LITIGATION · INDUSTRY HUB

AI for litigation firms: eDiscovery, case analysis, brief preparation and Swiss case law

How Swiss litigation firms and disputes boutiques use AI in eDiscovery, case analysis, brief preparation and case-law research – with BGFA and Art. 321 SCC as guardrails.

Researched & fact-checked by: DuneDive LLC · As of: 2026-05

Swiss litigation firms and AI

The Swiss bar in 2026 has a clearly identifiable litigation sector. It ranges from boutique firms with three to ten attorneys (e.g. Prager Dreifuss in business litigation, Lalive in international arbitration) to the disputes teams of large full-service firms (Walder Wyss, Baer & Karrer, Lenz & Staehelin, Homburger). Swiss seats in Geneva and Zurich are hubs for ICC, SCAI and ad-hoc arbitration with worldwide reach.

AI in litigation has particular significance in 2026. Unlike pure transactional advice, the work centres on document volume (case files with 10,000 to 1 million documents per matter), tight deadlines (reply, rejoinder, evidence motions) and precedent research across Federal Court rulings, Federal Administrative Court decisions, cantonal case law and foreign sources in cross-border matters. In all three areas AI is productively deployable in 2026 – provided attorney privilege (Art. 13 LFLP, Art. 321 SCC) and client confidentiality are respected.

International legal-tech providers are clearly more present in 2026: Harvey (LexisNexis integration, Anthropic-based), Spellbook for contracts, Hebbia for document analysis, Everlaw and Relativity for eDiscovery. Swiss firms in 2026 mainly test Harvey and their own RAG solutions over the firm's document base. The 2024 SAV guidance provides the frame – see also the separate industry page on AI for law firms for the general sector context.

Why litigation in 2026 cannot do without AI

Three pressure points hit litigation firms simultaneously.

First: document volume is no longer humanly feasible. A large-scale economic criminal proceeding (Banking Act breach, insider trading) or an international arbitration with acquisition background quickly produces 100,000 to 1 million documents. Classic manual review at 30 lawyer-days per 10,000 documents is no longer financeable – clients in 2026 no longer accept an offer that does not include AI support.

Second: legal-research depth is a competitive edge. A boutique with a well-indexed own precedent base plus open Federal Court, Federal Administrative Court and cantonal collections can beat a 50-lawyer full-service firm on specific legal questions. AI RAG over the firm library is the decisive lever in 2026 – unindexed, distributed firm libraries are a speed disadvantage.

Third: client and court expectations. Industrial clients in billion-franc proceedings in 2026 expect real-time reporting functionality, pre-review of opposing briefs in hours instead of days and evidence-motion preparation with clearly documented precedent support. Courts increasingly require evidence motions with electronic review – no longer feasible without eDiscovery software in 2026.

Fourth: 2024 SAV guidance as backbone. It is not anti-AI, it merely demands discipline: client data not in training models, EU/CH hosting with DPA and no-training, pseudonymisation where possible, audit-grade trail. Firms that master this can operate at the litigation depth expected in 2026.

The point: litigation in 2026 without AI means either too high billable hours or incomplete precedent support. Neither is an attractive position vis-a-vis the client and the court.

Where AI works productively in litigation in 2026

Five application clusters cover the bulk of realistically automatable litigation work today.

eDiscovery and document review. From tens of thousands to a million documents a review pipeline extracts relevant material for the proceeding: contracts, emails, internal memos, meeting minutes, technical reports. A classifier tags documents by relevance, privilege (attorney privilege), confidentiality level and time reference. The lawyer team reviews the tags, corrects and iteratively retrains the model. Tools 2026: Everlaw, Relativity AI, Reveal, plus own pipelines on open-weight models with local hosting for especially sensitive matters.

Case analysis and facts synthesis. From released documents an agent generates a structured statement of facts with timeline, actor relationships, critical evidence and open points. The litigator reviews, supplements and refines. The synthesis serves as the basis for briefs and evidence motions. Precondition: all client data go exclusively to EU/CH-hosted models or local hosting.

Brief preparation with disclaimer. From the facts synthesis, the firm's clause library and the precedents found, a first draft of the brief is generated. The lawyer reviews every statement, every citation, every precedent. Important: the brief is signed and owned by the lawyer – it is not "AI-generated" but "lawyer brief with AI preparation". The Federal Supreme Court and cantonal courts in 2024/2025 warned against blind takeover of AI drafts.

CH case-law indexing with RAG. Federal Court collection (1875 to present), Federal Administrative Court decisions, cantonal case law, the federal statute collection and international sources (CJEU, ECtHR, standard sources for Swiss-seated arbitration) are indexed in a proprietary vector database. Questions like "Which Federal Court decisions exist on burden of proof for waiver of claim renewal?" return grounded answers with citations. See Retrieval-Augmented-Generation.

Expert-witness preparation. From files, medical, technical or financial reports an agent generates a briefing note for the expert: relevant facts, open questions, precedents on evidence weighting. The lawyer reviews and supplements. Saves hours of briefing time and makes the expert meeting more efficient.

Across all applications: a multi-LLM gateway with routing by data classification. Client data go exclusively to EU/CH-hosted models (Mistral Large EU, Anthropic Claude via AWS Frankfurt) or local hosting (Llama 3.x, Mistral); general precedent research without case reference may go to cheaper US models.

How a litigation firm starts with AI – in 6 steps

01Inventory and AI guideline: capture all tools already in use (Harvey, Spellbook, Microsoft Copilot, private ChatGPT accounts). Draft an AI guideline based on the 2024 SAV guidance, BGFA Art. 13, Art. 321 SCC. A client-consent clause in the mandate contract.
02CH case-law RAG as pilot: indexing Federal Court, Federal Administrative Court, cantonal case law, federal statutes. Low risk (no client data), high benefit for all lawyers. Four to eight weeks.
03Decide hosting architecture: multi-LLM gateway with routing by data classification. Client data exclusively to EU/CH hosting (Hetzner Zurich, Infomaniak, AWS Frankfurt with DPA) or local hosting (Llama 3.x, Mistral) for especially sensitive matters.
04eDiscovery pilot in a concrete proceeding: explicit client consent, clear procedural goal (review pre-prioritisation, privilege flagging), a documented audit trail. Eight to twelve weeks, three months of accompanied production.
05Own firm RAG over the internal corpus: past briefs, clause libraries, internal memos. Access strictly per case authorisation. Set up consistent with guidance, with logging and source-citation duty.
06Quarterly review and client reporting: hours saved, error rate (especially hallucinated precedents), client feedback. Refine the AI guideline iteratively. Adjust prompts or knowledge base when recurring error classes appear.

Where litigation firms should start in 2026

Three stages, in this order.

Stage 0 – Inventory and AI guideline. Which tools are already in use (Harvey, Spellbook, Microsoft Copilot, in-house scripts)? Which US-hosted tools are used privately by junior lawyers? An inventory and an internal AI guideline based on the 2024 SAV guidance, BGFA Art. 13 and Art. 321 SCC are the precondition. A client-consent clause in the mandate contract.

Stage 1 – CH case-law RAG as pilot. Realistic is indexing of the public Federal Court collection, cantonal case law and relevant federal statutes. Low risk, since no client data are touched; high benefit, since the research tool is available to all lawyers. Four to eight weeks of implementation.

Stage 2 – eDiscovery pilot in a concrete proceeding. After the research pilot, a real pilot in a live proceeding with high document volume. Precondition: explicit client consent, EU/CH hosting or local hosting for especially sensitive content, a documented audit trail. Eight to twelve weeks.

Stage 3 – Own firm RAG over the internal corpus. Past briefs, orders, clause libraries, internal memos are indexed – strictly per case authorisation. This is the workflow with the greatest medium-term leverage: a 10-lawyer boutique with well-indexed own precedent suddenly becomes faster than a 50-lawyer firm without.

Important: every pilot with client data requires written client consent in advance. A standard clause in the mandate contract permitting AI use with EU/CH hosting and offering opt-out is the usual 2026 solution.

Where AI does not belong in litigation in 2026

Three areas where reservation in 2026 is legally and ethically required.

Client data in models without a DPA. A case analysis in ChatGPT-free, an NDA review in Bard without an enterprise DPA, a Federal Court search with identifiable client reference in a US-hosted tool without a no-training clause – all three are 2026 BGFA Art. 13 and Art. 321 SCC-relevant incidents. The 2024 SAV guidance is explicit: client data do not belong in US training models.

Final brief without human last review. A first draft may come from the model. The final version – the statements, the citations, the precedents, the prayers for relief – belongs to the responsible lawyer. "Hallucinated cases" are not theory in 2026 but a documented phenomenon that has already led to sanctions in the US in 2023/2024 and in Europe in 2024. Precedent verification is mandatory.

Voicebot recording of witness statements. Witness questioning at preparation stage (before formal evidence-taking) is delicate. Recording and AI transcription are problematic in law (revised FADP, CrimPC) and in evidence law – an automatically generated transcription can be classified as non-authentic in formal proceedings. Recommendation: classic secretarial minutes or audio recording with explicitly consented, documented transcription.

Particularly delicate and not finally settled in 2026: model predictions of case success probabilities (win/loss probability). Such models exist technically but are ethically and communicatively delicate – they can skew client decisions on settlement willingness. If at all: only as internal discussion basis, never as client recommendation.

Trade-offs

STRENGTHS

eDiscovery review cuts file mountains from weeks to days of lawyer effort
CH case-law RAG brings Federal Court, Federal Administrative Court and cantonal knowledge to all lawyers
Brief preparation noticeably faster – the lawyer focuses on argument, not on text creation
Facts synthesis structures file mountains into clear timelines and actor maps
Boutique firms with well-indexed own libraries beat full-service firms on specific questions

WEAKNESSES

BGFA Art. 13 and Art. 321 SCC force careful hosting and routing architecture
Hallucinated precedents remain a real risk with disciplinary consequences
Client consent must be obtained explicitly – an administrative burden
eDiscovery tools are expensive; ROI only at high document-volume processing
Junior lawyers must still learn manual research and case analysis – a learning risk

FAQ

Are tools like Harvey AI or Spellbook suitable for Swiss firms?

With reservations yes. Harvey is deployed at several major Swiss firms in 2026, including with EU-hosting option via Anthropic Claude on Bedrock. Important: DPA with no-training guarantee, EU data residency and documented security configuration. Spellbook focuses on US law – for CH-specific contract analysis a custom RAG solution over the firm's clause library is usually the better choice. Final review of every output stays with the lawyer.

How do we prevent hallucinated precedents?

Three layers. First: RAG over verified sources (Federal Court collection, federal statutes, cantonal case law) instead of pure model generation. Second: source duty in every output – every cited decision must be findable by case number and link. Third: a verification step by the lawyer team before any brief filing. Hallucinated precedents led to disciplinary sanctions in the US in 2023 and in Europe in 2024 – the verification duty is non-negotiable in 2026.

What does the Federal Supreme Court say about AI use in briefs?

The Federal Supreme Court has repeatedly warned in 2024/2025 against blind takeover of AI output. There is no explicit ban, but responsibility for the content of a brief stays entirely with the lawyer – that includes precedent verification, fact checking and plausibility review. Filing hallucinated decisions risks disciplinary consequences and possibly client liability. 2026 practice: AI preparation – yes; human responsibility – stays full.

How do eDiscovery reviews fit with the revised FADP?

eDiscovery typically touches large volumes of personal data of employees, business partners and third parties. Processing must be proportionate, purpose-limited and traceably documented. 2026 practice: explicit definition of the review purpose, EU/CH hosting or local hosting of the review platform, a documented audit trail, deletion after proceeding closure. For especially sensitive personal data (health, biometric) additional protective measures and a DPIA. See dpia-für-ki-systeme.

Sources

FITS YOUR STACK?

What this looks like in your business – a 30-minute intro call.

Book a call