AI for Swiss fiduciary offices: a practical guide

The fiduciary sector and AI: an overview

According to the Swiss Federal Statistical Office (NOGA 692000), the Swiss fiduciary sector comprises roughly 12,900 companies providing bookkeeping, payroll, VAT, year-end accounting, tax returns and SME advisory. About 80 percent of these offices employ fewer than ten people. The dominant industry bodies are TREUHAND|SUISSE, with more than 2,000 member firms, and EXPERTsuisse for licensed auditors.

AI is no longer a hype topic in this sector. In 2025, TREUHAND|SUISSE launched TREUHAND|SUISSE GPT, a Swiss-hosted AI offering with cantonally pre-loaded "constructors". Practical seminars on AI automation are running across multiple sections through 2026, and BDO and other large fiduciary firms have published transformation whitepapers.

The question in most offices is therefore no longer "whether" but "where first and how without risk". This page gives a sober overview: which workflows are ready, where the legal limits sit and in what order an introduction makes sense.

Why now

Three pressure points hit the sector simultaneously in 2026. First: the revised Anti-Money-Laundering Act (GwG) is set to enter into force on 1 July 2026. Fiduciaries acting as financial intermediaries (asset management, registered-office mandates, company formation) gain new due-diligence and reporting duties to the new central beneficial-owner register. The transition is twelve months – documentation effort per mandate goes up noticeably. Existing legal entities must report their beneficial owners within one year; this also hits long-running mandates.

Second: skills shortage. The annual cost of one trained bookkeeping FTE in 2026 runs CHF 95,000 to 130,000 gross, and quickly exceeds CHF 140,000 for an experienced clerk with VAT specialisation. Repetitive work – receipt capture, VAT pre-check, payroll triage, dunning – easily ties up one or two full-time positions in a five-person office. That work has moved from routine to half-automated within the last 18 months. If an apprentice or junior cannot be found, freed-up senior time can shift into client-side advisory.

Third: client expectations. Clients expect photo-upload of receipts, VAT quarterly responses in 48 instead of 14 days and transparent interim status. Offices that do not deliver lose SME mandates to software competitors like Bexio Pro, Abacus Cloud and new AI-native challengers that speak directly to end customers and bypass the classic fiduciary office.

The sector point is not "AI replaces the fiduciary" – it is "AI handles triage, the fiduciary decides". That is where the 2026 efficiency boundary actually moves. Offices from five staff upwards that implement this can handle 30-40 percent more mandates without new FTE; smaller offices mainly recover processing time that flows into client advisory.

Typical AI workflows for fiduciary offices

Five workflows cover the bulk of what is realistically automatable today. Each has a dedicated deep-dive in this knowledge base.

Receipt recognition and pre-posting. A client sends 30 PDF and photo receipts a month. An OCR pipeline with a downstream LLM classifier reads supplier, date, gross, VAT rate and proposes an account and a cost centre. The fiduciary checks and books. 3-5x faster than manual capture. See AI receipt recognition with OCR.

Payroll triage. Expense forms, social-insurance notifications, withholding-tax queries and sickness certificates pile up in a shared inbox. An AI classifier sorts by case type, summarises and proposes the next action. Sensitive cases (wage garnishment, pregnancy notification) go straight to a human, unclassified. See AI payroll triage.

VAT preparation. Before quarterly close, an agent checks the general ledger for common error patterns: missing VAT rates, input-tax corrections, mixed-use, own-consumption. The fiduciary receives a "review this" list, never an automatic posting. See AI VAT preparation.

Dunning with escalation logic. Overdue receivables are classified (forgotten, liquidity issue, dispute). Stages 1 and 2 are sent automatically by the agent, stage 3 (legal action) always with fiduciary approval. See AI dunning automation.

AML/KYC screening for new mandates. Before accepting a mandate, an agent checks: PEP lists, sanctions, negative press, beneficial owners through the new transparency register. The output is a risk recommendation, not a verdict. See AI AML/KYC screening.

Underneath all five sits a RAG system over your own guidelines, VAT industry booklets, cantonal tax information and the internal client history. That keeps answers consistent across the office and grounded in your own practice rather than generic model knowledge. More on this in "Your own knowledge with RAG".

How a fiduciary office starts with AI – in 6 steps

1. 01Inventory current tools: which accounting software (Abacus, Bexio, Banana, Sage)? Which CRM? Which mail system? What interfaces already exist? Sketch data flows on a single sheet.
2. 02Prioritise three pain points: which three tasks consume the most staff time per week today? Not the most spectacular – the most regular.
3. 03Legal stocktake: AML status (financial intermediary yes/no), professional-secrecy equivalent (auditor Art. 730b), client contracts checked for data-processing clauses, revDSG record updated.
4. 04Pick one use case as a pilot: realistic starting points are receipt recognition, VAT pre-check or payroll triage. Bring in two willing clients, get written consent to AI usage.
5. 05Settle the hosting question: CH hosting (Hetzner Zurich, Infomaniak, Exoscale) or EU with TIA? Set up a multi-LLM gateway so sensitive data goes to EU models, non-sensitive to cheaper US models.
6. 06Measure the pilot over 12 weeks: error rate, processing time, staff acceptance. Only when all three metrics convince should the next use case begin. Otherwise adjust or stop.

In what order to start

A three-stage sequence has proven itself, ordered not by tool logic but by risk level.

Stage 1 – Light audit. An external stocktake of current processes, software in use (Abacus, Bexio, Banana, Sage, Klara), data flows and legally sensitive points. Output: a report with three to five concrete pilot candidates, effort and benefit estimates. Two to five days. See AI Readiness Audit.

Stage 2 – Pilot with one use case. Realistic choice for a 5-person office: receipt recognition + pre-posting with two willing clients. Four to six weeks of implementation, then three months of accompaniment with weekly reports on error rate, time saved and acceptance. Only when this use case runs cleanly does the next one start.

Stage 3 – Managed service or in-house. Once three to five workflows are live, the step to your own monitored infrastructure makes sense – either as a managed service with monitoring SLA or as a trained in-house setup with a dedicated tech owner. See Managed Service Monitoring and n8n Workflow Automation.

The most common trap: tackling five workflows in parallel. That overloads the team, hides errors and burns the budget. Sequencing and pace discipline matter here more than the tool choice.

Where AI does not yet belong in a fiduciary office

Three areas where AI is best held back in 2026.

Tax assessment at the edge. When a situation is ambiguous in law (relocation with mixed business activity, equity rights in employee shareholding, liquidation considerations with tax deferral), the fiduciary's responsibility is not legally delegable. AI may gather material, pre-structure arguments, look up precedents – the judgement stays with the human.

Confidential client conversations. Crisis mandates (bankruptcy preparation, inheritance dispute, divorce-driven asset split) live on trust. AI in the sense of voice recording or automatic minutes is not only a data-protection risk here, it is a relationship risk. The sector position in 2026: such conversations stay unrecorded or, with explicit consent, in classic notes.

Audit-relevant procedures. If the office performs ordinary or limited audits (Art. 730b CO, EXPERTsuisse standards), the audit procedures themselves cannot be transferred to AI. Preparation, document search and initial analysis – yes. The audit judgement itself and the audit documentation – no.

The rule of thumb: AI handles triage and pre-structuring, the fiduciary handles judgement and responsibility. Where that line blurs, you face either liability or client trust loss.

Trade-offs

FAQ

Related topics

Sources

1. TREUHAND|SUISSE – TREUHAND|SUISSE GPT: KI-Lösung für die Treuhand-Praxis · 2026-03
2. EXPERTsuisse – Weiterbildung: Künstliche Intelligenz für Treuhänder, Wirtschaftsprüfer und Steuerberater · 2025-11
3. EXPERTsuisse / TREUHAND|SUISSE / SVIT – Gemeinsame Stellungnahme zur GwG-Revision · 2026-03
4. Bundesamt für Statistik – Wirtschaftsstruktur Unternehmen (NOGA 692000 Wirtschaftsprüfung und Steuerberatung) · 2026-02
5. Bratschi – TJPG und GwG: Neue Transparenz- und Sorgfaltspflichten ab 2026 · 2025-12