Why the inventory comes before classification
Article 6 and Annex III of the AI Act require classifying every AI system by risk level. But classification is a "per system" exercise — you can't apply it to something that's not on any list. Most companies underestimate how many AI systems they actually use, because they only count what they formally built or purchased, forgetting the AI each team adopted on its own: the CRM's AI plugin, the writing assistant built into email, or plain ChatGPT used without anyone in management knowing. This is called shadow AI, and it's the number one reason a "from memory" inventory always falls short.
What counts as an "AI system" under the Regulation
The AI Act's definition is deliberately broad — it's not limited to custom-built machine learning models. In practice, anything that makes decisions, generates predictions, produces content, or issues recommendations from input data, with varying degrees of autonomy, counts as an AI system. That includes, among others:
- Generative AI features built into software you already use (writing assistants, auto-summaries, code copilots).
- Customer service chatbots, internal or public-facing.
- Scoring or recommendation engines (credit, customer, pricing).
- Candidate screening or evaluation tools in hiring.
- Fraud or anomaly detection systems.
The fields every entry needs
| Field | Why it matters |
|---|---|
| Name and vendor | Identifies the tool and who's responsible for its maintenance and updates. |
| Function / use case | What specific decision or task it performs — the input the next step's risk classification needs. |
| Data processed | If it processes personal data, this system is also on GDPR's radar, not just the AI Act's. |
| Your company's role | Are you a provider (you build it) or a deployer (you use it)? AI Act obligations differ by role. |
| Preliminary risk classification | Filled in during the next step — the inventory leaves the field ready so you don't have to rebuild the list later. |
| Internal owner | Who in your company answers if a regulator, a client, or an audit asks about this system. |
| Date added and last reviewed | Without a review date, there's no way to know if the inventory is still reliable. |
The special case: EU database registration (Article 71)
Don't confuse your internal inventory with the public Article 71 registration. They're different things with different obligated parties. The internal inventory is yours, you need it whether or not a high-risk system exists, and no one outside your company sees it. Article 71 registration is a public EU database where the provider (or its authorized representative) has to register any system that, after classification, comes out high-risk under Annex III — and for certain systems used by public authorities, the deployer also registers additional information. If your internal inventory doesn't currently have any system classified as high-risk, Article 71 simply doesn't apply to you yet — but you need the inventory to know that for certain, not to guess it.
How to start without specialized software
You don't need governance software to get started. A spreadsheet with the columns from the table above is enough for a first version. What matters is the collection process: ask each department separately — marketing, HR, customer service, finance — what AI tools they use day to day, not just the ones "IT signed off on." It's common for the first round to double the number of systems leadership thought it had. Once the inventory is closed, the natural next step is classifying each entry by risk level.
Frequently asked questions
What if my company doesn't know all the AI systems it uses?
That's the usual starting point, not an exception. Most companies discover, while building the inventory, AI tools no department ever officially signed off on — mostly generative AI each team adopted on its own. The inventory exists precisely to bring those tools into the light before classifying anything.
Is the internal inventory the same as the Article 71 EU database registration?
No. The internal inventory is your own working document, and you need it whether or not a high-risk system exists. Article 71 registration is a public obligation that only applies once, after classification, a system comes out high-risk under Annex III — that's when the provider (or its authorized representative) must register it in the EU's public database.
Do I need to inventory tools like ChatGPT that my employees use?
Yes, if it's used for work, even if it's not a system your company formally built or purchased. Article 4's AI literacy obligation explicitly covers any AI use by staff, and you can't train people on it or govern that use if you don't know it exists.
How often should I update the inventory?
At minimum, a quarterly review, plus an immediate update whenever a team adopts a new tool. An outdated inventory gives a false sense of compliance — worse than knowing you have a gap.