The EU’s New AI Transparency Rules Now Reach Swiss Companies
The EU began applying the AI Act’s Article 50 transparency rules on 2 August 2026. A Swiss company running an AI chatbot for customers in Germany now has to look at what those customers actually see.
Article 50 requires providers of certain interactive AI systems to design them so that people know when they are interacting with AI, unless the circumstances already make that obvious. It also covers machine-readable marking of AI-generated or manipulated content and requires deployers to disclose deepfakes and several other specified uses.
A legal review establishes which rules apply. It does not reveal whether the German chatbot displays the required information on a mobile phone.
Someone has to open the product and test it.
Start with the AI systems that customers, users or other people outside the company encounter. A Swiss retailer might run a service chatbot on its French website, generate product imagery through an external AI platform and publish AI-edited video through its communications team. A software company might have built an AI assistant into a product sold in Germany.
First, find the systems.
Marketing bought one tool. Customer service contracted another. A product team connected a model through an API. A country office started using a separate platform for translations. In a large company, nobody necessarily holds a complete list.
A spreadsheet is enough for the first inventory. Record the system, provider, business owner, countries where people encounter it, the output it produces and whether an external person interacts with it directly.
Then compare that record with the live product. A customer-service chatbot marked “AI disclosed” in the inventory might display the notice on desktop but lose it on mobile. The German version might contain the wording while the French translation omits it. A vendor might have replaced its widget since the company last reviewed the service.
Generated media creates another set of obligations. Take a Zurich communications team filming an executive. The team uses AI software to remove background noise from the recording. It later creates another version in which a synthetic rendering of the executive delivers translated speech.
Both files may come from the same production software, but the audience encounters something different in the second video. Article 50 sets disclosure requirements for deepfakes, including AI-generated or manipulated image, audio or video content that resembles existing people, objects, places or events and could appear authentic.
The communications team controls the easiest point at which to label that material while it still holds the master file. After regional teams cut shorter versions, compress the video and upload copies to several platforms, correcting the disclosure becomes a distribution problem.
Machine-readable marking requires the company to follow the file itself.
Providers of systems that generate synthetic audio, image, video or text must mark their outputs in a machine-readable format and make them detectable as artificially generated or manipulated. A Swiss business using a third-party generator therefore needs to know what the provider inserts into the file.
Ask the provider what marking it adds, which formats retain it and what happens during export or editing. Then send a generated file through the company’s normal publishing chain.
An image compressor that removes the mark before publication defeats the technical control even if the generator inserted it correctly.
The European Commission published its Article 50 guidelines on 20 July, less than two weeks before the obligations started to apply. AI systems placed on the market before 2 August received a limited transition only for the Article 50(2) marking and detection requirement. Providers of those systems have until 2 December 2026.
That leaves little room for companies whose vendor changes depend on procurement, product releases or several internal teams.
The European timetable also differs from Switzerland’s regulatory route. Switzerland has not copied the EU AI Act wholesale. Swiss rules therefore do not settle the position of a company offering an AI-enabled service to people in the EU.
Put geography into the inventory. A forecasting model used by employees inside a Swiss factory does not raise the same Article 50 issues as a chatbot available to consumers in Milan. Record where people encounter each system instead of giving every AI tool in the company one regulatory label.
Suppliers cannot complete the exercise for their customers either. A vendor might state that its chatbot complies with the AI Act and supply the technical disclosure function. The Swiss company still decides where to deploy it, which languages customers see, how the interface presents the disclosure and what happens to content after employees export it.
Ask the vendor which Article 50 controls sit inside the product and which remain with the customer. Open the live version afterwards and check.
National market-surveillance authorities now enforce most Article 50 obligations. The AI Office has a narrower enforcement role in specified cases. Breaches covered by the relevant AI Act penalty regime can attract fines of up to €15 million or 3 percent of worldwide annual turnover for the preceding financial year.
The first compliance check costs considerably less. Open the German website. Start a conversation with the chatbot. See what the customer sees.
