AI takes on structured, repeatable work in legal practice: contract analysis, comparison against standards, research in statutes and case law, bulk review of documents, and quality control before a document goes out. Solid benchmarks show a divided picture, and the requirements of professional diligence, legal privilege and data protection remain fully in place.
This article sorts three questions: what is actually established today, where the risks lie, and what applies to Swiss firms under professional rules.
What benchmarks actually show
The evidence is better than the sweeping claims suggest, and more differentiated.
The Vals Legal AI Report of 27 February 2025 tested four legal AI tools across seven task types against a control group of lawyers. AI led in four of the seven tasks: document Q&A at 94.8 against 70.1 percent, document summarisation at 77.2 against 50.3 percent, transcript analysis at 77.8 against 53.7 percent, and data extraction at 75.1 against 71.1 percent. In the remaining three tasks the human comparison group stayed ahead. An October 2025 extension covering classic legal research found the tested systems averaging seven points above the lawyer baseline.
For contract review specifically, the January 2024 study "Better Call GPT: Comparing LLMs Against Lawyers" is the relevant one. It reports F-scores at the level of junior lawyers, with markedly different processing times: senior lawyers averaged 43.5 minutes per document, junior lawyers 56.2 minutes, outsourced review teams 201 minutes, and GPT-4 under 5 minutes.
The qualification is the same in both studies, and it matters more than the headline figures: the advantage arises on structured, clearly bounded tasks. On complex, context-heavy and multi-jurisdictional questions, human judgement remains superior.
Where AI actually applies in a firm
The use cases differ less by practice area than by the structure of the task. Four patterns cover most of the work.
Analysis of a single contract. Systems such as Risk Review identify the contracting parties and analyse from a party perspective rather than generically, with attribution, relevance and severity per finding, plus concrete drafting options. At CASUS a review takes one to three minutes and produces around twenty findings on average (as of August 2026).
Comparison against a standard. Benchmark measures a document against an internal playbook or recognised best practice for SPAs, NDAs and data processing agreements, and names gaps such as a liability clause without a cap or undefined IP ownership.
Reviewing many documents in parallel. In transactions, compliance checks and contract audits, an AI Data Room extracts self-defined fields across large document sets into a table that carries over into Excel and internal reporting. It does not automatically find every clause; it extracts along the fields and clause topics specified.
Research in statutes and case law. At CASUS, Legal Research draws on more than 1.8 million court decisions and statutory provisions from Switzerland, Germany and Austria, including over 660,000 from Switzerland alone, and returns citations rather than general web sources. For Switzerland the relevant considerations are highlighted in the result; German and Austrian law do not know considerations in that sense.
Added to this is pre-dispatch quality control, which checks cross-references, definitions, numbering and open placeholders without touching legal meaning. How these steps play out inside the document itself is covered in the article on Word workflows.
Risks and limits
Hallucinations, including in specialist tools
The best-known risk is the invented citation, and it is not confined to general chatbots. The Stanford RegLab study "Hallucination-Free?" of May 2024 tested commercial legal research tools across 202 legal questions and found hallucination rates of 17 percent for Lexis+ AI, 33 percent for Westlaw AI-Assisted Research and 43 percent for GPT-4. An answer counted as hallucinated even when it stated the law correctly but cited a source that did not support the claim.
The practical conclusion is not abstention but verifiability: a tool that surfaces its citations and lets the reader jump into the source can be checked; a freely composing chatbot cannot.
Mata v. Avianca shows what skipping that check costs. In those proceedings before the U.S. District Court for the Southern District of New York, two lawyers filed a brief citing six court decisions fabricated entirely by ChatGPT. On 22 June 2023 the court sanctioned them 5,000 US dollars. The fault lay not in using the tool but in failing to verify and in standing by the citations after the court had prompted them.
Professional diligence and legal privilege
The professional rules under Art. 12 BGFA apply regardless of the tool in use. An AI output is not legal work product but an aid whose verification remains part of the duty of care.
On privilege, the construction matters for Swiss firms. Under Art. 13(2) BGFA the lawyer ensures that auxiliary persons observe professional secrecy, and Art. 321 Ziff. 1 StGB extends criminal-law professional secrecy to them. A provider processing client matter data has to fit into that construction. Where this is not contractually secured, use is hard to reconcile with the duty of confidentiality.
Data protection requirements
The revised Data Protection Act governs the processing. Two provisions are particularly relevant to AI. Art. 22 revDSG requires a data protection impact assessment where risk is high, and names such risk expressly "in particular where new technologies are used". Art. 25 para. 2 lit. f revDSG grants data subjects a right to information about the existence of an automated individual decision and the logic on which it rests.
For disclosure abroad the Swiss regime stands on its own: under Art. 16 para. 1 revDSG it is the Federal Council that determines adequacy, not the European Commission, and standard data protection clauses under Art. 16 para. 2 lit. d revDSG must be approved, issued or recognised by the FDPIC. EU standard contractual clauses alone do not suffice; they need a Swiss addendum. A fuller treatment is set out in the article on AI and the Swiss Data Protection Act.
Regulation: what applies to Swiss firms
The relevant professional source is the Swiss Bar Association's guidance on dealing with AI, adopted in June 2024 and published in Anwaltsrevue 9/2024; the consolidated version carries the date 16 February 2025. What it requires in detail is covered in the article on the SBA guidance.
The German Federal Bar's guideline is frequently cited instead. It carries the status of December 2024 and was announced on 8 January 2025. It is substantively useful but binds German professional law and is not the governing source for a Swiss firm.
The EU AI Act, Regulation (EU) 2024/1689, takes effect through the mandates: it entered into force on 1 August 2024 and has been generally applicable since 2 August 2026. The AI literacy obligation under Art. 4 EU AI Act and the prohibition of certain practices under Art. 5 have applied since 2 February 2025. In Switzerland itself, the Federal Council adopted the key parameters for AI regulation on 12 February 2025; a consultation draft is to be prepared by the end of 2026.
Selection criteria in brief
Four questions largely decide the choice of tool: where the data sits and whether it is disclosed abroad, whether content is retained after processing or used for training, whether a third party can view the documents, and whether the output carries verifiable citations. To these add whether the tool runs inside the existing working environment, since switching systems costs time day to day.
A detailed comparison of the providers available on the Swiss market, with a full criteria matrix, is set out in the comparison of legal AI providers for Switzerland. Anyone specifically shopping on criteria will find the better entry point there than here.
Firms wanting to test the criteria against a real document can try CASUS for free; the first 14 days cost nothing and carry no minimum term. Such a test only becomes meaningful with a matter from live work.
Adoption inside the firm
Selection is the smaller part of the work; actual use is the larger one.
The return comes fastest on recurring, text-heavy tasks with a recognisable pattern: routine reviews, consistency checks, summaries and standard clauses, along with due diligence and the review of contract portfolios. On standard agreements such as NDAs, data processing agreements and framework contracts, checklists bite hardest.
A pilot with one or two use cases can be set up in hours or days. Reaching stable team-wide use realistically takes some weeks, and the effort then sits in standards, roles and approval processes rather than in the tool.
Two observations from initial conversations with Swiss law firms recur. First, the opening questions almost always concern data protection and security rather than functionality, most often whether documents must be anonymised before processing. With hosting in Switzerland, human review excluded and no retention after processing, that is not necessary. Second, adoption rarely fails on acceptance and often on test setup: evaluating with a template document instead of a real matter keeps the benefit invisible, because a clean template lacks precisely the problems the tool addresses.
Where no approved tool is available, staff reach for whatever is faster to hand. In legal work that is delicate, because client matter documents can end up with third parties unnoticed. The more effective answer is an officially approved tool with clear rules, not a ban.
FAQ
What can AI actually do for lawyers?
AI supports structured, recurring tasks: contract analysis from a party perspective, comparison against standards, initial research in statutes and case law, parallel review of many documents in due diligence, and formal quality checks before dispatch.
How reliable are AI results in legal work?
Reliable only with verification. The 2024 Stanford RegLab study found, across 202 legal questions, hallucination rates of 17 percent for Lexis+ AI, 33 percent for Westlaw AI-Assisted Research and 43 percent for GPT-4. Tools that surface citations can be checked; general chatbots cannot.
Does AI outperform lawyers?
It depends on the task. In the Vals Legal AI Report of February 2025, AI led the human comparison group in four of seven tested tasks and trailed in three. On complex, context-heavy and multi-jurisdictional questions, human judgement remains superior.
May lawyers use AI to process client data?
In principle yes, subject to professional secrecy and the requirements of the revDSG. Under Art. 13(2) BGFA the lawyer must ensure that auxiliary persons observe secrecy, and Art. 321 Ziff. 1 StGB extends it to them. What matters is where data is held, that third-party access is excluded, and that content is not retained after processing.
Does using AI require a data protection impact assessment?
Possibly. Art. 22 revDSG requires an assessment where risk is high and names such risk expressly in particular where new technologies are used. Whether high risk exists in a given case depends on the nature, scope and purpose of the processing.
Which rules apply to Swiss law firms?
Professionally, the Swiss Bar Association guidance on dealing with AI of June 2024, consolidated on 16 February 2025. The frequently cited German Federal Bar guideline binds German professional law. The EU AI Act applies to EU mandates and EU facts; Swiss regulation has been in preparation since the Federal Council decision of 12 February 2025.
What separates ChatGPT from specialised legal AI?
General language models compose freely and surface no verifiable citations. Specialised tools work against collections of statutes and decisions, trace statements back to the source, and analyse by party. Hallucination rates differ measurably, but they are not zero in specialist tools either.
Does legal AI pay off for small firms?
The strongest returns go to teams with high document volume, many versions and recurring standards, which often describes small and mid-sized firms. Where resources are tight, freed-up capacity shows up faster because it does not immediately require new hires.
