The short answer
A freely available AI tool is not automatically suitable for confidential business data. Before employees enter text, spreadsheets or documents into ChatGPT or other external AI services, they must consider company policies, confidentiality, data protection and the service’s contractual terms.
This applies even where the task appears harmless: an email draft or a summary of meeting minutes may contain names, health information, trade secrets or strategically important information.
Inputs that pose particular risks
Before transmitting information, employees should check whether the input contains information that must not be disclosed to an external provider. Examples of potentially sensitive inputs include:
- complete client files or contractual documents;
- personnel files, medical information and job applications;
- unpublished prices, quotations or business strategies;
- login credentials, technical security information and internal source code;
- information subject to statutory or contractual confidentiality obligations.
The risks depend on the particular service, its settings, the agreed terms and the intended use. The labels ‘Business’ or ‘Enterprise’ alone do not establish that use is legally approved.
What role does the employee’s duty of loyalty play?
Employees must safeguard their employer’s legitimate interests. This includes, in particular, handling confidential information and trade secrets with care. Within the limits of the law, employers may issue instructions on the use of work tools and AI services.
An explicit AI policy provides clarity. However, the absence of a policy does not automatically permit disclosure of company data. Conversely, permissible use should not be prevented merely by a blanket prohibition on every AI application. What matters is clear, risk-based guidance.
Data protection begins before uploading
Where a document contains personal data, using an AI service constitutes processing relevant to data protection law. The company must clarify, in particular:
- Which data is genuinely necessary for the task?
- For what purposes does the provider process inputs and outputs?
- What role does the provider have, and which contracts are required?
- Where is data stored or made accessible?
- What information must be provided to the individuals concerned?
- What options exist for deletion, security and access control?
Disabling model training does not answer all these questions. For example, it does not provide a conclusive answer regarding logging, retention, support access or other recipients.
Is removing names enough?
Not necessarily. A text may remain attributable to an individual even without names. A role, employer, unusual circumstances or a combination of details may make identification possible.
Replacing a name with ‘Person A’ is therefore not necessarily anonymisation. For a data-minimising workflow, it may be more appropriate to formulate an abstract question rather than upload an entire file.
Who is liable for an error in an AI output?
AI-generated text must not be treated as an accurate professional basis without review. Expert verification is particularly necessary for legal, financial or safety-related tasks. Plausible wording may contain incorrect information, non-existent references or inappropriate conclusions.
Whether an error has employment-law or liability consequences depends, among other things, on the employee’s duties, instructions, fault and the specific damage. There is neither automatic liability for every use of AI nor a general exemption on the grounds that the machine made the mistake.
May AI decide on job applications or dismissals?
Processing employee data is subject to particular restrictions. Automated assessments and decisions may also trigger information obligations and rights for the individuals concerned.
Particular caution is required where a decision is made exclusively by automated means and has legal effects or a comparably significant adverse impact. Human review must actually take place and must not merely be a formal sign-off. Discrimination risks and the quality of the data used must also be examined.
The requirements depend on the specific process. Using a writing assistant to draft a job advertisement is not the same as automatically rejecting applications.
Five elements of a useful AI policy
A practical policy should specify:
- Approved tools: Which services and accounts may be used for which tasks?
- Data categories: Which information may be entered — and which must not?
- Review: Who checks the results and bears professional responsibility?
- Transparency: When must AI use be disclosed internally or to third parties?
- Incidents: What should be done if sensitive data is transmitted accidentally?
A clear reporting process is more important than expecting mistakes never to occur. Following an erroneous transmission, relevant information should be preserved, responsible individuals informed and further risks assessed. Whether notification to affected individuals or authorities is required must be evaluated separately.
Three questions about AI use in businesses
Is ChatGPT generally prohibited for Swiss businesses?
No. Permissibility and requirements depend on the data, purpose, provider’s terms and specific use.
May employees use their personal AI accounts?
Not without checking company approval. Personal accounts may have different terms and offer less control than agreed business solutions.
Must every instance of personal-data processing be based on consent?
No. For private controllers, the Swiss Federal Act on Data Protection does not require the same general system of legal bases as the GDPR. Nevertheless, processing must be lawful, proportionate and consistent with other applicable rules.
Establish legal and organisational rules for AI use
Müller & Pfenninger helps businesses coordinate employment-law requirements, confidentiality and data protection when using external services in the workplace.
Note: This article provides general information about Swiss law and is not a substitute for individual legal advice. Provider terms and the current legal position have not been independently verified.