LAW FIRM · INDUSTRY HUB

AI for Swiss law firms: what works, what does not, and what the SAV guidance says

How Swiss law firms use AI in contract analysis, case research and client communication – without breaching attorney privilege under Art. 13 LFLP.

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

The Swiss bar and AI: overview

According to data from the Swiss Bar Association (SAV/FSA) and the cantonal bar registers, the Swiss bar comprises roughly 12,500 registered lawyers across some 3,500 firms. Walder Wyss is the largest Swiss firm in 2026 with more than 230 lawyers; the middle ground is boutique and mid-market firms with three to 30 attorneys. The bulk of firms are sole practices or two-partner setups.

The SAV approved a "Guidance on the use of artificial intelligence" on 14 June 2024. It is not binding like the rules of professional conduct but is the principal reference for the sector. The SAV is also dedicating Lawyers' Day on 19 June 2026 at Kunsthaus Zurich entirely to "AI as a tool in the legal profession". The Anwaltsrevue has been publishing technical articles regularly since 2024 on tool selection, liability and secrecy configuration; cantonal supervisory commissions point to the guidance as the state of the sector.

The upshot: AI in Swiss legal practice is no longer a question of whether but how. In 2026, firms are testing contract analysis, case research, dictation transcription, client communication and deadline triage – usually in parallel, often without a clear line. Several mid-market firms nominated their own AI owners ("Director of Legal AI") in 2025/2026; in smaller offices the question stays individual and is therefore more exposed to tool sprawl.

Why the issue cannot be deferred any further in 2026

Three realities are forcing the sector to take a position.

First: attorney privilege is strict, but not anti-AI. Art. 13 LFLP and Art. 321 SCC protect client data comprehensively. The 2024 SAV guidance is explicit: "At the current state of technology, transmission of client secrets to language models such as ChatGPT is not required, since they can be used without such transmission, and sharing would involve incalculable risks for clients." That does not mean "AI forbidden" – it means "client data not in training models".

Second: client expectations are moving. Corporate clients in 2026 expect first-pass contract review in hours rather than days, precedent research with clear source citation and real-time status reporting. A firm that needs twelve to 24 hours for a 40-page NDA review loses mid-market mandates to firms with running AI workflows.

Third: market competition. Klinkler, AdvoSys and Winjur as well as international legal-tech vendors (Harvey, Spellbook, eDiscovery solutions) are entering the firm tech stack directly. A firm that does not pick this consciously will have it slipped in by the IT provider – typically without configuration that fits attorney privilege and revDSG.

The sector answer is not "do not use" but "use deliberately, with clear rules, with hosting in the EU or Switzerland, with pseudonymisation, and with human approval at the sensitive points".

Typical AI workflows in a law firm

Five workflows are mature in 2026 for production use in a Swiss firm – provided hosting and routing are properly solved.

Contract analysis and clause review. An NDA, an SLA or a supplier contract is checked against the firm's clause library. The agent flags deviating clauses, unusual liability arrangements, missing standard clauses and proposes redlines. The lawyer reviews and decides. Precondition: contract material is pseudonymised, or the client has consented in writing to AI usage.

Case and precedent research with RAG. The firm's own library (judgements, briefs, orders, library extracts) is indexed in a vector database. Queries like "Which judgements do we have on non-compete in a C-level transition?" return grounded answers with citations. See "Your own knowledge with RAG".

Client query triage. Incoming emails are classified (new mandate, ongoing matter, general inquiry, complaint), summarised, linked to the known case file and provided with a draft reply. The lawyer decides what goes out. See AI client queries.

Dictation and meeting transcription. Recordings are transcribed locally or in EU hosting (Whisper, voice-agent architecture). The transcription goes into the file; the audio is deleted after review. Sensitive for client conversations – see "When not".

Deadline and task tracking. From an order or pleading, an agent extracts deadlines, hearings and next steps and writes them into the firm software (AdvoSys, Klinkler, Winjur). The lawyer reviews the extraction. Reduces missed-deadline risk – but does not move it away from the human.

Across all workflows: a multi-LLM gateway. Sensitive content goes to EU- or CH-hosted models (Mistral Large EU, Anthropic Claude via AWS Frankfurt, local Llama 3.x); non-critical research can go to cheaper US models.

How a law firm starts with AI – in 7 steps

01Inventory the software landscape: which firm system (AdvoSys, Klinkler, Winjur, iManage)? How do client data flow today? Which US-hosted tools are already in use?
02Draft an internal AI guideline: built on the 2024 SAV guidance, plus professional-secrecy law plus revDSG. Two to four pages, signed by the managing partner.
03Decide on hosting architecture: CH hosting (Hetzner Zurich, Infomaniak), EU hosting (Hetzner Falkenstein, AWS Frankfurt with DPA) or hybrid. Plan a pseudonymisation layer before the model call.
04Draft a client-consent clause: include in mandate agreements, request in writing from existing clients, respect refusal and handle manually-only.
05Pick a pilot use case: contract analysis for a specific contract type, or client-query triage at the inbound mailbox. Four to eight weeks, with clear success metrics.
06Index your own knowledge: after a successful pilot, place the firm library (judgements, briefs, guidelines) into a RAG system. Access strictly per case authorisation.
07Quarterly review: error rate, hours saved, client feedback. Refine the AI guideline iteratively. Stay consistent with the bar association (SAV) and your supervisory commission.

Where a law firm should start

Three stages, in this order.

Stage 1 – Light audit. An external stocktake: which software is in use (AdvoSys, Klinkler, Winjur, in-house DMS), how do client data flow through the network today, which data already leave Switzerland, what internal rules on AI exist. Output: a report with three pilot candidates and a legal-status assessment. Two to five days. See AI Readiness Audit.

Stage 2 – One use case as a pilot. Realistic for a 3-to-10-person firm: contract analysis for a specific contract type (e.g. employment or supplier contracts), with a clear pseudonymisation step before the model call. Four to eight weeks of implementation. In parallel: write the internal AI guideline.

Stage 3 – Scaling with your own knowledge base. Once the first use case runs cleanly, case research with RAG follows. This is the workflow that makes the biggest medium-term difference – a 10-person firm with a well-indexed own library is suddenly faster than a 50-person firm without.

Important: the SAV guidance explicitly recommends that firms issue internal instructions and set rules for AI use. These should come before the first production use – not after.

Where AI does not belong in legal practice in 2026

Three areas where reservation is not "conservative" but legally and ethically required.

Initial client meetings and confidentiality conversations. Criminal defence, divorce mandates, conflict conversations between disputing parties – these live on confidentiality and presence. Automatic voice recording for AI transcription is not only legally delicate here (Art. 13 LFLP, Art. 321 SCC) but relationship-damaging. If needed: a classic file note, or explicitly consented transcription.

Settlement and detail negotiations. When a lawyer sits in settlement talks for the client, judging the other side, the personalities involved and the choice of the next argument is a confidential-human task. AI may gather material beforehand, pre-structure argument trees – the negotiation itself belongs to the human.

Final brief review before filing. An AI draft may handle the first or second pass. The final review before filing belongs to the responsible lawyer, preferably not on screen, but on paper or with four-eye control. The lawyer is liable – the AI is not.

Particularly sensitive: client identification and instruction in criminal proceedings (CrimPC). The Federal Supreme Court explicitly warned in 2025 against blind AI use; the duty of legal instruction remains personal.

Trade-offs

STRENGTHS

First-pass contract review goes from days to hours – the client feels the difference
Your own case library becomes searchable instead of sitting in senior lawyers' heads
Junior lawyers get a better research tool and become productive faster
Deadline tracking becomes systematic – missed deadlines from cascading orders become rarer
The 2024 SAV guidance creates a clear frame – firms can decide with confidence

WEAKNESSES

Attorney privilege forces careful hosting, pseudonymisation and routing
Client consents must be obtained explicitly – administrative overhead
Wrong AI recommendations, taken unchecked, lead to immediate lawyer liability
Initial effort for guideline, audit, pilot: 10-20 days; the AI guideline needs senior time
Juniors might forget how to research manually without AI – a training risk

FAQ

Do we breach attorney privilege by using ChatGPT for research?

If client data (names, facts of the matter, identifiable cases) enters the prompt: yes, with high probability. The 2024 SAV guidance is clear: such data do not belong in a US training model. For purely abstract legal research ("Current case law on non-compete clauses in Switzerland") without case reference: usually unproblematic. As soon as a specific client becomes identifiable, pseudonymisation or a switch to EU/CH hosting with a no-training contract is required.

Which software is compatible with attorney privilege?

It is less about the software than the configuration. A Microsoft 365 tenant with EU data residency and no model training, an Anthropic Claude call via AWS Bedrock Frankfurt with a DPA, a locally hosted Llama 3.x model – all three are compatible if the configuration is clean. A direct OpenAI free account without an enterprise DPA is not. A firm clause library on CH servers plus multi-LLM routing by data classification is the usual 2026 solution.

Do we have to inform clients in advance that we use AI?

The SAV guidance recommends transparency; a legal duty does not exist in general. However, as soon as client data go to third parties (cloud providers) – and that includes EU hosting – this should be open in the mandate agreement. Best practice 2026: a standard clause in the mandate that permits AI use with EU/CH hosting; an opt-out for clients who decline.

Is the firm liable if the AI proposes a wrong contract change?

Yes, without limitation. Liability for legal advice cannot be delegated – neither to staff nor to software. That is precisely why the workflow principle "AI proposes, lawyer decides" is more than mere etiquette. In practice: never pass AI output unchecked to a client; always document lawyer approval in the audit trail; adjust prompts or knowledge base when recurring error patterns appear.

Sources

FITS YOUR STACK?

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

Book a call