AI chatbot on your SME website: what the Swiss nFADP requires

Does the nFADP apply to an AI chatbot?

Yes, as soon as it processes personal data. The FDPIC said so on 9 November 2023 and confirmed it on 8 May 2025: the Act is directly applicable to AI-supported data processing. In January 2026 it clarified that generative AI as such is not subject to the FADP. It becomes subject to it once it processes personal data.

On a website, that is nearly always the case: according to the FDPIC, the information customers give a customer-service chatbot will usually include personal data. A name and an email address are enough, since any information relating to an identifiable person counts (Art. 5 let. a FADP).

Your business is the controller. The chatbot platform, the automation tool, the model provider and the host process the messages on your behalf, in principle as processors (Art. 9 FADP). The duties below therefore sit with you.

Do I have to tell visitors they are talking to an AI?

The Act does not mention chatbots, but the FDPIC derives from the transparency principle that users of a language model have a right to know whether they are speaking or corresponding with a machine, and whether what they type is used to improve the system or for other purposes. Anyone who provides or uses an AI system must also make the purpose and functionality of the processing transparent, data sources included.

On top of that comes the duty to inform when collecting data (Art. 19 FADP), in a precise, transparent, comprehensible and easily accessible form (Art. 13 DPO). For a chatbot, in practice:

• A visible notice in the chat window, before the first message, saying it is an automated assistant (AI) and naming the provider that processes the messages, with its country.
• A link to the privacy policy, right in that window.
• No first name or photo that would suggest a human adviser.
• A simple way to reach a person, by email or by phone.

What should the privacy policy say about the chatbot?

Art. 19 FADP sets a minimum, and the FDPIC adds to it for AI. For a chatbot or a form processed by AI, your privacy policy should state:

• Your business's name and contact details, as controller (Art. 19 para. 2 let. a FADP).
• The purpose, for example answering visitors' questions or drafting replies to incoming requests (let. b).
• The recipients: chatbot platform, automation tool, model provider, host (let. c).
• The country of each recipient abroad, with the safeguard used or the Art. 17 exception relied on (Art. 19 para. 4).
• That visitors are talking to an AI, and whether their messages are used to train or improve a model (FDPIC position).
• How long conversations are kept: you will have to say so if someone makes an access request (Art. 25 para. 2 let. d FADP), so you may as well write it down from the start.

An AI reads the messages from your contact form: what changes

Visitors never see an AI, but the rules are the same. If form submissions go through a workflow (n8n or Make, for example) and then through a language model's API that sorts them or drafts a reply, that provider is a recipient you must name in your privacy policy, with its country.

Proportionality (Art. 6 para. 2 FADP) and data protection by default (Art. 7 para. 3) mean sending the model only what it needs. To sort a request, the message body is often enough: contact details can stay in your CRM. The FDPIC gives individuals the same advice: remove personal data before entering a request into an AI-supported program, where possible.

Purpose matters too (Art. 6 para. 3): any further processing must stay compatible with what you announced, and someone asking for a quote does not expect their message to train a third party's model. Check in the API plan's terms whether your data is used for training and where it is processed. Finally, an AI-generated summary saved in your CRM must be accurate (Art. 6 para. 5): having a person read it is still the best safeguard.

Some companies publish this flow. Neuronia, a Lausanne studio that connects the software of SMEs in French-speaking Switzerland to n8n workflows and language models, with a human validation step at sensitive points (I built its website with Safwan Abdirahman), has an “IA et nLPD” page on it. The page links four working rules to articles of the Act, including proportionality (Art. 6 para. 2) and data protection by design (Art. 7), and shows, for each option, where an email's text is read, for example by the API provider, often outside Switzerland, or on a server at an agreed location.

An AI provider in the US or the EU: what does Art. 16 say?

Sending a message's text to a model hosted abroad is a disclosure of personal data abroad. It is freely permitted to countries the Federal Council has found to have an adequate level of protection, listed in Annex 1 of the Data Protection Ordinance (DPO): all EU and EEA states, the United Kingdom, Israel and New Zealand, among others.

For the United States, only companies certified under the Swiss-U.S. Data Privacy Framework are covered, since 15 September 2024. Before you connect a US model, check on dataprivacyframework.gov that the entity you contract with is listed with “Swiss-U.S. DPF” certification. Otherwise you need another safeguard under Art. 16 para. 2 FADP, in practice the EU standard contractual clauses recognised by the FDPIC, which can be used without notifying the FDPIC.

That leaves the Art. 17 FADP exceptions, such as each visitor's explicit consent or a direct connection with a contract. I would not build a chatbot on them. To keep messages in Switzerland, an open model on a Swiss server, or installed on your own premises, avoids any disclosure abroad for that step.

Processor contract and security: what to check

Handing messages to a provider requires a contract (Art. 9 FADP), often the provider's data processing addendum (DPA). The provider may only process data as you yourself would be permitted to, and may only pass the processing to a third party with your prior approval (para. 3). It is up to you to check that it can guarantee data security (para. 2).

Watch out for professional secrecy: using a processor is ruled out if a statutory or contractual duty of confidentiality prohibits it (Art. 9 para. 1 let. b). A medical practice or a law firm should settle this with its legal adviser.

On security, Art. 8 FADP requires measures appropriate to the risk, and Art. 3 DPO requires, among other things, that authorised persons only access the data they need. For a chatbot: API keys kept out of the website's public code, restricted access to conversations and workflow logs, regular updates, and a retention period after which everything is deleted (Art. 6 para. 4 FADP).

Fines of up to CHF 250,000 target, in principle, the responsible individual, and only for intentional breaches, for example failing to inform (Art. 60 FADP) or disclosing data abroad without safeguards (Art. 61).

Automated decisions: when to keep a human in the loop

A chatbot that tells people your opening hours decides nothing. Art. 21 FADP targets a decision based exclusively on automated processing that has a legal consequence for the person or a considerable adverse effect on them. An AI that rejects a job application on its own, or turns down a rental application without an employee ever seeing it, may fall into that category.

You must then inform the person. On request, they can express their point of view and ask for the decision to be reviewed by a natural person. These duties only fall away if the decision is directly connected with a contract and the person's request is granted, or if they explicitly consented to the automation (para. 3). The right of access also covers the logic behind the decision (Art. 25 para. 2 let. f).

The simplest set-up for an SME: the AI sorts and drafts, a person decides. A draft that an employee approves before sending is no longer a decision made exclusively by the machine. Keep that approval step at least for refusals and for anything that commits the business, such as a price or a deadline.

Health data and other sensitive data in the chat

Visitors can describe their health in a chat window. That is sensitive personal data (Art. 5 let. c FADP), and processing it on a large scale requires a prior data protection impact assessment (Art. 22 FADP). The FDPIC points out that high-risk AI processing is permitted in principle, but with that assessment and appropriate safeguards.

For a physiotherapy practice or a dental clinic, the safest approach is to ask visitors in the chatbot window not to describe their health there, and to offer another channel. If the tool has to handle that kind of content, have the set-up checked by a specialist before it goes live.

Is a Swiss AI law on the way?

It is being prepared, but nothing is in force. On 12 February 2025, the Federal Council decided that Switzerland should ratify the Council of Europe Convention on AI, which it signed in March 2025, and amend Swiss law accordingly, while continuing to regulate AI sector by sector, for example in healthcare and transport. The Federal Department of Justice and Police (FDJP) is drafting a bill to be submitted for consultation by the end of 2026, covering transparency, data protection, non-discrimination and supervision, among other things.

According to the Federal Chancellery, Switzerland does not yet have any overarching legislation dealing specifically with AI, so the nFADP is what applies. If you also target customers in the EU, European law may apply on top, and my guide to the nFADP and websites explains when the GDPR applies.

AI chatbots and the nFADP: the checklist

Before you put a chatbot online or connect an AI to your form:

• List every tool that sees the messages (chatbot platform, workflow, model, host) and its country.
• Outside Annex 1 DPO, put a safeguard in place (Swiss-U.S. DPF certification, standard clauses recognised by the FDPIC) or choose a model hosted in Switzerland.
• Sign each provider's processor contract and read its terms on training and data retention.
• Show in the chatbot window that it is an AI, with a link to the privacy policy.
• Update the privacy policy: recipients, countries, safeguards, use of messages, retention period.
• Send the model only the text it needs and set a retention period for conversations and logs.
• Keep a human approval step before any refusal or commitment sent to a customer.
• Know how to answer an access request covering stored conversations, as a rule within 30 days (Art. 25 FADP).

Want a chatbot or an AI connected to your website?

At Kollowo, I design and maintain websites for SMEs in French-speaking Switzerland. If you want to add a chatbot or connect your contact form to an AI tool, write to me: we'll look at where the messages will go before I write the first line of code.