Using generative AI in legal work compliantly means securing data sovereignty, meeting data protection law (Swiss revDSG and GDPR), ensuring no client data is used for model training, and keeping human oversight over AI-assisted outputs. CASUS, a Swiss legal AI platform, retains no data after a session (Zero Data Retention), hosts exclusively in Switzerland and the EU, and transfers no data to the US.
Why compliance in legal AI is not optional
Generative AI is entering law firms and in-house legal departments for contract review, research, and document analysis. The regulatory picture is sharpening faster than many teams anticipated: Switzerland’s revised Federal Act on Data Protection (revDSG / nDSG) has been in force since 1 September 2023, the EU’s GDPR applies to many Swiss organisations with EU-facing operations, and the obligations of the EU AI Act have been phasing in since 2025; following the Digital Omnibus on AI, the high-risk obligations for Annex III systems apply from 2 December 2027.
Deploying generative AI in legal work without addressing compliance questions risks data breaches, loss of client trust, and fines of up to CHF 250,000 under Art. 61 FADP, for example where processing is entrusted to a processor without meeting the requirements of Art. 9 FADP. That applies to small teams and in-house legal functions just as much as large firms.
Under the Swiss FADP, cloud processing of personal data – even where the cloud provider is located in the EU – remains subject to the processing principles of Art. 6 FADP. Processing may only be entrusted to a processor by contract or by law, and only if the data are processed as the controller itself would be permitted to and no statutory or contractual duty of confidentiality prohibits it (Art. 9 FADP). The rules behind these checks are explained in our guide to AI and the Swiss Data Protection Act, including revDSG duties, contracts, and vendor-risk questions.
According to the TrustArc Global Privacy Benchmarks Report 2025, AI is now the single biggest privacy challenge organisations report (47%), and there is a 16-point gap between organisations that consider themselves AI-ready and those that do not. AI can help close that gap, but only if the deployment itself meets applicable standards.
What “generative AI legal compliance” actually means
Generative AI legal compliance describes the lawful deployment of AI systems that use large language models (LLMs) to generate, analyse, or summarise text. In a legal context, that means contracts, pleadings, legal opinions, or compliance documents.
Compliance here is not only about the AI system being technically secure. It covers three levels:
Data protection: Where does data go? Who has access? Is client data used for model training?
Liability and professional duty of care: Who is responsible for AI-generated outputs? What oversight do lawyers retain?
Copyright and ownership: Who owns AI-generated text? Can it be legally used and attributed?
These are not hypothetical questions. In March 2025, the U.S. Court of Appeals for the D.C. Circuit confirmed in Thaler v. Perlmutter (No. 23-5233, decided 18 March 2025) that copyright protection requires human authorship; purely AI-generated works are not protectable. For pleadings and contracts produced under professional liability, that distinction is directly relevant.
On 12 February 2025, the Swiss Federal Council decided to implement the Council of Europe’s AI Convention in Swiss law, mainly through sector-specific amendments, for example on transparency, liability and risk assessment; a consultation draft was announced for the end of 2026. Until then, existing law applies to AI-assisted legal work: the FADP, the professional rules under Art. 12 of the Swiss Federal Lawyers Act (BGFA) and, for liability towards clients, Art. 97 in conjunction with Art. 398 OR (Swiss Code of Obligations).
The compliance checklist: what to clarify before deploying AI
Data protection and data sovereignty
Where is data processed? (Switzerland / EU, or US / third countries)
Is client data used to train the underlying model?
Is Zero Data Retention in place – meaning no storage after the session ends?
Is No Human Review guaranteed – no vendor staff access to content?
Are data processing agreements in place under Art. 9 revDSG and/or Art. 28 GDPR?
For transfers abroad: is the destination listed as adequate by the Swiss Federal Council (Annex 1 of the Data Protection Ordinance, e.g. US companies certified under the Swiss-U.S. Data Privacy Framework), or are standard contractual clauses recognised by the FDPIC in place?
Model and vendor assessment
Which base model is used, and how transparent is the vendor about it?
Is there ISO 27001 certification or comparable documented security evidence?
Are security audits or penetration tests conducted and documented regularly?
Internal governance and bar association rules
Is there an internal AI usage policy?
Are staff trained to critically review AI outputs?
Are AI-generated results always approved by qualified lawyers before entering work product?
Do workflows reflect the duty of care (Art. 12(a) BGFA), the duty to perform the mandate personally (Art. 398(3) OR), the Swiss Bar Association (SAV/FSA) code of conduct and its guidance on AI?
Copyright and IP
Does AI use in drafting pleadings or contracts require disclosure to the court or client?
Are no copyrighted texts reproduced without review?
Legal teams that want to work through these points systematically can try CASUS at no cost. The platform is built to meet these requirements from the ground up: Sign up for free.
DSG vs. GDPR: a risk matrix for a Zurich-based firm with EU clients
For a law firm based in Zurich that advises both Swiss and EU clients, two legal regimes apply in parallel. The table below maps where obligations diverge and where they overlap – a framing that does not typically appear in generic legal-AI commentary.
Topic | Swiss revDSG (since 1.9.2023) | EU GDPR (article) | Practical consequence for Zurich firm |
|---|---|---|---|
Legal basis for processing | Art. 6 revDSG: lawfulness, good faith, proportionality, purpose limitation | Art. 6 GDPR: lawful basis (consent, contract, legitimate interest) | Where the GDPR applies (Art. 3 GDPR), each processing activity needs a legal basis under Art. 6 GDPR; the revDSG has no equivalent permission principle but applies processing principles instead |
AI vendor as processor | Art. 9 revDSG: contract or law (no written form required); controller must ensure the processor can guarantee data security | Art. 28 GDPR: written contract, sub-processor authorisation | One processing agreement can cover both regimes if it meets Art. 9 revDSG and, where the GDPR applies, Art. 28 GDPR |
Breach notification duty | Art. 24 revDSG: notify EDÖB if high risk; no fixed deadline (promptly) | Art. 33 GDPR: 72-hour deadline to supervisory authority | The GDPR deadline is stricter; a single process covering both regimes is needed |
Third-country transfers | Art. 16 revDSG: adequate protection or SCCs | Art. 44 et seq. GDPR: adequacy decision or SCCs | Since 15 September 2024, US companies certified under the Swiss-U.S. Data Privacy Framework are recognised as adequate (Annex 1 DPO); under the GDPR, the EU-U.S. Data Privacy Framework applies. For non-certified US recipients, SCCs and a transfer risk assessment remain necessary; with CASUS this assessment does not arise, as no data is transferred to the US |
Client access rights | Art. 25 revDSG | Art. 15 GDPR | Broadly similar in substance but deadlines and exceptions differ |
Sanctions | Art. 60–63 revDSG: fines up to CHF 250,000 against the responsible individuals (intentional conduct only) | Art. 83 GDPR: up to EUR 20m / 4% annual turnover | The revDSG primarily targets individuals; the business itself can be fined up to CHF 50,000 (Art. 64 revDSG). GDPR fines target the legal entity |
Profiling / automated decisions | Art. 21 revDSG: information, right to state one’s position and to request review by a natural person | Art. 22 GDPR: right to human review | Relevant only for decisions based solely on automated processing with legal effect on individuals; an AI-assisted contract analysis reviewed by a lawyer is not such a decision |
What this means in practice: A Zurich M&A boutique advising a German corporate group on a Swiss acquisition must choose AI tools that technically satisfy both regimes. Legally, transfers to the US are permitted under both regimes to recipients certified under the Data Privacy Framework or with additional safeguards such as SCCs. That still requires assessment, contracts and ongoing monitoring, and for professional secrecy holders the risk of access by US authorities remains. A provider like CASUS with hosting exclusively in Switzerland and the EU removes these questions: no data is transferred to the US.
How CASUS implements compliance requirements technically
CASUS implements the technical requirements above at the architectural level – not as an afterthought.
Zero Data Retention: Inputs and outputs are not stored after a session ends. Client data is never used for model training.
Hosting in Switzerland and the EU: No data is transferred to the US, not even via sub-processors. Legally, US transfers would be possible under the Data Privacy Framework or with SCCs; CASUS goes further and rules out transfers to the US and the risk of access by US authorities from the outset.
No Human Review: Neither CASUS staff nor third parties have access to content. Abuse monitoring can be disabled via opt-out.
Traceable outputs: In Legal Research, CASUS references specific sources drawn from more than 8 million legal documents from six countries. Outputs are source-based, structured, and traceable. Relevant reasoning sections from decisions are shown inline in answers, with no click-through required. That lets lawyers check every source immediately and makes it efficient to meet the duty of care under Art. 12(a) BGFA.
For Swiss law, CASUS also searches open-access commentary from Onlinekommentar and company information from ZEFIX, the Swiss Central Business Name Index. View the current source overview.
One concrete workflow example: an in-house lawyer at a Basel-based pharmaceutical company reviews incoming supplier contracts against the company’s internal standard. Using the Benchmark module, she compares an incoming draft against the internal playbook and receives a structured report within minutes: which clauses are missing (for example, no liability cap, IP ownership undefined), which deviate, and a percentage match score. The legal judgment and sign-off remain hers; what changes is the initial analysis: it is ready in a structured form within minutes, instead of a manual clause-by-clause comparison.
In AI contract review, risks are assessed from each party’s perspective and prioritised by severity (low / medium / high). That lets lawyers focus review time where it matters, while retaining final responsibility.
What remains legally unsettled – and why it stays relevant
Generative AI in legal work still operates in a partially unresolved regulatory space. The EU AI Act classifies AI systems by risk level; high-risk applications under Art. 6 EU AI Act – including AI systems intended to be used by or on behalf of judicial authorities (Annex III, point 8) – face transparency obligations, conformity assessments, and logging requirements. Typical law-firm use for research or drafting is generally not high-risk, but organisations within the Act’s scope have been subject to the AI literacy duty (Art. 4) since 2 February 2025. Following the Digital Omnibus on AI, in force since 27 July 2026, the Annex III obligations apply from 2 December 2027 and the Annex I product rules from 2 August 2028. For Swiss firms with EU clients, this can become directly relevant depending on how the system is used.
In Switzerland, liability for AI-generated legal work is governed by Art. 97 OR (breach of contractual obligation) and Art. 41 OR (tort). A law firm that passes AI output into a pleading without review and thereby harms a client faces a claim under Art. 97 OR – the breach of professional duty of care would be difficult to contest. Under established Swiss professional-conduct principles, a lawyer’s duties of care are not reduced by the use of technical aids; the lawyer remains responsible for the correctness of the result. This principle predates generative AI but applies directly to it.
The question of platform liability for AI-generated content is also unresolved in Switzerland. There is no equivalent to the US Section 230 of the Communications Decency Act in Swiss law; platform liability would be assessed under general OR principles. For law firms, this means liability exposure does not shift to the AI vendor by contract alone – it stays with the professional who signed the work product.
Practical implications – what actually goes wrong
Technical compliance is solvable. What is more often underestimated in practice are the substantive pitfalls that LLMs introduce specifically in the Swiss legal context.
LLMs trained on broad international datasets have a known weakness with jurisdiction-specific limitation and notice periods. One concrete example: under Art. 210 OR, warranty claims in Swiss sales contracts are subject to a two-year limitation period from delivery - which is distinct from the duty under Art. 201 OR to give notice of defects immediately upon discovery. Both are regularly conflated with each other, and with the general ten-year limitation period for contractual claims under Art. 127 OR. LLMs tend to smooth over this distinction and produce clause language calibrated to German law or US warranty concepts rather than to Swiss OR. Firms that adopt AI-generated contract drafts without review can expose clients who rely on the two-year period without realising the clause actually sets a different standard.
The same applies to default interest clauses: the statutory rate under Art. 104 OR is 5%. Models trained predominantly on international data sometimes replace this with EURIBOR-based or US-prime-rate references, simply because those appear more frequently in training data.
Neither of these failures is an argument against using AI. They are arguments for qualified review. The AI Chat with Agent Mode can be used to identify such clauses and check them against Swiss OR standards – but the final judgment requires someone who knows the OR.
Tension with bar association rules: Lawyers must perform the mandate personally (Art. 398(3) OR) and with due care (Art. 12(a) BGFA), and must protect professional secrecy (Art. 13 BGFA, Art. 321 Swiss Criminal Code). Both create friction with AI deployment. Personal performance does not prohibit technical tools, but it requires the lawyer to substantively own the result – quality review is not optional. Legally, it is sufficient to bind the AI vendor to confidentiality as an auxiliary person and to observe the outsourcing rules (Art. 13(2) BGFA), or to obtain the informed client’s consent. CASUS goes further: Zero Data Retention, no training on client data, no human review and hosting exclusively in Switzerland and the EU. Mandate information is not stored after the session and is seen by no one, which minimises the disclosure risk from the outset.
The Swiss Bar Association has published its own guidance on the use of AI (SAV board, 14 June 2024; updated version February 2025). It recommends an internal AI policy, independent review of outputs and compliance with the outsourcing rules, and points to a possible duty to inform clients. In addition, the existing code of conduct applies in full.
Next steps: putting compliance into practice
Compliance in AI deployment is not a one-time project. The technical foundation – Zero Data Retention, hosting in Switzerland and the EU, no data transfer to the US, no vendor staff access to content – is built into CASUS from the start. Teams that work with source-based legal research or structured contract review get traceable outputs that integrate into existing quality assurance processes.
More on CASUS security and data protection standards: Security. Platform overview: About.
Those who want to try the platform without commitment: Sign up for free.
FAQ
What is generative AI legal compliance?
Generative AI legal compliance describes the lawful use of AI systems that generate or analyse text in a legal context. It covers data protection conformity under revDSG and GDPR, liability questions under Art. 97 OR, transparency obligations, and internal governance – including professional duty obligations under the SAV/FSA code of conduct and Art. 12 BGFA.
Can generative AI legally process client data?
Yes, provided the applicable requirements are met: a processing agreement with the vendor under Art. 9 FADP and, where the GDPR applies, Art. 28 GDPR; lawful transfers abroad (Art. 16 FADP, Art. 44 et seq. GDPR); a vendor bound to confidentiality as an auxiliary person (Art. 13(2) BGFA); and no training on client data. Where the GDPR applies, each processing activity also needs a lawful basis under Art. 6 GDPR. CASUS goes beyond these minimum requirements: hosting exclusively in Switzerland and the EU, no transfers to the US, Zero Data Retention and no human review.
What does Zero Data Retention mean for AI tools?
Zero Data Retention means inputs and outputs are not stored after a session ends. The vendor holds no data and does not use it for model training. In CASUS, this is enabled by default. For professional secrecy under Art. 13 BGFA, the law does not require Zero Data Retention but a confidentiality undertaking by the vendor as an auxiliary person; Zero Data Retention goes further, because no mandate data remains with the vendor after the session.
Must AI use be disclosed in court filings or to clients?
In Switzerland, there is currently no statutory disclosure obligation to courts or opposing counsel. The duty of care under Art. 12(a) BGFA does require lawyers to personally own the work product they submit; passing AI output on without review would be professionally problematic. Whether Swiss courts will introduce a formal disclosure requirement is an open question. Towards clients, the SAV guidance on AI points to a possible duty to inform, for example where the client requires or legitimately expects the lawyer to perform the mandate personally.
Are AI-generated texts protected by copyright in Switzerland?
The Swiss Copyright Act (URG) requires human creative activity as a precondition for protection. Under prevailing academic opinion, purely AI-generated texts do not qualify. The Federal Supreme Court has not yet ruled on the question directly. For comparison, the US Court of Appeals for the D.C. Circuit confirmed in Thaler v. Perlmutter (No. 23-5233, 18 March 2025) that copyright requires a human author. Anyone submitting AI-generated text as their own work carries the IP risk.
What drafting errors do LLMs typically make in Swiss contracts?
LLMs trained on broad international data regularly miss or conflate Swiss-specific rules. Known failure patterns: mixing up the two-year warranty limitation period under Art. 210 OR and the duty to notify defects immediately under Art. 201 OR with the general limitation period under Art. 127 OR (ten years); replacing the statutory default interest rate of 5% under Art. 104 OR with EURIBOR or US-prime references; producing IP clauses calibrated to US copyright structures rather than URG. These errors are avoidable through qualified review – they are not unavoidable features of AI use.
How does CASUS differ from tools like Harvey or ChatGPT?
CASUS hosts exclusively in Switzerland and the EU, transfers no data to the US, retains no session data, and is built specifically for Swiss law firms and in-house legal teams. With tools such as Harvey or ChatGPT, processing location and retention depend on the product version and contract; for US-hosted setups, firms must themselves assess and contractually secure third-country transfers, possible access by authorities and retention periods, in addition to the confidentiality undertaking of the vendor under Art. 13 BGFA. With CASUS, third-country transfers and retention periods are not an issue, and research draws on more than 8 million legal documents from six countries, including Swiss legislation and case law.
What should an internal AI usage policy for a law firm include?
At minimum: which tools are permitted (with verification against Art. 9 revDSG and Art. 28 GDPR), how outputs are documented and labelled, when manual review is mandatory, how Swiss OR-specific deviations in AI-generated drafts are handled, and who holds accountability for liability questions under Art. 97 OR. Where the GDPR applies (Art. 3 GDPR): an assessment of the lawful basis under Art. 6 GDPR for each processing activity.
Does the EU AI Act apply to Swiss law firms?
It applies directly only to entities with EU operations or EU market exposure. Swiss firms advising EU clients or providing services in the EU may fall within scope. Following the Digital Omnibus on AI, the high-risk obligations for Annex III systems apply from 2 December 2027; typical AI-assisted legal advice by law firms is generally not a high-risk use case, whereas AI systems used by or on behalf of judicial authorities can be. Switzerland is not adopting the EU AI Act; in February 2025 the Federal Council decided to implement the Council of Europe’s AI Convention through mainly sector-specific amendments and announced a consultation draft for the end of 2026.
