AI draft for tax optimisation of legal entities

What is the AI tax draft?

A language model connected to Federal Tax Administration (ESTV) guidelines, circulars, and canton-specific practice produces draft variants for typical tax optimisation questions of legal entities. Examples: should the owner-manager of a GmbH receive dividends or salary? How far does the participation deduction (Art. 69 DBG) reach for a holding with a 12-percent stake in a subsidiary? What is the best loss-offset strategy for an AG with a prior-year loss and positive current result?

The use-case is explicitly draft generation, not advice. The model delivers a structured initial calculation with assumptions, source references and sensitivities. The tax advisor reads, checks the assumptions, corrects where needed, adds client-specific factors (family situation, AHV liability, succession planning) and signs the final document. The client never receives the raw model output.

The value lies in reduced preparation time. A 3-hour pre-analysis becomes a 45-minute review of an AI draft. The final tax advisory service stays human, with full professional liability – as Swiss legal-profession rules and FINMA standards require.

Why it matters

Swiss tax law is federal: federal DBG plus 26 cantonal tax codes plus around 2,300 municipal incentive schemes. ESTV guidelines change every year (rate adjustments, practice changes, circulars), and cantonal practice diverges sharply – Zug and Schwyz treat participation disposals differently from Geneva or Zurich. This complexity exhausts fiduciary offices: no one can hold all cantonal practices in their head.

AI-supported drafting addresses this through RAG: the current guideline of the relevant canton is supplied at calculation time. If Canton Aargau publishes a new practice for holding-company tax status in March 2026, that change is in the vector database – and the draft accounts for it automatically.

Second benefit: variant comparison. Instead of one recommendation the model delivers three to five scenarios (dividend 100 percent / salary 100 percent / mixes 60-40, 40-60, 20-80) with exact tax loads, AHV impact and liquidity effects. The tax advisor discusses the variants with the client – the client understands the trade-off better, because he sees numbers instead of recommendations.

Third point: documentation of the decision basis. In a later assessment discussion with the tax administration there is a traceable calculation with source references. If the client raises a liability claim ("you advised me wrongly"), there is an audit trail.

The tax advice itself remains the task of the licensed fiduciary expert. No one should believe the AI is the advisor.

How it works

The pipeline has five components.

Capture client data: Structured input of the relevant figures – financial-year profit, balance-sheet positions, participation quotas, owner salary, prior-year losses, canton of seat. Input comes via form or directly from the ERP. Client identifiers stay pseudonymised; the language model sees "Client 4711", not "Müller AG".

RAG lookup of tax practice: A vector database holds DBG articles, ESTV guidelines (especially KS 5, KS 30 and 31), cantonal tax laws, and relevant circulars. For each query the relevant passages are pulled – typically 8 to 15 chunks per calculation. Sources are cited in the draft.

Calculation engine: The actual tax maths runs NOT in the language model but in a maintained Python library. The model structures the question, calls the right calculation, interprets the result. Example: "Dividend CHF 100,000 in Canton Zug, owner resident in Aargau" – the engine computes Zug withholding tax, Aargau cantonal tax, federal tax, AHV exemption. The model describes the result in plain text.

Variant generation: The model runs three to five scenarios and contrasts them in a table. Per scenario: gross load, net liquidity, AHV impact, sensitivity to +/-10 percent profit change.

Human review and finalisation: The tax advisor reads the draft, checks the assumptions (canton of seat correct? participation quota right? prior-year loss usable?), adds factors the model does not know (family situation, succession plans, AHV optimisation), and signs. The final document goes to the client.

NOT to be sent as an end product: the raw model output NEVER goes directly to the client. Tax optimisation with professional liability requires the review by a licensed person.

Drafting workflow in 7 steps

1. 01Capture and pseudonymise client master data (legal form, canton of seat, financial year, owner setup).
2. 02Structure the tax question: which optimisation axis (dividend/salary, participation deduction, loss offset, seat relocation)?
3. 03RAG lookup: current ESTV guideline, relevant circulars (KS 5, KS 30, KS 31), cantonal practice.
4. 04Call the calculation engine: a Python library, deterministic, tested – do not let the language model do the maths.
5. 05Generate variants: 3 to 5 scenarios with gross, net, AHV impact, sensitivity. Cite sources in the draft.
6. 06Tax-advisor review: check assumptions, add client-specific factors, spot any special situations.
7. 07Finalisation and signature by the licensed fiduciary expert. Keep draft audit trail (model, prompt, sources) for 10 years (Art. 957a CO).

When to use

The use-case fits recurring standard questions: annual dividend-vs-salary optimisation for owner-managed SMEs, holding structuring on participation changes, loss-offset strategy after crisis years, seat relocation between cantons.

The method works well for mandates with a clear legal structure (GmbH, AG, holding with max. 3 participation levels) and a stable business model. With annual repetition the one-time pipeline setup amortises fast.

Particularly valuable in cantons with complex special-tax practice (Zug, Schwyz, Nidwalden, Geneva, Vaud). The gap between federal and cantonal law is widest here, and the RAG lookup on cantonal practice saves significant research time.

For one-off, exotic questions (such as splitting a family holding with 7 heirs across 3 cantons) the model still produces a draft – but the human review takes 4 hours, not 45 minutes. The value drops.

Combination with tax-assessment procedures: the draft can serve as the basis for a pre-ruling with the cantonal tax authority – not as the request itself, but as internal preparation.

When not to use

DO NOT send raw output to clients as a final product. The raw model output is a draft, not an advisory document. Ignoring this risks a liability claim and a FINMA-register entry – the client relies on an "advisor" who is not an advisor.

DO NOT use for special tax cases outside the curated RAG sources: withholding tax for expats with multiple residences, cross-border inheritance optimisation, tax-status switches under holding-privilege abolition. These cases require specialist knowledge not in any ESTV guideline.

DO NOT use when the language model uses US or UK third-country hosting and the mandate is under strict data-protection rules (legal mandate with Art. 321 SCC professional secrecy, FINMA mandate). Tax data is particularly protected personal data under the revised revDSG – third-country transfer needs a TIA and standard contract clauses.

DO NOT use when the client is not informed. Revised revDSG requires transparency: if AI procedures are used in advisory preparation, that must be stated in the mandate contract or the data-protection declaration.

Also DO NOT use as an argument before the tax office ("the AI says"). In the assessment procedure, the legal reasoning counts, not the source of the calculation.

Trade-offs

FAQ

Related topics

Sources

1. Eidgenössische Steuerverwaltung – Kreisschreiben und Wegleitungen Bundessteuern · 2026-04
2. Cosmos Verlag – Steuerpraxis Schweiz (Sammelwerk Bund und Kantone) · 2026-03
3. Bundesgesetz über die direkte Bundessteuer (DBG, SR 642.11) · 2026-01
4. EXPERTsuisse – Steuerkommission Praxis-Empfehlungen · 2026-04
5. Kantonalsteueraemter – Praxis-Festlegungen (Beispiel ZG, ZH, GE) · 2026-05