AI in law firms and fiduciary offices: three workflows that respect professional secrecy

Starting point: the billable hour is getting scarce

A survey by the Swiss Bar Association and the Swiss LegalTech Association shows how far apart firms are when it comes to their filing: 63% of the firms surveyed work with practice management software, but only 47% with document management software.¹ The survey ran from October 2021 to January 2022 and drew 532 participants. More than half therefore look for documents in the folder tree, in the mailbox or in a paper file. What that costs is visible in an international comparison, a so-called benchmark. Average utilisation there sits at 38% of an eight-hour day: roughly three billable hours, with five hours going unbilled.² That figure comes from international data, not from a Swiss survey; the direction is familiar to every firm nonetheless.

In fiduciary practice the picture is the same, with different labels. A 2016 study by TREUHAND|SUISSE and bexio found that fiduciaries spend more than 60% of their working time on administration; the same study expected that share to fall to around 40% by 2025.³ On top of that comes pressure from the client side: in the 2022 bureaucracy monitor of the State Secretariat for Economic Affairs, 60% of more than 1,500 companies surveyed perceived their administrative burden as rather high or high.⁴ What burdens the client lands on the fiduciary’s desk. There it meets a capacity limit, and that limit does not stretch during the year-end season.

The profession knows this and is looking for the way forward: on 14 June 2024 the Swiss Bar Association adopted guidance on dealing with artificial intelligence and has updated it since.⁵ So the question is no longer whether a firm uses AI, but under which conditions it may. This article is neither legal nor tax advice.

Use case 1: opening the case file at the start of a mandate

A new mandate arrives with several hundred pages: the prior representative’s files, two years of correspondence, enclosures in no particular order, part of it scanned. You load the whole set into the mandate’s workspace. What comes back is a chronology of events, an overview of the parties involved and a list of the disputed points, each with a reference to the place in the file the statement comes from.

The reference is the actual point. You are not reading a summary you have to trust; you jump in seconds to the page where the statement stands. What follows from it is your decision: the legal assessment, the argument and the strategy stay entirely with the lawyer. The AI colleague provides the overview you would otherwise put together over the weekend.

What does not happen: no deadline is calculated and nothing is filed. A language model can miss a consideration or attribute a reference to the wrong place. That is why every output is a draft with evidence for you to check, and not a statement to rely on.

Use case 2: checking contracts and year-end documents against your own practice

In the law firm an NDA, an employment contract or the other side’s standard terms land on the desk. The AI colleague compares the document with your own clause and template library and returns a list of deviations: which clause is missing compared with your template, which one differs, which wording you have already negotiated differently in comparable cases. The result is a pointer, not an approval decision.

In fiduciary practice it is the preparation of the year-end file. Contracts, minutes of the general meeting, loan and leasing agreements are reviewed and condensed into a structured summary; alongside it comes one bundled list of open questions for the client instead of six separate emails over three weeks. A knowledge base of circulars from the Federal Tax Administration, VAT information and your own fact sheets supplies the source reference. That way the question to the client comes with its reasoning.

The limit is clear: the AI colleague gives no legal advice and makes no approval decision. It shows where a document deviates from your practice and names the place in your own library where you can read it up.

Use case 3: client communication in German, French, Italian and English

A firm writes all day long: a request for an extension of a deadline, a follow-up for missing documents, notice of the fee note, a plain-language summary of a judgment or a contract for a client without legal training. Each of these letters is routine and still costs half an hour, because it has to be clean.

From one sentence of context and the firm’s model letters, the draft appears in minutes. It comes in German, French, Italian or English, in the tone your firm uses. That is more than translation: the salutation, the structure and the phrasing come from your own templates and from the earlier correspondence in the mandate. You read, cut, add and approve.

No letter leaves the firm without approval. Who produced which draft and who changed it is recorded in the audit log, the record of every access. Permissions per tool decide who can reach which template library at all.

Connectors: what can be linked to your industry software

An AI colleague is only as useful as what it is allowed to reach. The link to another program is called a connector. It runs over MCP, an open standard for connecting external systems. The prerequisite is a documented interface, an API in technical language: the vendor publicly describes how other programs may retrieve data. Some vendors already ship a ready-made connecting piece for this, an MCP server. Permissions apply to each individual tool, and every call is logged. This is the state of play for common law firm and fiduciary software, checked in September 2026:

Is your software missing from the list? Much of the daily work runs with documents and needs no system connection at all, from opening a case file to correspondence. How the AI connection to a law firm or fiduciary software works in principle via MCP, and which systems are prepared, is described on the integrations page.

Professional secrecy and data protection: the precondition, not the extra

For a law firm the starting point is not a data protection argument but the criminal code. Article 321 of the Swiss Criminal Code names lawyers explicitly, and in the same sentence it names their auxiliary persons. Whoever discloses a secret entrusted to them in their profession faces, upon complaint, up to three years of imprisonment or a monetary penalty.⁸ On top of that comes Article 13 of the Lawyers Act (BGFA): the duty of confidentiality applies without any time limit and towards everyone, and the firm has to make sure that its auxiliary persons observe it as well.⁸ It does not end with the mandate, and it does not end with removal from the register.

From this follows the construction every firm knows that has ever brought in an IT company: an IT service provider is treated as an auxiliary person of the firm and is therefore not an unauthorised third party. The Swiss Bar Association ties this to conditions in its guidance on IT outsourcing and cloud computing: careful selection of the provider, contractual protection of professional secrecy, and use of the data solely to perform the contract.⁹ For an AI tool that means, concretely: your own instance rather than a shared account, that is a system belonging to your firm alone; a contract on commissioned data processing; an explicit commitment that customer data is not used for training; and a disclosed chain of subcontractors. Whether that is enough in a particular mandate is for the firm itself to judge.

Kisuno is built for this. Operation runs in data centres of Green in Lupfig that are certified to ISO 27001, or on your own server. The data processing agreement, which Swiss data protection law calls commissioned processing in Article 9,⁸ states explicitly, in its section on training, that no models are trained on customer data, and the subprocessor list is published openly on the website. Telemetry, the automatic reporting of usage data to the provider, is off by default. Every access is recorded in the audit log, and you can export it. Permissions apply down to individual tools.

Fiduciaries are not in the catalogue of Article 321 of the Criminal Code, but auditors bound to secrecy under the Code of Obligations are.⁸ Anyone running an audit mandate therefore falls under the criminal provision themselves. In the rest of fiduciary work, confidentiality follows from the mandate relationship with the client. On top of that comes data protection law: payroll and personnel data are ordinary personal data, while individual elements such as health information or trade union membership count as sensitive personal data under Article 5 of the Data Protection Act.⁸ And the documents stay around for a long time: books of account, accounting vouchers, the annual report and the audit report have to be kept for ten years under Article 958f of the Code of Obligations, counted from the end of the financial year.⁸ Anyone who pastes payroll data into a chat window whose operator they do not know has a problem that no terms of use can define away.

That is why operation in Switzerland or on your own server is not a preference for a law firm but the precondition for the use to come into question at all. What you need in a conversation with the supervisory authority, the audit client or the client themselves is not a leap of faith but a contract, a list and a log. All three exist. One limitation belongs in the same conversation: Kisuno offers the current models from OpenAI (ChatGPT) and Anthropic (Claude); they run in the AWS data centre in Frankfurt. AWS is a US company, so the US CLOUD Act applies even with servers in Frankfurt, and we say so openly. For customers with higher data-protection requirements, open frontier models such as GLM, Kimi and DeepSeek are available, operated by European companies on servers in Europe; the CLOUD Act does not apply there. Anyone who has to rule that out for a mandate chooses the European model tier with GLM, Kimi and DeepSeek or runs the model and the instance themselves. This article is neither legal nor tax advice.

Frequently asked questions from law firms and fiduciary offices

Sources

All figures in this article come from the following sources, last checked in September 2026. The utilisation figure is marked as an international benchmark and does not come from a Swiss survey.

1. Swiss Bar Association and Swiss LegalTech Association, survey on digitalisation in Swiss law firms, conducted from October 2021 to January 2022, Anwaltsrevue 10/2022: sav-fsa.ch
2. Clio, Legal Trends Report 2025, appendix on hourly rates and key performance indicators; international comparison data, not a Swiss survey: clio.com
3. TREUHAND|SUISSE and bexio, study on digitalisation in fiduciary services (2016), share of administrative work and the expectation for 2025: netzwoche.ch
4. SECO, Bürokratiemonitor 2022 (bureaucracy monitor), media release of 1 March 2023: admin.ch
5. Swiss Bar Association (SAV), guidance for dealing with artificial intelligence, adopted on 14 June 2024, version of 16 February 2025: digital.sav-fsa.ch
6. Vertec, Knowledge Base: Vertec MCP server: vertec.com
7. justitia.swiss, information for developers on connecting to the platform: justitia.swiss
8. Wording of the cited provisions in the Classified Compilation of Federal Legislation: Art. 321 Criminal Code, Art. 13 BGFA, Art. 5 and Art. 9 Data Protection Act, Art. 958f Code of Obligations
9. SAV, guidance on IT outsourcing and cloud computing: digital.sav-fsa.ch
10. Justitia 4.0, factsheet on the entry into force of the BEKJ: justitia40.ch