Professional services firms sell judgement, and judgement is expressed in documents: opinions, filings, audit files, reports, advice letters. Generative AI is very good at producing documents that look right. That combination explains both why law, accounting and consulting firms have adopted AI quickly and why some of the most public AI failures have come from these professions.

For a Swiss fiduciary, law office or consultancy with a handful of partners, the lessons from the past two years are concrete and cheap to apply.

A pattern, not an accident

The early cases came from US courtrooms. In February 2025, Reuters reported that AI-invented case law had led courts to question or discipline lawyers in at least seven cases over two years, including a fine for two New York lawyers in 2023 and a threatened sanction against lawyers at a large personal injury firm. The American Bar Association had already told its members that their duty to stand behind filings extends to unintentional misstatements produced through AI. Andrew Perlman, a law school dean who supports AI in legal work, called unchecked AI citations "incompetence, just pure and simple".

The problem then moved upmarket. In October 2025, The Register reported that Deloitte would repay the final instalment of an AU$440,000 contract with an Australian government department after an academic found fabricated references, phantom footnotes and an invented quotation from a court judgment in a review it had delivered. The corrected version disclosed that a GPT-4o-based tool chain had been used. The department said the substance and recommendations of the review were unchanged.

In June 2026, the Financial Times, republished by Swissinfo, reported that a KPMG publication on agentic AI contained false claims about named organisations, including UBS and Swiss Federal Railways, both of which disputed what had been written about them. KPMG withdrew the report from websites while it investigated and said its staff were required to follow AI guidelines that include human oversight and source verification. The same article noted that EY had retracted a study in May 2026 over fake footnotes.

The detail worth noting is that these were large, well-resourced firms, and in KPMG's case the rules requiring human verification already existed on paper. The rules were not the problem. The checking did not happen.

The Swiss angle

Swiss courts are seeing a related effect from the other side. A Swissinfo summary of NZZ and SRF reporting in October 2026 described Zurich's administrative court receiving far longer, polished complaints, apparently drafted with AI, that the court must still read in full. In two cases it imposed higher procedural fees, in one case above CHF 3,000. The court's president rejected using AI to review the submissions, and official secrecy rules currently prevent the court from using standard AI tools.

The message for advisers is that AI makes it easy to produce more text, and the recipient bears the cost of reading it. Clients, authorities and courts will increasingly notice the difference between work that is long and work that is right.

Why this happens

Language models generate text from statistical patterns rather than by looking facts up. Unless a system is connected to a verified source and constrained to quote from it, a plausible citation and a real one look identical in the output. The risk is highest exactly where professional work is most valuable: specific references, precise figures, named authorities and quotations.

A checklist for a small professional-services firm

  • Separate drafting from sourcing. AI may draft structure and prose. Every citation, figure, statute, ruling and quotation is added or confirmed by a person from the primary source.
  • Click every reference. Before anything leaves the firm, someone opens each cited source and confirms it says what the document claims.
  • Name the reviewer. Each client deliverable has a named person who signs off and is accountable for its accuracy, as before AI.
  • Disclose where it matters. Agree with clients when you will tell them AI was used, particularly for analysis that underpins advice.
  • Protect client confidentiality. Professional secrecy and the Federal Act on Data Protection apply. Client files do not go into consumer AI tools; use systems where you control where data is processed and stored.
  • Write a one-page policy and audit it. A policy is only useful if someone checks, on a sample of real work, that it was followed.
  • Prefer shorter. If AI makes a memo three times longer, edit it back down. Length is not diligence.

The bottom line

AI can take hours out of research, first drafts and document review in a small professional firm. The firms that benefit will be the ones that treat every AI-generated reference as unverified until a person has checked it. The cost of that discipline is small. The cost of skipping it, as several of the world's largest advisers have now learned, is borne in public.

Want to identify where AI can save time in your practice without putting client trust at risk? The free AI Readiness Audit maps your recurring tasks and ranks what is worth automating first.

Sources

  • Reuters (republished by The Star), "AI 'hallucinations' in court papers spell trouble for lawyers", 19 February 2025. link
  • The Register, "Deloitte refunds Aussie gov after AI fabrications slip into $440K welfare report", 6 October 2025. link
  • Swissinfo (Financial Times), "KPMG report contained AI hallucinations on benefits of . . . AI", 12 June 2026. link
  • Swissinfo, "Switzerland Today" newsletter item "For courts, AI can add more work rather than reduce the burden", 15 October 2026. link