Using AI without putting your data at risk: the method for a Swiss SME

A Swiss SME using AI does not need a new law to have obligations: the revised Federal Act on Data Protection (nFADP), in force since 1 September 2023, applies as soon as an employee puts a client email or a personnel file into an AI tool. The good news is that the work comes down to a handful of written decisions. Here is the method, without legal jargon, and what to check.

What the nFADP already requires

Five obligations matter for everyday use of AI:

• Purpose and proportionality. You know what each use is for, and you only put in the data that is needed. "Saving time" does not justify uploading your entire client history.
• Processors (Art. 9). An AI provider that processes personal data on your behalf is a processor: you need a contract governing that processing and its own subprocessors. The consumer terms of free tools do not provide this.
• Disclosure abroad (Art. 16 and 17). If data is processed outside Switzerland, the country must offer adequate protection (as the European Union does) or the contract must contain safeguards, such as standard contractual clauses. For the United States, companies certified under the Swiss-U.S. Data Privacy Framework have been recognised since 15 September 2024.
• Information (Art. 19). Your privacy notice lists the categories of recipients, including AI providers, and the countries where data is processed.
• Impact assessment (Art. 22). It is required before any processing likely to result in a high risk, for example involving health data or assessing individuals.

The law also provides for criminal fines aimed at the individuals responsible, not just the company. Our analysis of the nFADP, the AI Act and generative AI covers these points in detail, including the case of companies that work with the European Union.

Step 1: list the uses, not all your data

There is no need to inventory every piece of data in the company. List what will go into AI: for each planned use, the documents involved, where they come from (internal, client, supplier, candidate) and who is responsible. Half a day is enough; it is part of the AI Workshop.

Step 2: sort into three levels

The trap is classifying everything at the strictest level: it feels reassuring, but the tool becomes unusable and staff go back to their free accounts. A sound classification protects what needs protecting and lets people work on the rest.

Step 3: choose and configure the tool

Check these once when accounts are opened, then whenever the plan or the provider changes:

• No training on your content. On the business plans of the main assistants (ChatGPT Business and Enterprise, Claude Team and Enterprise, Microsoft 365 Copilot, Gemini for Google Workspace), this is the default. There is no opt-out to hunt for, but there is a clause to read.
• The data processing agreement and the provider's list of subprocessors.
• The place of processing and the legal basis for the transfer: an adequate country, Data Privacy Framework certification or standard contractual clauses.
• Retention: how long conversations and files are kept, and who can delete them.
• Access: who administers the workspace, centralised sign-in (SSO) where possible, accounts removed on the day someone leaves.
• Connectors: which external tools (email, file storage) the assistant can read. At the start, as few as possible.

Our guide ChatGPT Free, Plus, Business or Enterprise applies this checklist to ChatGPT.

Step 4: professional secrecy and your commitments

Professional secrecy does not depend on the nFADP. The professions covered by Article 321 of the Swiss Criminal Code (notably lawyers, notaries, doctors, auditors and their assistants) may only entrust information covered by secrecy to a third party within a framework compatible with that secrecy: a suitable contract, a solution hosted under your control, or the client's consent, depending on the case. A fiduciary firm is also bound by its duty of discretion towards its clients, and a bank by banking secrecy.

Reread your client contracts too: a non-disclosure agreement or a clause prohibiting undisclosed subcontracting may rule out a use that the law would allow.

Step 5: review before anything goes out

Any AI output that leaves the company (for a client, a supplier or an authority) is reviewed by a person. For agents that send messages, the rule is strict: approved drafts at the start, automatic sending only for one low-risk category, after at least four weeks of approved drafts and on a written decision. Our guide to agents explains it in detail.

Step 6: plan for incidents and keep track

Write down what happens if protected data goes into the tool or a mistake reaches a client: who deletes, who informs, who decides. If a data security breach results in a high risk for the people concerned, it must be reported to the FDPIC as soon as possible.

Then review the rules whenever there is a new tool, a new provider or a new use: a short check-in every month at the start, then every quarter.

The three mistakes that make security purely decorative

1. A written rule that the tools contradict. The policy bans personal accounts, but no business tool is provided. The result: undeclared use.
2. The "everything is protected" classification. The tool becomes useless and the rules get bypassed.
3. Removing the review too early. You remove it once use is stable and measured, not because you are in a hurry.

What it costs

The sheet on permitted and prohibited data is part of the AI Workshop (CHF 1,900 excluding VAT, 4 hours). Configuring the tool with your IT team, checking contracts, the rules and ongoing follow-up fall under AI Adoption Support, at a fixed price on quotation or as a subscription from CHF 1,800 a month for 10 hours, with a three-month minimum. This guide is a working method, not legal advice on your situation.