2024 was the year of big promises. 2025 was the year of pilot projects. And 2026 — the year when AI actually delivers value in Swiss payroll. But not everywhere, and not in the way the marketing slides claimed.
We run a payroll platform for Swiss SMEs and see every day where AI measurably improves the work — and where it is just window dressing. Here's an honest status report.
Where we stand in 2026
Three things have changed since 2024:
- Models are cheap enough to embed into production workflows. What needed a research budget two years ago now runs at cents per payroll calculation.
- Hallucinations haven't gone away, but they're manageable. With structured prompts, validation layers and confidence scores, risks can be managed well — if you build it seriously.
- Swiss payroll data stays in Switzerland. Microsoft Azure OpenAI in the Switzerland region, Mistral models on-prem, Swiss-hosted embeddings — the compliance hurdle is gone.
The difference between an AI feature that really helps and one that only looks cool: validation layers, deterministic fallbacks, clear confidence thresholds. Ignore those and you build nice demos — not production tools.
5 concrete use cases that deliver value today
We focus on areas where AI demonstrably saves hours or reduces errors. We ignore marketing speak.
When a new employee joins, the AI suggests the correct WHT tariff based on place of residence, religious affiliation, marital status and secondary income — across all 26 cantons. HR only has to validate instead of researching themselves. Saves 10–15 minutes per new hire.
Every monthly salary is compared with the previous 12 months. Plausibility checks flag when a gross salary deviates by 30%, a wage type is missing or a BVG contribution is inconsistent. Before go-live, not after the payout.
"How much holiday do I have left?" — the most frequent HR question in every Swiss SME. The AI chat pulls the answer from the payroll data within seconds. HR is freed from the helpdesk role.
Expense receipts, medical certificates, paternity confirmations — the AI classifies them automatically and suggests a posting or wage type. Manual sorting disappears.
When switching to a new platform, the AI interprets old data structures — even from Excel lists or PDF wage statements — and maps them automatically. That makes onboarding tools noticeably faster.
How we use AI at payrollnow in concrete terms
For us, AI is not a marketing lever but a concrete engineering component — and always embedded in a deterministic payroll engine with ELM 5.1 via partner. Live in production today:
- Onboarding tool with data classification — when switching from third-party systems, the AI interprets old data structures (Excel, PDF wage statements, Bexio/Abacus/SAP exports) and maps them automatically. Migration in under 24 hours, including 12-month back-testing and an optional parallel run.
- Withholding-tax tariff suggestions on new hire — based on master data, place of work, place of residence and family situation. Suggested by AI, confirmed by the account manager. Works across all 26 cantons, including DTA logic for cross-border workers.
- Anomaly detection in payroll runs — every monthly salary is compared with the previous 12 months. Plausibility checks flag deviations, missing wage types, inconsistent BVG contributions — before the payout runs.
- Employee chat in the employee app — self-service answers for leave, wage statement, BVG, expenses. Pulls the answers directly from the payroll data of the asking employee. Typically reduces HR email volume by 60–80%.
- Event-stream architecture — every change (hire, sick, bonus, termination) generates structured tasks for HR, the account manager and the platform. AI classifies, prioritises and suggests next steps.
Important: every AI output comes with a confidence score. Below a threshold (typically 80%), verification by the account manager is required — a person with solid Swiss payroll expertise who carries overall responsibility for your mandate. So we don't let the AI generate payroll data autonomously. We let it suggest, check and document — and a human stays in the loop for every critical posting.
What AI does not replace
Payroll is a regulated field. Use AI where compliance and determinism rule and you're building time bombs.
These tasks don't belong in an AI's hands:
- Social-security calculations. OASI, BVG, accident insurance (UVG), sick-pay insurance (KTG) — those are deterministic rules, formally defined. No LLM intuition, just arithmetic code.
- Applying the WHT tariff. Suggesting yes, applying no. The final tariff choice must be reviewable and audit-ready.
- Wage statements. Structured reports with statutorily prescribed fields. No room for AI creativity.
- Filings to authorities. ELM 5.1, eOASI/DI, FSO — all schema-based XML/JSON transmissions. AI has no business there.
AI is a lever for speed, comfort and anomaly detection. It is no replacement for the regulatory logic that defines the Swiss payroll system.
What 2027 will add
Three trends we're currently working on, scheduled to go into production next year:
- Predictive compliance. AI spots early when cantonal WHT tariffs change or CBA adjustments are coming — and suggests the necessary configuration changes.
- Voice interface. Employees simply ask their payroll question out loud — the app answers. Already works in DE/EN, being extended to FR/IT.
- Custom workflow builder. Accounting firms or HR teams describe a process in natural language, and the AI builds the workflow. Democratises custom integrations.
Frequently asked questions
Does our payroll data stay in Switzerland?
Yes. We host on Microsoft Azure in the Switzerland region, and AI models run in the same data centre. Data does not leave the country.
Can we switch off the AI features?
Yes. Every AI feature is optional and can be disabled in the platform settings. Some customers only use the deterministic payroll engine — that's a valid mode too.
What happens if the AI gets it wrong?
Confidence scores and validation layers catch most errors. In live payroll runs every output is validated against the deterministic engine anyway. If a discrepancy remains, a plausibility warning is raised — no automatic go-live.