How an SME takes inventory of its AI tools: the document to have before anything else

The step all the others assume

Our guides on the AI Act, the European regulation on artificial intelligence (Regulation (EU) 2024/1689), end with the same advice: take inventory of your uses of AI. The one on 2 December 2026 (/conseils/ai-act-2-decembre-2026) asks you to write to the providers of your generative tools; you still have to know which ones you use. This guide describes the document itself: what it contains, how to fill it in, how to keep it up to date.

It is written for an SME that uses AI tools without having developed them: a writing assistant, a sorting module in the HR software, a chatbot on the website. The European regulation calls this role “deployer”. Drawing up the list requires no technical skill; above all, it means asking the right questions of the right people.

What the law requires, and what it does not

In Switzerland, no law today imposes an “AI register” as such. Swiss law specific to AI does not exist yet: the preliminary draft implementing the Council of Europe Framework Convention on AI is due to go to consultation in spring 2027 (Federal Council, 28 October 2026). The AI Act, for its part, does not require an inventory under that name.

What does exist is the register of processing activities of the Federal Act on Data Protection (FADP, art. 12). It contains at least the identity of the controller, the purpose of the processing, the categories of data subjects and of personal data processed, the categories of recipients and, where data is disclosed abroad, the country concerned and the safeguards provided (art. 12 para. 2 FADP). An exception covers companies with fewer than 250 employees, provided their processing entails a limited risk of harm to the personality of the data subjects (art. 12 para. 5 FADP and art. 24 DPO). If the European GDPR applies to your activity, its exception is much narrower (art. 30 GDPR).

The AI Act applies to a Swiss SME only in the cases set out in its art. 2: when it sells, in the European Union (EU), products or services that integrate an AI system (art. 2(1)(a)), or when the output produced by its systems is used there (art. 2(1)(c)). In those cases, several obligations presuppose knowing which tools run in the company: AI literacy of staff (art. 4, applicable since 2 February 2025, softened into an obligation of means by the Digital Omnibus on AI), transparency (art. 50, applicable since 2 August 2026) and, for high-risk AI systems under Annex III, the obligations of deployers under art. 26, applicable from 2 December 2027. The voluntary standard ISO/IEC 42001 calls for the same inventory.

An inventory of AI tools is therefore good practice, not a stand-alone legal obligation. But every line that touches personal data feeds the FADP register, and every line with an EU link serves to classify your obligations under the AI Act. That is why it comes before anything else.

Step 1: find every tool, including the ones nobody chose

The list is built from four sources. Invoices and card statements: subscriptions to assistants, to transcription or image generation tools. The administration consoles of your software: e-mail, office suite, customer relationship management, accounting, HR tool, many of which switch on AI features through updates. The website: chatbot, a form that sorts requests. And people: one short question to each team lead, “which AI tools do you use for work, even free ones, even on a private account?”.

That last question matters most. Unofficial tools, opened with a personal address and used for work, are often the ones where customer data goes in without a contract or settings. The AI-Karma assessment question on inventory targets them explicitly. Ask it without looking for someone to blame: the aim is to see, and a person who fears a sanction will say nothing.

One writing rule avoids most mistakes: one line per use, not per tool. The same chat assistant used to translate quotes and to pre-select job applications makes two lines, because the classification depends on the use, not on the software. Translation is usually minimal risk; pre-selection falls under Annex III, point 4 (employment), if the EU is involved.

Step 2: six questions for each line

Who uses it. The department or role (accounting, HR, sales) and the approximate number of people. They are the ones AI literacy targets (art. 4 AI Act) and, for a high-risk HR tool, the ones to inform as workers before it is put into service (art. 26(7), from 2 December 2027).

What for. The purpose, in one concrete sentence: “drafting replies to quote requests”, not “productivity”. It decides the classification, and it is also an entry in the FADP register (art. 12 para. 2 let. b).

Which data goes in. Four boxes are enough: no personal data; data on customers or prospects; data on employees or applicants; sensitive personal data within the meaning of art. 5 let. c FADP (health, religious or political views, ethnic origin, among others). As soon as personal data goes into the tool, the line falls under the FADP: informing people (art. 19), data security (art. 8).

Which provider, and where the data goes. The vendor's name, the hosting country, whether a processing agreement exists (art. 9 FADP) and what the provider does with the data sent, in particular whether it reuses it to train its models. Disclosure abroad is allowed only to a country whose legislation ensures an adequate level of protection, or with appropriate safeguards (art. 16 FADP); the FADP register records the country and the safeguards (art. 12 para. 2 let. g).

Which AI Act classification. Out of scope (no link with the EU), prohibited practice, high risk, transparency or minimal risk: the next step explains how to find it. Until it is done, write “to classify” rather than leaving the box empty.

Who is responsible. A named person, not a department, who answers for the line: they know who uses the tool, approve a change of use and review the line on the planned date. In an SME, this is often the person who leads the team using the tool, not IT.

Step 3: classify each use

First question, before any other: is the EU involved? Do you sell to customers located in the EU products or services that integrate this tool, is its output used there, does it assess people located there? If the answer is no for that line, the AI Act does not apply to it. The FADP still does.

If the EU is involved, the order of reading follows the regulation. First the prohibited practices of art. 5 (emotion recognition at work, social scoring, manipulation), banned since 2 February 2025, to which two prohibitions are added on 2 December 2026. Then the high-risk areas of Annex III: employment and HR (point 4, the most frequent category in SMEs), creditworthiness assessment, pricing in life or health insurance, education, access to essential services, biometrics; the obligations of deployers (art. 26) apply there from 2 December 2027. Then the transparency rules of art. 50: a chatbot facing the public, published generated content, deepfakes, applicable since 2 August 2026. The rest is minimal risk, and AI literacy of staff (art. 4) applies to every line.

One case is noted separately: an SME that markets an AI system under its own name or trademark, or changes its intended purpose, can become a provider within the meaning of art. 25. That requalification only operates towards high risk, but it then changes every obligation on the line.

Step 4: keep the document alive

An inventory drawn up once and filed away no longer describes the company a few months later: software adds AI features, teams adopt new tools. Three rules keep it current. Every new tool goes through the designated owner before use, which the AI usage policy can provide for. Every line carries a review date; the frequency is the company's choice, no text sets it. An abandoned tool stays on the list with the status “retired”, to keep a record of what was done with the data.

Then copy into the FADP register of processing activities, if you keep one, the lines that touch personal data: the purpose, the categories of data and recipients, the hosting country are already written there. The two documents stay separate, one listing tools, the other processing activities, but the first saves you from writing the second from memory.

The site's free tools to keep it

The AI system register (/registre) takes up these questions: use case, tool, provider, internal owner, status (in production, pilot or test, considered, retired), classification and deadline, personal data, hosting, processing agreement, reuse by the vendor, review date and a plan for when the tool stops. It has no dedicated column for users: write the department in the use case label, for example “HR: pre-selection of applications”. Everything stays in your browser; the register exports to CSV or JSON, prints to PDF, and its reviews can be added to a calendar (.ics file).

The classification module (/aiact) asks the questions of step 3, use by use, and adds the result to the register. The usage policy (/charte) sets the list of authorised tools and the rule for admitting a new one. If a line touches sensitive data or profiling, the /aipd module tells you whether an impact assessment is required (art. 22 FADP). The full assessment (/evaluation) places your inventory among your other practices: its highest level is a complete inventory, kept up to date, with a designated owner. And if the company does not use any AI tool yet, the “Starting well” page (/demarrer) gives the first actions in order.

This is not legal advice. For a use at the boundary of high risk or a possible requalification as provider, have your analysis validated by specialised counsel.

Frequently asked questions