI directory. It means the same thing.
The list supports three things you have to deliver anyway.
- EU AI Act. The prohibitions in Article 5 have applied since February 2, 2025, and two new ones follow on December 2, 2026. Under the revised Article 4, you must take measures to support the AI literacy of your people, taking into account the context in which the systems are used. From December 2, 2027, the deployer obligations for high-risk AI apply. For all of this you need to know which systems you have.
- Data protection. If an AI processes personal data, the processing belongs in the record of processing activities under Article 30 GDPR, and where the risk is high, with an impact assessment under Article 35. The AI inventory and the record of processing should refer to each other.
- Works council. In Germany, section 90(1) no. 3 of the Works Constitution Act (BetrVG) requires the employer to inform the works council in good time about planned work processes, “including the use of artificial intelligence”. If you don't know where AI runs, you can't do that.
Many companies are not there yet. In the Bitkom survey from September 2026, 37% saw themselves affected by the AI Act as users, and another 34% were still checking.
What counts as an AI system and what does not
The AI Act defines an AI system in Article 3(1). In short: a machine-based system that operates with some degree of autonomy and infers from its input how to generate predictions, content, recommendations or decisions. On July 29, 2025, the Commission published guidelines on this. They break the definition down into seven elements. The deciding one is usually inference: does the system learn or reason by itself, or does it only follow fixed rules a human has programmed?
According to the guidelines, these are not AI systems, for example:
- database queries and spreadsheets without AI features, such as “all customers who bought product X last month”,
- software that only describes and displays data, such as a sales dashboard that gives no recommendations,
- classical heuristics with fixed rules,
- simple predictions that only carry an average forward.
The guidelines are not binding and do not give an exhaustive list. For the inventory, one simple rule helps: when in doubt, include it and note in one sentence why a tool is not an AI system. And don't worry about a long list. According to the Commission, the vast majority of AI systems are not subject to any particular obligation under the regulation. The inventory is there to find the few for which that is different.
Where AI sits in your company
AI enters the company by four routes today. Only the first one reliably goes through IT.
- Your own applications. The chatbot on your website, an analysis a service provider built, an agent that pre-sorts orders.
- AI in purchased software. CRM, ERP, ticketing, accounting, applicant management, phone system. Such features often arrive with an update and are sometimes already switched on.
- Approved assistants. ChatGPT, Microsoft Copilot or Gemini with a company account, plus the extensions and agents that run inside them.
- Private accounts. Tools that employees use on their own initiative for work. The term shadow AI has taken hold for this.
How widely AI now runs is shown by the Bitkom survey from September 2026. 57% of companies with 20 or more employees use AI, up from 20% in 2024. Most often in customer contact, but areas that rarely come to mind first in an inventory have grown considerably.
Where companies use AI, Bitkom 2026
Controlling and accounting rose from 17 to 25% in one year, the IT department from 2 to 9%. So don't just ask sales and marketing, ask every department.
The fourth route is the hardest. In the Bitkom survey from October 2025, 25% of companies knew for certain that employees use private AI accounts at work. Another 17% assumed so. Only 29% were sure it does not happen with them, compared with 37% a year earlier.
Do employees use private AI at work? Companies' answers
The same survey shows why. Only 26% of companies gave their employees access to generative AI, and only 23% had rules for AI tools. Anyone without an approved tool goes and finds one.
The AI inventory in six steps
1. Assign responsibility, involve the works council early
One person keeps the list together, usually from IT, data protection or compliance. Data protection, purchasing and one contact per department contribute. Inform the works council before you start. In Germany it has a right to be informed about AI, and if it has to assess a use of AI, an external expert is deemed necessary under section 80(3) BetrVG.
2. Review the lists you already have
Much of it is already written down somewhere. Go through the software and licence list, the contracts and data processing agreements, the record of processing activities, credit card and expense reports with subscriptions, and the applications that employees have allowed to access their company account. Your single sign-on admin console shows these grants.
3. Evaluate network data in aggregate
Proxy, firewall or DNS logs show which AI services are called from the company network. Only evaluate them in aggregate, per service and not per person, and agree on it beforehand with the works council and data protection. Technical facilities that can monitor behaviour or performance are subject to co-determination under section 87(1) no. 6 BetrVG.
4. Ask departments and suppliers
A short questionnaire to every department is enough: Which tools with AI do you use, for what, with which data, with which account, and who decides based on the result? Make clear that this is about an overview, not about blame. Otherwise you won't get honest answers. Also ask the vendors of your most important software which AI features are included, which of them are active and where the data goes.
5. Record and classify
Every finding gets a row with the fields from the checklist below. Then you classify: Is the use prohibited? Does it fall under Annex III, such as employment, creditworthiness or education? Does a labelling obligation under Article 50 apply? Or does it stay at AI literacy? One example that often goes unnoticed: AI that screens or filters job applications is listed explicitly in Annex III no. 4. According to Bitkom, 6% of companies were already using AI for the first screening of incoming applications in summer 2026. How to determine role, risk class and obligations in detail is covered in our post on the AI Act gap analysis.
6. Maintain it instead of counting again
An inventory goes out of date quickly. Three rules keep it current: no new AI without its own row before it starts, including trial accounts. Major updates of purchased software are checked for new AI features. Once a year the whole list is reconciled with the record of processing activities and the licence list.
Checklist: these fields belong in the AI inventory
A spreadsheet is enough. What matters more than the tool is that every column is filled in.
| Field | Example | What you need it for |
|---|---|---|
| Name, vendor, version | AI feature “suggested reply” in the ticketing system | requesting documents from the vendor |
| Type | own, in purchased software, assistant, private | shows who has influence over the system |
| Purpose | suggest replies to customer enquiries | basis for the risk class |
| Department, owner | customer service, head of customer service | clear responsibility |
| Users | 14 employees | training needs under Article 4 |
| Input data | customer emails with names and order numbers | data protection, trade secrets |
| Output and decision | draft, a person checks and sends it | human oversight, liability |
| Role under the AI Act | deployer | which obligations apply |
| Risk class with reasoning | no particular obligation, not an Annex III case | proof that it was checked |
| Contract, data protection, location | data processing agreement in place, data centre in the EU | link to the record of processing |
| Works council, training | informed on March 3, training completed | proof of co-determination and AI literacy |
| In use since, next review | since May 2026, review in May 2027 | keeping the list current |
What you should do now
- Name a responsible person and inform the works council about the project.
- Set up the table with the twelve fields and fill it with the systems you already know.
- Go through the licence list, contracts, subscriptions and app grants. That usually brings the most findings.
- Ask every department and your main vendors, without assigning blame.
- Classify every finding, first against the prohibitions, including the two new ones from December 2, 2026.
- Offer an approved tool and a short rule, so that private accounts become unnecessary. Which ChatGPT plan is suitable is covered in our guide (in German).
Frequently asked questions about the AI inventory
Is an AI inventory required by law?
Not in so many words. For companies that use AI, the EU AI Act does not prescribe a separate AI register. It does require you to rule out prohibited practices, promote the AI literacy of your people and, from December 2027, meet the obligations for high-risk AI. On top of that come the record of processing under the GDPR and, in Germany, informing the works council. All of this assumes you know your AI systems. The inventory is the simplest way to get there.
Do ChatGPT and Copilot belong in the AI inventory?
Yes. Anyone who uses such assistants at work is a deployer. Record the company account as a system and the main uses as separate rows if they differ clearly. An assistant that drafts texts for the newsletter has to be assessed differently from the same assistant used to pre-sort job applications.
What do we do with shadow AI we find?
Record it, don't punish it. Find out which purpose the private tool serves and offer an approved alternative with a company account. Set out in a short policy which tools are allowed and which data must not go into them. Bitkom recommends exactly these two steps to companies: clear rules and company AI offerings for employees.
How often do we need to update the AI inventory?
Continuously with every new AI and every major update, plus a full review once a year. There is no fixed deadline. What matters is that new systems get a row before they start.
What is the difference between an AI inventory and a gap analysis?
The inventory answers the question of which AI runs where. The gap analysis builds on it: it assigns role, risk class and obligations to every system and shows what is still missing. Without an inventory there is no gap analysis.
An AI inventory can be done in a few weeks if you start with the lists that already exist and ask departments without reproach. If you would like a second pair of eyes on your list, we are happy to look at it together.