fairlane.systems

VAT PREPARATION · USE CASE

AI-assisted VAT preparation: classifying receipts, suggesting input-tax codes, checking the net tax rate method

AI classifies receipts by VAT code, proposes input-tax deduction, and flags net-tax-rate branches. The fiduciary checks, corrects and books.

Researched & fact-checked by: · As of: 2026-05

What it is

The quarterly VAT return is the recurring bottleneck in every fiduciary office. Per client SME 200-2000 receipts arise per quarter: supplier invoices, restaurant receipts, fuel receipts, hotel invoices, hosting bills from EU suppliers with reverse-charge obligation, investments with input-tax correction. Each receipt must be classified: standard rate 8.1% (since 1.1.2024), reduced 2.6%, accommodation 3.8%, input-tax recoverable yes/no, mixed use with allocation key.

AI VAT preparation means: a pipeline reads each receipt's OCR data, proposes the right input-tax code based on supplier and line items, checks plausibility (does the rate match the business purpose?), and passes the pre-booked receipt to the case handler for approval and final posting in the ERP (Abacus, Bexio, Sage, Topal). The machine proposes, the human decides.

Important: the system does not replace VAT due-diligence. It accelerates routine classification. The fiduciary remains professionally responsible for the return under Art. 65 VAT Act.

Why it matters

Three reasons make VAT preparation tedious in 2026. First: rate change since 1.1.2024. Standard rose from 7.7% to 8.1%, reduced from 2.5% to 2.6%, accommodation from 3.7% to 3.8% - to finance AHV. Receipts from the transitional year 2023/2024 must still be allocated cleanly to the service date. Second: mandatory annual input-tax adjustment for mixed use (Art. 30 VAT Act). Third: the net-tax-rate method (Art. 37 VAT Act) is open to businesses up to CHF 5.024 million turnover and CHF 108,000 tax; knowing the right branch codes (e.g. 6.5% for architects, 4.2% for food retail) is a discipline in itself.

At the same time client pressure rises. Many SMEs send receipts by email with a photo rather than sorting them cleanly. Others use Dext, Pleo or proprietary receipt apps - data quality varies widely. The fiduciary becomes the recipient of a chaotic receipt flow and must turn it into a clean VAT return.

From a compliance angle several duties apply: Art. 957a CO demands ordered, traceable bookkeeping. Art. 70 VAT Act demands ten-year retention. The FTA 2024 guideline requires electronic receipts to meet the GeBüV (business-records ordinance), i.e. immutable archival. An AI pipeline must fit into this retention regime, not circumvent it.

How it works

The pipeline has four stages.

Stage 1 - OCR and structuring: The receipt (PDF, JPG, HEIC, optionally ZUGFeRD XML from EU suppliers) is read by an OCR model. For high quality, Google Document AI Invoice Parser, Azure Document Intelligence, or the open-source surya-ocr and marker work well. For Swiss receipts, recognising the QR-bill (SIX standard, mandatory format since 30.6.2020) is the anchor: from the QR code the recipient, IBAN/QR-IBAN, amount and reference number are reliably extracted. The remaining receipt (line items, VAT disclosure, rate) is structured via the OCR model.

Stage 2 - supplier match: The UID (CHE-xxx.xxx.xxx per UIDV) is matched against the ERP master data. If the supplier is new, a creation proposal is generated. If known, the historical booking pattern is loaded.

Stage 3 - AI classification: A language model (Claude Sonnet or Mistral Large) receives the OCR text, the supplier's historical bookings, and the FTA VAT guideline as RAG context. Output is JSON: {rate: "8.1", code: "210", inputTax: true, costCentre: "SALES", reason: "supplier XY is a hosting provider, recoverable for business purpose webshop"}. For reverse-charge cases (EU suppliers without CH VAT number) the acquisition-tax path is flagged.

Stage 4 - pre-posting and approval: The pre-booked receipt lands in the firm's ERP as a draft entry. The case handler reviews, corrects if needed, and approves. The system stores (a) OCR text, (b) AI proposal, (c) final entry, (d) employee decision. This four-element trail is the basis of the audit trail under Art. 957a CO.

For clients on the net-tax-rate method an extra stage runs: the system checks quarterly turnover against the FTA threshold (CHF 5.024 million) and warns once 80% of the threshold is reached - with a note on the methodology change in the following year.

Pipeline in 7 steps

  1. 01Intake: receipt arrives via email, Dext/Pleo connector or client portal. n8n stores the original immutably in a GeBüV-compliant archive (e.g. S3 with Object Lock).
  2. 02OCR: Google Document AI or Azure Document Intelligence structures the receipt. For QR-bills the QR block is read first (amount, IBAN, reference).
  3. 03Validation: UID format CHE-xxx.xxx.xxx is matched against ERP master data. QR-IID is checked against the SIX register.
  4. 04AI classification: the language model proposes tax rate, input-tax code and cost centre, based on supplier history and the FTA VAT guideline (RAG).
  5. 05Reverse-charge check: EU suppliers without a CH UID are flagged as acquisition-tax cases (Art. 45 VAT Act).
  6. 06Pre-posting: the proposal lands as a draft entry in the ERP (Abacus, Bexio, Sage, Topal).
  7. 07Approval and booking: case handler reviews, corrects if needed, approves. Audit log records OCR, AI proposal, final entry, approver.

When to use

Useful from about 500 receipts per client per quarter or about 5,000 receipts per firm per quarter. Concretely: a Lucerne fiduciary with 18 mandates classifying about 8,000 receipts per quarter; a Bern bookkeeping boutique with ten e-commerce clients producing three-digit receipt counts per month; a Ticino fiduciary with Italian-language receipts needing automatic language detection.

Also fits mandates with high receipt frequency from the EU (reverse-charge obligation) - the pipeline reliably sets the acquisition-tax marker.

When not to use

Not for very small mandates (< 200 receipts per quarter) - manual entry is faster. Not when the accounting software has no open API (some older Topal or NestSuite versions) - the import break wipes out the speed gain.

Not without clean master data. If the supplier master in the ERP is incomplete and the UID fields are empty, the AI cannot match reliably. First fill the UID fields, set the VAT code per supplier once cleanly, then automate.

Not for complex special constellations like property sale with ten-year input-tax adjustment, intercompany recharges with transfer-pricing aspect, or VAT group taxation. There the fiduciary stays directly on the receipt.

Trade-offs

STRENGTHS

  • 60-75% time saving per receipt on routine classification
  • QR-bill detection with high hit rate, IBAN/QR-IID validation against SIX
  • Reverse-charge cases flagged reliably (EU suppliers without CH UID)
  • Audit trail under Art. 957a CO structurally clean: OCR original, AI proposal, final entry, approver

WEAKNESSES

  • OCR quality on thermal receipts and HEIC photos varies
  • Master data must be cleaned before rollout (UID, VAT code per supplier)
  • Complex special cases (real estate, intercompany) remain manual
  • Integration with older ERP versions lacking open APIs is laborious or impossible

FAQ

How reliable is the AI on restaurant receipts?

On standard receipts with legible print over 95% correctly classified. It gets harder with hand-written tip entries, very small thermal receipts or EU-source receipts without a CH VAT number. The reduced 2.6% rate for takeaway versus standard 8.1% for on-site consumption requires context, which the AI usually infers from the receipt layout. On doubt the system marks "human check".

What happens for input-tax adjustment on mixed use?

The AI marks the receipt as "mixed use" and proposes a split (e.g. 70% taxable rental purpose, 30% private). The final ratio is set by the fiduciary. At year-end a separate pipeline aggregates all adjustment items and prepares a proposal for the annual correction under Art. 30 VAT Act - with references to the FTA methods (input-tax allocation key, lump-sum methods).

Is GeBüV compliance automatic?

Not automatic, but achievable. Preconditions: immutable original archive (S3 Object Lock, or a Swiss provider such as tresorit, Infomaniak Swiss Backup), traceable change history (every AI change and every human correction in the audit log), and secure identification of the receipt (hash signature). These three pieces satisfy the GeBüV (SR 221.431). The AI as such is not the GeBüV subject - the archive and the audit trail are.

Related topics

RECEIPT OCR · USE CASEAI receipt recognition for Swiss documents: structured capture of QR-bills, receipts and PDF invoicesPAYROLL TRIAGE · USE CASEAI triage in payroll: pre-sorting client questions on AHV, BVG, and withholding taxCLIENT TRIAGE · USE CASEAI triage for client queries: turning WhatsApp, email and phone into structured casesART. 957a CO · COMPLIANCEArt. 957a CO and AI bookings: audit trail, GeBüV, and 10-year retentionRAG ON YOUR OWN KNOWLEDGE · SERVICERAG on your own knowledge: answers from your documents – with sources, not made up

Sources

  1. ESTV - MWST-Info, Steuersätze und Steuerpflicht (Steuersatz-Erhöhung 1.1.2024) · 2026-02
  2. ESTV - Saldosteuersatz-Methode und Branchencodes (MWST-Info 12) · 2026-01
  3. SIX - Swiss QR-Bill Implementation Guidelines v2.3 · 2025-09
  4. GeBüV - Geschäftsbücherverordnung (SR 221.431) · 2024-01
  5. KMU-Magazin - MWST-Vorbereitung in Treuhandbüros: Engpässe und Automatisierung · 2025-10

FITS YOUR STACK?

What this looks like in your business – a 30-minute intro call.

Book a call