AI triage in payroll: pre-sorting client questions on AHV, BVG, and withholding tax

What it is

In a typical Swiss fiduciary office with 5-30 staff, 15 to 60 client queries around payroll arrive each day. The range is wide: registering a new hire with the compensation fund, correcting an AHV-relevant expense item, BVG enrolment of a new employee at 60%, withholding-tax obligation for a French cross-border commuter, family-allowance claim for separated parents. Each requires the right guideline, the right fund and the right deadline.

AI triage in payroll means: an automated pipeline reads every incoming message (email, client portal, optionally WhatsApp), classifies it by case type and urgency, retrieves the relevant guideline (BSV AHV/IV guideline, FTA withholding-tax guideline, cantonal family-allowance rules), drafts a reply and passes the package to the responsible case handler for approval. The human decides, the machine prepares.

The goal is not to replace the case handler. The goal is that each case saves 6-8 minutes of research time and that deadline discipline improves.

Why it matters

Three pressure points make payroll harder in 2026 than five years ago. First: complexity. Withholding-tax rates differ by canton, tariff codes depend on civil status, denomination, children and secondary income. A mistake quickly costs several hundred francs in correction work. Second: staff shortage. Qualified payroll specialists are a bottleneck profession across Switzerland in 2026; the industry magazine TREX already reported in 2023 an average 4-7 months to fill a payroll position. Third: deadline pressure. AHV contributions are due monthly as instalments; BVG enrolments must happen within 30 days of joining; withholding tax is remitted monthly or quarterly.

In the same period regulatory diligence requirements have risen. Art. 957a CO demands gapless, traceable bookkeeping; the revised revDSG (in force since September 2023) demands a documented data-protection impact assessment for automated decision-making in HR contexts; professional secrecy under SCC Art. 321a (which applies to fiduciaries by professional rules) requires confidential handling of payroll data.

AI triage addresses exactly this tension: higher throughput per case handler without sacrificing diligence. The precondition is that the human stays in the loop and every step is logged.

How it works

The pipeline runs in five layers. Layer 1 - intake: n8n receives emails via IMAP, portal messages via webhook, optionally WhatsApp Business via the official cloud API. Every message gets a unique case ID and is logged with timestamp.

Layer 2 - classification: A language model (Claude Haiku for cost control, or Mistral Large for EU hosting) receives the message and a fixed category list: AHV correction, BVG enrolment, BVG mutation, withholding-tax tariff, family allowance, sickness daily allowance, accident insurance, expense question, other. Output is a JSON object with category, urgency (deadline if detectable) and client number.

Layer 3 - RAG lookup: A retriever searches a Qdrant vector database containing BSV guidelines, FTA withholding-tax tariffs by canton, the client file (confidential, local) and the firm's practice notes. The top-8 passages move to the next step.

Layer 4 - draft generation: The language model writes a draft reply in Sie-form, cites the exact guideline passages and lists the relevant deadlines. If no clear answer is possible, the model explicitly returns: "Please consult case handler - guideline not unambiguous."

Layer 5 - human approval: The draft lands in the responsible employee's inbox (Outlook draft, or a fiduciary CRM such as Klara, Bexio or Abacus AbaClik). The human reads, edits, decides whether the reply goes out. Only after approval does anything leave the building.

Abacus payroll software (AbaPay/AbaLohn) has shipped, since version 2024, an open REST API for master data and open pay items. This lets the AI pull the last few payroll statements when a question like "why was the salary lower this month" comes in, rather than answering generically.

Pipeline in 6 steps

1. 01Intake: n8n pulls client mail via IMAP, issues a case ID, writes the audit-log entry.
2. 02Classification: Claude Haiku or Mistral Large assigns the message to one of nine categories (AHV, BVG, withholding, allowances, sickness, accident, expenses, mutation, other).
3. 03RAG lookup: Qdrant returns the top-8 passages from BSV guideline, FTA withholding-tax tariff for the relevant canton, client file and the firm's practice notes.
4. 04Abacus data pull: for statement queries the pipeline retrieves the last three payslips via the AbaPay REST API.
5. 05Draft: language model writes a Sie-form reply with citations and deadline note. On uncertainty it explicitly escalates to the case handler.
6. 06Approval: draft lands as an Outlook draft with the responsible employee. Only after manual review and send does the reply leave the building. Audit log records approver and timestamp.

When to use

AI triage in payroll concretely pays off from about 100 payroll statements per month. Below that threshold manual processing is simply faster than setting up the pipeline. Above 300 statements per month the investment usually pays back within six months.

Concrete constellations where it definitely pays: a Zurich fiduciary with 12 staff and 220 mandates processing about 800 statements per month; a Basel payroll desk with twelve cross-border-commuter mandates and recurring withholding-tax queries; a Ticino office with Italian- and German-speaking clients spending a lot of time on translation steps.

Also suited: constellations with recurring question patterns (BVG enrolment, family-allowance changes on relocation), constellations with deadline risk, and constellations with high client-load per case handler.

When not to use

AI triage is the wrong answer when the office processes fewer than 50 payroll statements per month - setup overhead exceeds the benefit. Also not suited for payroll desks that almost exclusively handle complex special cases (international assignments, group recharges, executive bonuses with equity components); each case is one of a kind and belongs in the hands of a specialist directly.

Not suited if internal guidelines are poorly maintained. A RAG system answers only as well as the source it finds. If the internal payroll checklist last received attention in 2022, the AI will give answers that lag current practice. Clean up first, then automate.

Not suited without prepared data-protection footing. Payroll data is particularly sensitive personal data under Art. 5 lit. c revDSG. Before any model sees it, you need: a written data-processing agreement with the LLM provider, a data-protection impact assessment, an update to client terms, and ideally an EU- or CH-hosted model (Mistral, Aleph Alpha, or local Ollama). A quick hop to OpenAI direct can trigger revDSG breaches with notification obligation.

Trade-offs

FAQ

Related topics

Sources

1. BSV - Wegleitung zur AHV/IV/EO/ALV (WML) · 2026-01
2. ESTV - Wegleitung zur Quellenbesteuerung des Erwerbseinkommens · 2026-03
3. Abacus - REST-API Lohnbuchhaltung (AbaPay/AbaLohn) · 2026-04
4. TREX - Der Treuhandexperte (Fachzeitschrift Treuhand|Suisse), Personalmangel im Treuhandwesen · 2025-11
5. EDÖB - Leitfaden Datenschutz-Folgeabschätzung nach revDSG · 2025-06