AI must be recognisable from August: what Swiss companies should check across chatbots, voice and content

A larger part of the EU AI Act becomes applicable on 2 August 2026. For a Swiss company, that can sound like a Brussels issue for a legal team. In practice, it may sit directly on the website: a chatbot serving EU customers, a voice assistant taking calls, or generated content published into a European market.

The sensible response is neither panic nor a forty-page policy nobody uses. It is a careful review of the moments where a person should reasonably know that AI is involved. Done properly, this improves compliance and removes awkward surprises from the customer experience.

Why a Swiss company may still need to pay attention

Switzerland is outside the EU, but European rules can still matter when a service is directed at people or markets in the EU. The exact scope depends on the system, the company’s role and how the offer reaches that market. This article is not legal advice; it is a map of the operational questions that deserve attention.

The most visible cases are systems that communicate directly with people. If a customer believes they are speaking to an employee while an automated system is responding, trust is already damaged before anyone discusses regulation. Transparency is not legal decoration. It is part of honest service design.

A disclosure does not have to spoil the conversation

Teams often worry that identifying AI will make an interaction feel cold. Usually the opposite is true. A short line such as “I’m AlpenAgent’s digital assistant” sets the role without forcing technical language on the customer. What matters next is whether the system understands the request and hands over cleanly when needed.

The bad experience is not the disclosure. It is a bot pretending to be human and then failing on a basic follow-up. A clear opening sets the right expectation. That is also the right logic for an AI phone assistant for Swiss SMBs: take reliable work off the team without inventing a false identity.

Where to look first

These questions belong in vendor selection, not after launch. Our AI vendor checklist helps teams check ownership, data routes and support before they commit.

Start with an inventory of visible AI touchpoints. Not every internal tool needs a public label. Focus on the places where customers, applicants, partners or the public interact with a system or receive material significantly shaped by one.

A small company does not need specialised software for this review. A simple table is enough if it has a clear owner. Someone must know where the disclosure appears, when it is shown and what needs another review after a supplier, channel or model changes.

• website, portal and WhatsApp chatbots
• voice AI on inbound or outbound calls
• automatically produced images, audio or video
• AI-generated public-interest communication
• assistants that prepare customer-facing decisions

Voice AI needs an early signal and a clean handover

Voice is more sensitive than chat because people read timing, tone and pauses instinctively. A natural voice must not mislead the caller about the system’s role. The explanation belongs near the start of the call, not buried at the end of a privacy page.

Then define the boundary. May the assistant only capture the request? Can it suggest appointments? Which commitments stay with the team? These are process and privacy questions. Our guide to privacy for AI phone assistants in Switzerland covers the data side in more depth.

Disclosure, privacy information and consent are different

A common mistake is trying to solve every obligation with one banner. Transparency, privacy information and possible consent are separate issues. Saying that AI is responding explains the interaction. It does not automatically explain what data is processed, how long it is kept or the legal basis for doing so.

Review the whole route: opening, data collection, transfer, storage and deletion. The more sensitive the request, the more important a human exit becomes. A question about opening hours should not use the same route as medical, legal or highly personal information.

Generated content needs a credible evidence trail

This does not mean placing a warning on every sentence that was polished with AI internally. Context and the risk of deception matter. A realistic manipulated video is clearly different from correcting grammar in a product description.

Still, a business should know how public content was produced. Who checked the facts? Where did figures come from? Was an image materially altered? Those questions support transparency and strengthen the proof layer that matters for visibility in AI search.

Practical changes on the website and behind it

The visible fix is often small: a clear assistant disclosure, accessible privacy information and an obvious path to a person. The internal work matters more. The inventory needs an owner and a review date so the explanation does not become stale after a vendor or channel changes.

Language matters too. “AI-powered conversational interface” sounds sophisticated but tells a customer little. “Digital assistant” or “AI assistant” is usually clearer. French, Italian and German versions should be written naturally rather than copied word for word.

• name the system’s role at the start
• link to relevant privacy information
• make human handover easy to find
• document an owner and review date
• test the disclosure in every language

A realistic sixty-minute review

Open the website, chat, call flow and key social channels as if you were a new customer. Note every place where AI responds or visibly shapes content. Ask three questions: Is the role clear? Is the data route understandable? Is there a sensible human exit?

Put unclear points on a short action list. Send legally complex cases to a specialist. Fix obvious user-experience gaps immediately. This turns a large regulation into manageable work rather than another policy document that never reaches the actual customer journey.

Conclusion: honesty is the better interface

The August deadline should not push Swiss businesses to cover every page with warnings. It should make AI use more deliberate. People can know what they are interacting with without the service becoming heavy or awkward.

Clear explanations, data routes and handovers deliver more than formal safety. They make the business feel credible. That trust determines whether customers experience a chatbot or voice system as useful help or as a trick.