The Swiss FADP and EU GDPR both regulate how AI services handle personal data, but their requirements differ. GDPR requires an Article 6 lawful basis for each processing purpose. Swiss private-sector rules centre on processing principles and protection against unjustified infringements of personality rights. Breach reporting also differs: Switzerland generally requires authority notification when a breach is likely to create high risk, while GDPR has a lower reporting threshold.
For businesses using AI assistants, recruitment tools or predictive systems, these differences affect training data, automated decisions, contracts and incident response.
This guide focuses on private businesses. Swiss federal bodies face additional requirements, while cantonal authorities may fall under separate laws.
Provision | Swiss FADP | EU GDPR | AI implication |
|---|---|---|---|
Territorial scope | Covers circumstances producing effects in Switzerland | Covers EU establishment and specified targeting or monitoring | Both laws may apply |
Lawful processing | Processing principles and personality-rights protection | An Article 6 lawful basis for each purpose | Assess each use separately |
Sensitive data and profiling | Specific sensitive-data and high-risk profiling rules | Article 9 conditions and profiling safeguards | Assess inputs and inferences |
Automated decisions | Information and human-review rights, with exceptions | Article 22 restricts qualifying decisions, with exceptions | Examine consequences for people |
Transparency | Explain purposes, recipients and relevant foreign disclosures | Detailed Article 13 or 14 information | Disclose training and reuse |
DPIA | Required where processing may create high risk | Required where processing is likely to create high risk | Assess the actual use case |
Providers and transfers | Processor safeguards and foreign-disclosure rules | Article 28 contracts and Chapter V transfer rules | Review vendors and destinations |
Breach notification | High-risk breaches reported as quickly as possible | Report unless unlikely to create risk; normally within 72 hours | Apply separate reporting tests |
Privacy roles | Adviser generally optional; representative sometimes mandatory | DPO and representative required in specified circumstances | Assess appointments separately |
Enforcement | Certain intentional offences carry personal criminal fines | Administrative fines can reach €20 million or 4% of worldwide annual turnover | Liability and enforcement differ |
The FADP applies to circumstances producing effects in Switzerland, even when they originate abroad. An overseas AI provider handling information about people in Switzerland may therefore need to assess Swiss obligations.
GDPR applies to processing in the context of an EU establishment’s activities. It can also cover businesses without an EU establishment when processing relates to offering goods or services to people in the EU or monitoring their behaviour there.
A Swiss recruitment business offering services to applicants in Germany, for example, may need to consider both laws when using AI to assess applications.
Citizenship and server location do not determine the answer. Hosting a model in Zurich does not remove GDPR obligations, and an EU citizen’s involvement does not automatically bring every overseas activity within GDPR scope.
Assess where the business operates, whom it serves and whose behaviour it monitors. The GDPR compliance guide explains the broader obligations once GDPR applies.
The most important distinction is how private businesses establish whether processing is lawful.
Swiss private businesses must process personal data lawfully, in good faith and proportionately. Collection must serve a specified purpose recognisable to the individual, and further processing must remain compatible with it.
They do not have to select a GDPR-style lawful basis for every activity.
Articles 30 and 31 address infringements of personality rights and their justification. An infringement may arise from processing contrary to applicable principles or an individual’s express wishes. Consent, an overriding private or public interest, or legislation may justify it.
Where consent is required, it must be voluntary and informed. When relying on consent for sensitive personal data or private-sector high-risk profiling, it must be explicit. This does not make consent compulsory for every activity involving sensitive information.
GDPR Article 6 requires an applicable lawful basis for each processing purpose:
Consent.
Necessity for a contract or requested pre-contractual steps.
Compliance with a legal obligation.
Protection of vital interests.
A public-interest task or exercise of official authority.
Legitimate interests, subject to the required assessment and limitations.
Including an activity in contractual terms does not automatically make it necessary for the contract. A commercial benefit alone does not establish legitimate interests.
Special-category data, such as health information, also requires an Article 9 condition. Criminal-conviction and offence data is governed separately by Article 10.
Assess these requirements before allowing unrestricted AI uploads. A document summarisation tool may receive medical records or employee complaints unless its use is controlled. The Swiss FADP and EU GDPR establish the underlying requirements.
Permission to process a prompt to answer a question does not automatically cover reusing it to improve a model.
Under Swiss FADP, assess whether training is compatible with the purpose made recognisable at collection. Consider proportionality, transparency, individuals’ express wishes and whether an infringement of personality rights requires justification.
Under EU GDPR, establish a lawful basis for training and consider further-processing rules. Legitimate interests requires an assessment of the specific interest, necessity and impact on individuals. Sensitive data may require an additional Article 9 condition.
For the service itself, distinguish between:
Generating the requested response.
Keeping conversation history.
Detecting abuse or investigating incidents.
Personalising future responses.
Training or fine-tuning models.
Retrieval-augmented generation retrieves documents to help answer a query; it does not inherently train the model. It still processes data and can expose information through poorly configured permissions.
The provider’s role may change with the purpose. Following customer instructions may make it a processor, while using customer data for its own training may make it a controller for that activity.
Effective AI data privacy controls address training, logs, retention and access separately. A training opt-out does not confirm that all other processing has stopped.
AI-generated summaries, profiles and decisions do not all trigger the same obligations. The processing and its consequences matter.
Swiss FADP Article 21 addresses decisions based exclusively on automated processing that have legal consequences or considerably affect an individual. Subject to exceptions, the controller must inform the person and, on request, allow them to express their position and obtain review by a natural person.
Exceptions include explicit consent to the automated decision and certain contract-related decisions where the individual’s request is granted. A connection to a contract alone is insufficient.
GDPR Article 22 generally protects individuals against solely automated decisions producing legal or similarly significant effects. Exceptions concern contract necessity, authorisation by applicable law and explicit consent, with relevant safeguards.
For the contract and consent exceptions, safeguards include human intervention, an opportunity to express a viewpoint and the ability to contest the decision.
An assistant suggesting interview questions differs from a system automatically rejecting candidates. A reviewer who accepts recommendations without meaningful assessment may not constitute genuine human involvement.
Give reviewers enough information, time and authority to change the outcome. Privacy information must explain relevant purposes and, where required, meaningful information about the logic, significance and expected consequences of automated decision-making under GDPR.
A DPIA is required when processing meets the applicable high-risk threshold. Using AI alone does not determine the answer.
Assess the information involved, scale, affected individuals and likely consequences. Employee monitoring or sensitive-data profiling may create greater risks than an assistant working with public product descriptions.
A data protection impact assessment for AI should examine:
The purpose, data flows and necessity of processing.
Risks such as discrimination, disclosure or inaccurate inferences.
Safeguards and the risk remaining after their implementation.
Swiss FADP Article 23 generally requires prior consultation with the FDPIC where high residual risk remains. A private controller may use the adviser exception when statutory conditions are met and it consults an appropriately qualified, independent data protection adviser.
The exception affects consultation with the FDPIC, not the obligation to conduct the DPIA. Adviser requirements include independence, expertise, absence of incompatible duties and publication and notification of contact details. Appointment alone is insufficient; the adviser must actually be consulted.
Under GDPR Articles 35 and 36, high residual risk requires prior consultation with the supervisory authority. Consulting a DPO does not replace that requirement.
An AI provider’s contract should establish what it may do with personal data, who can access it and what happens when the service ends.
Swiss FADP Article 9 requires the processor to handle data only as the controller is permitted to do. Relevant confidentiality restrictions must allow outsourcing, the controller must verify appropriate security, and subprocessing requires prior authorisation.
Disclosure abroad also requires adequate protection, an appropriate safeguard or an applicable statutory exception.
GDPR Article 28 requires a binding processor arrangement covering specified matters, including instructions, confidentiality, security, subprocessors, assistance, deletion or return, and compliance information. Third-country transfers require a separate Chapter V assessment.
Review the agreement alongside the actual architecture:
Where prompts, files, logs and backups are processed.
Who can access them, including support teams.
Which subprocessors participate and how changes are communicated.
Whether deletion and return commitments cover relevant data stores.
The data processing agreements (DPAs) should reflect those arrangements and assign practical responsibilities.
EU transfers to Switzerland can rely on the EU adequacy decision within its scope; Standard Contractual Clauses are not automatically necessary. Swiss disclosures must be assessed under Switzerland’s own framework.
Review international data transfers across the full provider chain, including onward disclosures. The FDPIC’s cross-border transfer guidance explains the Swiss requirements.
Both laws give individuals rights over personal data, but their conditions and legal routes differ.
Access and correction: Both contain access and rectification rights. Relevant AI data may include conversation logs and profiles linked to a person, alongside original uploads.
Deletion and objection: GDPR expressly addresses erasure, restriction and objection in Articles 17, 18 and 21. Swiss private-sector protection operates through processing principles and personality-rights remedies, including requests to prohibit processing or delete or destroy data. These are not an identical set of rights.
Portability: Swiss FADP Article 28 covers data disclosed by the individual where automated processing is based on consent or directly connected to concluding or performing a contract. GDPR Article 20 covers qualifying data provided by the individual where automated processing is based on consent or contract, specifying a structured, commonly used and machine-readable format. Both rights have limitations.
Build rights-handling procedures around actual data locations. Deleting chat history may leave related information in a support ticket or retrieval index.
Retention should reflect purpose. Temporary uploads, conversation histories and security logs may need different schedules, implemented across active systems and backups.
For trained models, anonymity requires a case-specific assessment. Removing names from training data does not automatically make the model anonymous.
Avoid assuming that every deletion request requires retraining—or that model-related data is always unaffected. Examine the applicable rights and technical circumstances.
Security controls should address uploads, separation between customer environments, document-level retrieval permissions and privileged access. Test whether manipulated prompts or connected tools can expose protected information.
An AI-related breach may involve files exposed to another customer, an accessible conversation archive or unauthorised disclosure through a connected tool.
Notification issue | Swiss FADP | EU GDPR |
|---|---|---|
Authority reporting threshold | Breach likely to create high risk to personality or fundamental rights | Report unless unlikely to create risk to rights and freedoms |
Authority deadline | As quickly as possible | Without undue delay and, where feasible, within 72 hours of awareness |
Informing individuals | Where necessary for their protection or required by the FDPIC, subject to exceptions | Without undue delay where high risk is likely, subject to exceptions |
Processor reporting | Notify the controller as quickly as possible | Notify the controller without undue delay |
The Swiss authority-reporting threshold is higher. Applying a universal “72-hour rule” would also misstate Swiss law.
Where both regimes apply, assess notification separately, preserve evidence and document the response.
If an assistant reveals one customer’s support history to another, establish whose information appeared, its sensitivity, who received it and whether further disclosure is possible. Those facts drive the risk assessment.
A Swiss private business generally does not have to appoint a data protection adviser. GDPR Article 37 requires a DPO in specified circumstances, including qualifying large-scale monitoring or large-scale processing of special-category or criminal-offence data as core activities.
A representative performs a separate role.
Under FADP Article 14, a private controller established abroad must appoint a Swiss representative when processing meets all four conditions:
It relates to offering goods or services to people in Switzerland or monitoring their behaviour there.
It is extensive.
It takes place regularly.
It involves high risk to individuals’ personality or fundamental rights.
Swiss users alone do not establish this obligation. GDPR Article 27 separately addresses representatives for controllers and processors subject to Article 3(2), with limited exceptions.
Certain intentional FADP offences can attract criminal fines of up to CHF 250,000 for responsible individuals. Cantonal authorities handle prosecution; the FDPIC does not impose these fines. A limited statutory mechanism permits a business fine in certain cases where identifying the individual would be disproportionate.
GDPR administrative fines have different tiers. The higher maximum reaches €20 million or 4% of total worldwide annual turnover, whichever is higher. Neither regime treats every operational error as an automatic maximum-fine case.
Begin with a defined use case, such as summarising support tickets.
Identify the data and purpose. Record expected inputs, outputs, affected people and likely sensitive information.
Establish scope and roles. Determine applicable laws, who controls each purpose, and the GDPR lawful basis or Swiss processing analysis.
Inspect the provider. Review training reuse, retention, access, subprocessors, contracts and transfers. If your team lacks the expertise to assess these arrangements, a GDPR compliance service can help review the provider and identify gaps.
Assess safeguards. Decide whether a DPIA, updated privacy information or meaningful human review is necessary.
Test the configuration. Check permissions, deletion, training settings and incident escalation before using real customer data.
Assign responsibility for reviewing changes. A new connector or training feature can alter the original assessment.
Use the GDPR compliance checklist to track wider implementation alongside the AI-specific review.
The EU AI Act adds rules for AI systems and models without replacing GDPR or FADP obligations.
Its high-risk classification and a data protection high-risk assessment answer different questions. A system outside the AI Act’s high-risk categories may still require a DPIA.
As of October 2026, most general AI Act provisions apply. Under the revised timetable, Annex III high-risk requirements apply from 2 December 2027, while relevant high-risk systems connected to Annex I regulated products follow on 2 August 2028. Specific transitional rules still matter.
Assess these duties alongside data protection, including whether the business acts as an AI provider or deployer.
Swiss private businesses do not need consent for every use of personal data in AI. They must follow the FADP’s processing principles and assess whether any infringement of personality rights has a valid justification.
For example, using an assistant to summarise customer correspondence and reusing that correspondence to train a model involve different purposes. When consent is relied on for sensitive personal data or high-risk profiling, it must be explicit.
GDPR compliance does not automatically establish compliance with Swiss FADP. Existing security controls, processing records and privacy procedures can be reused, but they need to reflect Swiss requirements.
Review foreign-disclosure information, Swiss representative obligations and breach-reporting procedures in particular. The two laws also take different approaches to lawful processing, so a GDPR lawful-basis assessment should not be copied across without considering Swiss rules.
Employees can enter customer data when the intended use has been assessed and meets applicable privacy and confidentiality requirements. An approved account alone is not enough; the information uploaded and the provider’s use of it also matter.
Before allowing uploads, establish:
Which customer information is necessary for the task.
Whether prompts or files are retained or reused for training.
Which account settings and access restrictions employees must use.
How accidental uploads should be reported.
Remove unnecessary identifiers before submission where possible.
Swiss hosting tells you where certain infrastructure is located, not whether the service meets every data protection requirement. A locally hosted system can still collect excessive information, retain it too long or reuse it without adequate transparency.
Check the actual processing arrangements, including access from abroad. Hosting is one factor in the assessment, rather than proof of compliance.
A DPIA is required when the processing meets the applicable high-risk threshold. An assistant working with public product descriptions and a system profiling patients require different assessments.
Consider the sensitivity and volume of data, the people affected and the consequences of errors or misuse. Document the screening decision, even when it concludes that a DPIA is unnecessary.
Training may be lawful, but permission to answer a prompt does not automatically cover model training. The provider must assess the training purpose, its role and the applicable processing and transparency requirements.
Check whether training is enabled, which inputs it covers and what an opt-out changes. A statement in the terms that data “may improve our services” does not settle those questions.
Profiles and embeddings can be personal data when they relate to an identifiable person. This includes inferred information, such as a customer’s interests or an applicant’s suitability score, even when the inference is inaccurate.
An embedding is not automatically anonymous because it appears as numbers. Assess whether it can be linked to someone or reveal personal information when combined with other reasonably available data.
A single notice can cover both laws if it addresses their applicable requirements clearly. Shared information can appear together, with jurisdiction-specific explanations where the rules differ.
For an AI service, describe what happens to prompts and uploads, including retention, training reuse and relevant disclosures. The notice should match the service’s actual settings and practices rather than describe every possible use in vague terms.
Türkçe
English
