The Swiss legal AI market splits into two groups: international document platforms such as Harvey, Legora or Spellbook, and Swiss research specialists such as Omnilex, Jurata or Swisslex. Comparing both without knowing the difference means comparing tools built for different jobs. Five criteria decide the outcome, and Switzerland weighs them differently than Germany.
Two provider groups, two starting points
The groups differ less in feature count than in where they came from.
International document platforms grew out of contract and transaction work. They review documents, draft clauses, work inside the text. Their legal sources are rarely tailored to a single market, and their target clients are large firms and legal departments handling cross-border mandates.
Swiss research specialists start from the opposite end. Their foundation is the legal source: Federal Supreme Court decisions, cantonal case law, commentaries, journals. Swisslex has been indexing Swiss legal literature since 1986 and has since added semantic search and dialogue-based research. Omnilex and Jurata AI are younger and were built around AI from the start.
The practical consequence: a research tool answers how the Federal Supreme Court assesses a clause. A document platform rewrites the clause. Both are legal work, but they are different steps.
Why German comparison lists mislead in Switzerland
Most German-language overviews are written for the German market. The most comprehensive one lists seventeen tools and bases its data protection assessment on the BRAO and the guidance of the German federal bar. Switzerland does not appear in it.
That leaves out exactly the points where a Swiss selection process fails: professional secrecy under Art. 321 of the Swiss Criminal Code, which by its first paragraph extends to auxiliary persons; the duty under Art. 13 BGFA to ensure auxiliary persons observe it; and the requirements of the revised Data Protection Act for disclosing personal data abroad under Art. 16 revDSG. A tool that passes the German standard can still be ruled out in a Zurich firm.
Which five criteria decide the outcome in Switzerland
Data residency and professional secrecy
The central question is not whether a provider is GDPR-compliant, but where data sits, who can see it, and whether the provider contractually assumes the position of an auxiliary person. Anyone working with client data needs a solid answer on three points: storage location, processing location of the AI models, and how abuse monitoring at the model provider is handled. The last point is regularly overlooked, although it determines whether third parties can view prompts when something goes wrong.
How early this criterion applies is described by Dr. Martin E. Looser, partner at Küng Rechtsanwälte: “For us as a Swiss law firm, it was non-negotiable: our data stays in Switzerland.” The requirement is not technical but one of professional conduct law, and it rules providers out before the feature comparison even starts.
Multilingual coverage across three legal languages
Swiss law exists in three official languages, and decisions are published in the language of the proceedings. A tool covering only German-language sources misses French and Italian case law on the same legal question. For firms with mandates across the language border, that is not a convenience feature but a question of completeness.
Cantonal case law
Federal Supreme Court decisions are widely accessible and included almost everywhere. The difference shows one level down: higher cantonal courts, commercial courts, cantonal administrative courts. In construction, tenancy and employment law in particular, cantonal practice decides the case, and coverage varies considerably between providers.
Source fidelity
An answer without a verifiable citation is useless for a file note. It becomes useful once every statement can be traced back to the specific reasoning and it stays visible what the system relied on. When evaluating, test with a legal question whose answer is already known.
Fit with the existing workflow
Practical value depends on whether the tool works where the document is written. Providers such as Spellbook or CASUS run as an add-in inside Microsoft Word, which removes the switch between applications and the copying back of text. A browser-only interface forces one media break per change, which adds up on a thirty-page contract with forty adjustments.
What the first AI Legal Research Challenge showed
On 17 June 2026, Weblaw ran the first AI Legal Research Challenge at the 12th Weblaw Forum LegalTech in Zurich. Seven providers took part: Omnilex, Opencaselaw, Silex, Jurata, Unplex, LibraTech and Weblaw Lawsearch. All systems received identical inputs, queries were run centrally in one continuous session and fully recorded.
The revealing part is not the outcome but the choice of test criteria. The systems were assessed on how they handle multilingualism, cantonal case law and source fidelity. Those are the same three points where tools diverge in Swiss daily practice, and they appear in no international feature list. Anyone setting up their own evaluation can borrow the setup: identical questions to all candidates, in one session, documented.
Sorting providers instead of ranking them
A single ranking across both groups helps little, because it mixes different jobs. Sorting by focus is more useful:
Focus | Typical providers | Fits when |
|---|---|---|
Swiss legal research | Swisslex, Omnilex, Jurata AI, Weblaw Lawsearch | legal questions and citations drive the daily work |
International document work | Harvey, Legora, Spellbook | transactions and English-language contracts dominate |
All-in-one platform with Swiss legal grounding | CASUS | document work in Word and legal questions arise in the same flow |
Pricing models are only partly comparable. Omnilex publishes CHF 1,600 to 2,000 per user per year; Swisslex and Jurata AI do not list prices. With providers who keep pricing private, ask early about minimum seats and contract term, since both affect total cost more than the per-user rate.
How the selection works in practice
A Zurich commercial firm with twelve lawyers typically proceeds like this: first, identify which step actually consumes time. If it is legal questions without a clear citation, the path leads to a research tool. If it is reviewing draft contracts, to a document platform.
Then comes testing with real matters, not demo data. Three to five closed mandates with known outcomes reveal more within a week than any product demonstration. Contract review comes last: data processing under Art. 9 revDSG, storage location, sub-processor arrangements, and whether the provider contractually accepts the position of an auxiliary person bound by professional secrecy.
A common mistake is evaluating by feature count. A tool offering twenty functions and convincing in none costs more time than one that handles a single task reliably.
Where CASUS fits into this picture
CASUS belongs to the third category: a platform that brings document work and legal sources into the same flow rather than covering one of them.
On the document side there is Risk Review, which analyses risks from each party’s perspective and assigns a severity to every finding, Benchmark for measuring a document against a playbook with a percentage match, Proofread for cross-references, definitions and numbering, and the AI Data Room, which extracts defined fields from many documents into a table. Dossiers keep the context across up to 1,000 documents, so questions can be answered across a whole collection. Changes are applied in Word with the formatting intact.
On the research side, Legal Research covers 1.8 million decisions and articles from the DACH region; for Switzerland, the relevant reasoning passages are highlighted in the result. Data is stored in Switzerland, model inference runs in Europe, and CASUS staff are bound by professional secrecy as auxiliary persons under Art. 321 of the Swiss Criminal Code. Details are on the security page.
Clients include firms from both language regions, among them Studio Legale e Notarile Niccolò Salvioni in Ticino.
Anyone wanting to check the fit can test CASUS free for 14 days. The useful test uses a closed mandate whose outcome is already known. Pricing is CHF 145 per user per month, or CHF 120 with annual billing (as of August 2026).
A broader side-by-side of individual providers is in the independent comparison of the ten best legal AI tools, and a deeper look at research is in the best legal research tools for Switzerland.
FAQ
Which legal AI tools suit Swiss law firms?
It depends on the task. For research in Swiss legal sources, Swisslex, Omnilex, Jurata AI and Weblaw Lawsearch are established. For contract and document work, international platforms such as Harvey, Legora or Spellbook and CASUS come into question. What matters is which step consumes the most time in daily practice.
How does selection in Switzerland differ from Germany?
The decisive rules are professional secrecy under Art. 321 of the Swiss Criminal Code and Art. 13 BGFA rather than the German BRAO, the revised Swiss Data Protection Act alongside the GDPR, and the requirements around multilingual coverage and cantonal case law. German comparison lists do not cover these points.
What is the AI Legal Research Challenge?
A comparative test run by Weblaw, held for the first time on 17 June 2026 in Zurich. Seven providers of legal research solutions received identical tasks under time pressure. The test covered multilingualism, cantonal case law and source fidelity.
How can a firm check whether a tool protects professional secrecy?
Three questions get you there: Where is data stored and where is it processed? Has abuse monitoring at the model provider been waived, so that no human review takes place? Does the provider contractually accept being bound by professional secrecy as an auxiliary person?
Is a general AI tool such as ChatGPT enough for legal work?
For drafting language without client details, it can be. For research it lacks the connection to Swiss legal sources with citations, and for client data it lacks contractual commitments on storage location and confidentiality.
How long does a proper evaluation take?
Two to four weeks are enough when testing with real, closed matters. Longer pilots usually fail not for technical reasons but because nobody finds time for structured testing.
