Compliance · 5 August 2026
AI in the law firm: what you may enter and who is liable
An AI law firm won a case in England - a human had to appear in court. What Swiss law firms must settle before client data goes into an AI tool.
In brief
- On 14 May 2026 an authorised AI law firm won a £7,000 case in England. The AI did the pre-trial work; a human had to appear in court.
- The Swiss Bar Association's AI guidance names three routes by which client data may enter an AI system: on-premise, outsourcing rules, or client consent.
- Liability stays with the lawyer in every case. The guidance leaves no room: you cannot argue that the AI made the mistake.
- The English regulator specifically barred the system from proposing case law - the very task you would expect a law-firm AI to handle first.
As of 5 August 2026. On 14 May 2026 Wandsworth County Court heard the first case prepared by an authorised AI law firm, and ruled in its client’s favour. This article sets out what happened and answers the question it raises for Swiss law firms: what may be entered into an AI tool, and who is liable for the result? I am not a lawyer, and this article is not legal advice - for concrete questions, the cantonal supervisory authority or a specialised firm belongs at the table. Sources at the end.
What happened in England and Wales
On 14 May 2026 Wandsworth County Court heard a freelance HR consultant’s claim against a hospitality business. The court found in her favour, awarded £7,000 and dismissed the counterclaim. The case had been prepared by Garfield AI, a firm authorised by the regulator for England and Wales whose legal work is done by an AI. According to the firm, the claimant paid around £400 in fees.
By the same account, the AI handled all the pre-trial work: correspondence, issuing proceedings, document production, four witness statements and the trial bundles. Shortly before the hearing began, Garfield instructed a junior barrister for the courtroom appearance, Dominic Li of One Essex Court. The hearing lasted three hours.
One detail from the press release is rarely quoted and changes the picture: the hearing was one “in which both sides were represented by barristers”. The defendant had also instructed a solicitor, the English role that prepares a case but does not appear in court. In the courtroom itself, then, the two parties looked identical. What differed was how their cases had got there.
Li put it this way: “the advocacy at trial remained essential and a fundamentally human exercise.” And Garfield’s chief executive Philip Young said something you would not expect from a vendor in his position: “AI did not replace the judge, the barrister or the legal system.”
Two caveats belong with this. The Wandsworth judgment is not published, and the fullest account comes from the winning party itself; the Law Gazette reports the same details but was no closer to the courtroom. More than 600 claims started and more than £500,000 recovered are the firm’s own figures. And whether judgment was given on the day of the hearing or weeks later, the sources tell differently.
What the regulator ruled out from the start
The Solicitors Regulation Authority authorised Garfield as a licensed body on 14 March 2025 and made the authorisation public on 6 May 2025. In doing so it drew three lines that barely feature in the coverage of the court win. They show how a regulator makes such a model authorisable at all.
First, the system does not act on its own: “Garfield is not autonomous and will only take a step where the client has approved it, and furthermore there are supervision and monitoring processes in place.”
Second, responsibility stays with named humans. “Under our rules, named regulated solicitors will still ultimately be accountable for the firm delivering high professional standards. This means they will also be responsible for all the system outputs and for anything that goes wrong.”
Third, the regulator explicitly forbade the system one particular task: “The system will not be able to propose relevant case law, which is a high-risk area for large language model machine learning.” Proposing case law is therefore off the table - the task you would ascribe to a law-firm AI first.
Why the case sounds more radical in the headlines than it is
I checked five German-language reports on the Garfield case, from JUVE through beck-aktuell and Golem to t-online and Business Insider. The usual charge against such coverage does not apply here: all five mention that a human represented the claimant in court. t-online even gives it its own subheading.
What gets lost is subtler. Only JUVE writes “barrister”. The others translate the role as “a young lawyer”, “a junior member of the bar” or “a human lawyer” - and with that, the distinction between solicitor and barrister disappears. That distinction matters here: Garfield is authorised as a solicitor firm and had to hand the courtroom appearance over. A structural limit becomes, in translation, a staffing choice.
Second, agency shifts. The press release says “Garfield then instructed Dominic Li” - the firm instructed. beck-aktuell reports this correctly. Golem and t-online write that the system instructed the lawyer. Third, none of the five carries the sentence that both sides were represented by barristers.
The mechanism is clearest at beck-aktuell itself: the strapline reads “Algorithm instead of lawyer”, and two paragraphs later the text says Garfield instructed a human lawyer. The headline deletes what the body holds.
What you may enter into an AI tool
For Swiss firms the practical question is what may be entered into an AI system at all. The AI guidance issued by the Swiss Bar Association (SAV) names exactly three permissible routes: running the software on your own network, complying with the outsourcing rules, or obtaining a consent and waiver declaration from an informed client. The association’s board adopted the guidance on 14 June 2024; the version in force is dated 16 February 2025.
The first route, under section 2.1, is in-house operation, and it carries two conditions: the software is installed and run on the firm’s own network, and it is ensured that no data whatsoever leaves that internal network or is stored outside the firm’s infrastructure. The second is compliance with the outsourcing rules where the application is obtained through a provider. The third is the consent and waiver declaration from an informed client.
01 · Own network
On-premise
Software runs on the firm's network. No data leaves it, and nothing is stored outside the firm's infrastructure.
No consent required
02 · Provider
Outsourcing rules
The service is bought in. The provider counts as an auxiliary and must be bound to professional secrecy by contract.
Contract required
03 · Client
Waiver declaration
The informed client consents with regard to professional secrecy and data protection law.
Briefing required
Without one of these three routes the guidance is unambiguous: confidential information, trade secrets, personal data of employees, clients or business partners in any format, and content protected by intellectual property rights must not be entered into AI systems.
What has to be settled beforehand, under the same section, is what happens to the input data - in particular who has access to it and where it is stored, including intermediate storage. Which providers contractually exclude training on customer data, and from which paid tier, I have set out in detail in the article on business subscriptions to ChatGPT, Claude and Copilot. The general criteria under the revDSG (the revised Swiss Federal Act on Data Protection) - data location, processing agreement, retention - are in the article on using AI in a Swiss SME.
Why your AI provider is an auxiliary
The term everything hinges on in Switzerland is the auxiliary, in German Hilfsperson. An auxiliary in Swiss legal practice is anyone drawn into the division of labour within a law firm who thereby gains access to client secrets. On the prevailing view cited by the SAV, an external software service can occupy that role too.
The Lawyers Act puts it briefly in Art. 13(2) BGFA: lawyers must ensure that professional secrecy is observed by their auxiliaries. The provision does not subject the auxiliary to professional secrecy directly; it obliges the lawyer to ensure the auxiliary observes it. Enforcement runs through the cantonal supervisory authority, against the registered individual.
Criminal law reaches further. Art. 321(1) of the Swiss Criminal Code lists, alongside other professions, lawyers, defence counsel and notaries as well as their auxiliaries, and covers a secret that was entrusted to them in their professional capacity or that they became aware of in practising it. The penalty is a custodial sentence of up to three years or a monetary penalty. Two points matter for AI use. What was merely perceived in the course of practice is covered too - precisely what a tool sees in passing. And the offence is prosecuted only on complaint.
Renderings of Swiss statutes in this article are mine; the binding texts are the German, French and Italian versions, linked in the sources below.
That a cloud provider falls under this is not in the AI guidance but in the older SAV guidance on IT outsourcing and cloud computing from June 2019, which states that on the prevailing view cloud providers count as auxiliaries (Hilfspersonen) of lawyers. And on drafting: the IT service provider must be told that it is subject to professional secrecy as an auxiliary, and must be bound to confidentiality by contract. What else belongs in such a contract - notice periods, data export, transition arrangements - is set out in the article on vendor lock-in with AI platforms.
Who is liable when the AI gets it wrong
The AI guidance answers the liability question in section 2.3 with a sentence that needs no interpretation: as contractor, the lawyer is liable for any defective performance of the mandate, and cannot argue that the AI made the mistake.
A duty to check comes with it. Section 2.2 states that it is of the utmost importance to review the output independently and critically and to correct or supplement it - and explains why this is not self-executing: an AI is not capable of reviewing the results it has generated itself.
What is at stake is set out in Art. 17 BGFA. The cantonal supervisory authority can issue a warning or a reprimand, impose a fine of up to CHF 20,000, suspend practice for up to two years, or ban it permanently. These do not exclude one another: under Art. 17(2) BGFA a fine may be imposed in addition to a practising ban.
Could an AI law firm be authorised in Switzerland?
The question carries a skewed premise, because Switzerland does not authorise law firms at all. Under Art. 5 and 6 BGFA the cantonal register of lawyers registers individual people, and supervision is directed at the registered person. There is no route here for authorising a firm as such, as the English regulator did with Garfield.
That does not make company structures unlawful for a firm. The Federal Supreme Court recognised the law-firm company in 2012, holding that whether independence is given depends not on the legal form of a firm but on its concrete organisational structure. The limit is control. In 2017 Geneva refused registration to a law-firm company in which one of 39 partners was an unregistered tax expert who held shares and sat on the board; the Federal Supreme Court upheld that refusal.
That answers the original question. A Swiss law-firm company controlled entirely by registered lawyers and using AI as a tool would be permissible. Garfield’s specific structure would not, because the English licensing type presupposes participation by people who are not authorised lawyers. The obstacle is the ownership structure, not the technology.
Frequently asked questions
May I use ChatGPT for a draft contract if I leave the client’s name out?
No, not on that basis alone. What counts under the AI guidance is confidential information, trade secrets and personal data in any format. A draft contract contains terms, party roles and facts from which those involved can be reconstructed. Without one of the three permissible routes the entry remains problematic, even anonymised.
Is a business subscription enough to keep my data out of training?
No, that is only part of the question. The training exclusion concerns onward use of your data. Professional secrecy additionally turns on who has access and where storage happens - both of which the AI guidance requires you to settle beforehand.
Do I have to tell my clients that I use AI?
That depends on the route you choose. Anyone relying on the consent and waiver declaration must inform the client - that is the third of the three routes and cannot be had without it. Anyone running the software on their own network, with no data leaving it and nothing stored outside the firm’s infrastructure, does not need that declaration.
Is there an authorised AI law firm in Switzerland?
None has become known to date, in Switzerland, Germany or Austria (as of August 2026). That does not mean one would not be allowed: all three countries permit lawyer-owned firm companies, and only the circle of owners is restricted. What is demonstrable is simply that no report exists so far.
Practical conclusion
The English case shows both sides. For a claim of £7,000, £400 in preparation costs were enough according to the firm, while the other side instructed a solicitor and a barrister; what their representation cost is not public. At the same time the regulator forbade the system from proposing case law from the outset, and a human took the courtroom appearance. The AI took on the work that needs a lot of time and little judgement.
For a Swiss law firm this leaves an afternoon’s task: go through the tools you use and assign each one to one of the three routes in the AI guidance - own network, outsourcing rules or client consent. Anything that fits none of them must not see client data. Writing that assignment down is the basis for everything else.
For fiduciaries, consultancies and other businesses with confidentiality duties, the AI guidance is not a binding standard - it addresses lawyers, and the list of professions in Art. 321 of the Criminal Code does not name fiduciaries. The same exercise still works as a grid, because the underlying questions are identical: who sees the data, where does it sit, and who carries responsibility for the result. If you want to work that out as a team rather than decide it alone, it is the substance of an AI workshop for your team.
Sources (checked 5 August 2026):
- Garfield AI: AI Lawyer Wins First Court Trial (22 June 2026)
- SRA: SRA approves first AI-driven law firm (6 May 2025)
- SRA register: Garfield.Law Ltd, SRA number 8010904 (retrieved 5 August 2026)
- Law Gazette: AI-powered law firm Garfield claims first county court victory with debt recovery judgment worth £7,000 (23 June 2026, John Hyde)
- SAV: Wegleitung für den Umgang mit künstlicher Intelligenz (adopted by the board 14 June 2024, version in force 16 February 2025)
- SAV: Wegleitung für IT-Outsourcing und Cloud-Computing (June 2019)
- Lawyers Act (BGFA), SR 935.61, Art. 13 and Art. 17 (as of 1 July 2025)
- Swiss Criminal Code, SR 311.0, Art. 321 (as of 12 June 2026)
- BGE 138 II 440: law-firm companies (7 September 2012)
- BGE 144 II 147: control by registered lawyers (15 December 2017)
- beck-aktuell: Algorithmus statt Anwalt - KI-Kanzlei gewinnt ersten Prozess (24 June 2026)
- JUVE: KI-Kanzlei Garfield AI gewinnt erstmals Gerichtsprozess (23 June 2026)
- t-online: KI-Kanzlei Garfield AI gewinnt Prozess vor Gericht in Grossbritannien (23 June 2026)
- Business Insider: KI-Kanzlei gewinnt erstmals Fall vor Gericht (25 June 2026)
Tags