Payslip extraction

A payslip (in France, the bulletin de paie) is issued monthly by an employer's payroll department or an outsourced payroll provider, and it lands on the desks of HR teams, accountants, lenders, and compliance officers who need to verify income, deductions, and employment status. Mortgage brokers check them for loan applications. Auditors check them for payroll compliance. Fraud teams check them because payslips are one of the most commonly forged documents in circulation. Automated extraction turns a dense, regulation-heavy layout into structured fields that can be validated against known rules and cross-checked against other records, without a human reading every line.

Fields to extract

Why this document is hard to process

The French bulletin de paie is one of the densest single-page documents in circulation. A standard payslip carries 20 to 40 line items covering gross salary, dozens of social contribution lines (health, retirement, unemployment, CSG, CRDS), cumulative year-to-date totals, and paid leave counters, all packed into a table with tiny fonts and inconsistent column widths depending on the payroll software used. There is no single template: large firms often run SAP or Sage, while small businesses may generate payslips from generic accounting tools or, worse, from a Word template someone built a decade ago. Layouts shift accordingly, and template-based extraction breaks the moment a new software vendor is introduced or a company switches providers.

Fraud is the other half of the problem. Payslips are routinely altered to inflate income for loan or rental applications: numbers are edited in PDF editors, fonts are subtly mismatched, totals don't reconcile with the itemised lines above them. Detecting this requires more than reading text; it requires checking internal arithmetic (does gross minus contributions equal net taxable, does net taxable roughly match net paid), checking formatting consistency, and comparing against known payroll software output patterns. Scanned or photographed payslips add noise, and multi-page slips (common for cadres with variable pay elements) require joining continuation pages back to the header data. See document fraud detection for the mechanics.

How Sygnet processes it

Sygnet first classifies the incoming file as a payslip, distinguishing it from adjacent document types like bank statements or contracts, then infers the schema based on payroll structure rather than a fixed template. This matters because payslip layouts vary by software vendor and company size, and a rigid template approach fails the first time a new format appears; our OCR vs VLM comparison explains why a multimodal model handles this better than classic OCR pipelines.

Each extracted field carries a confidence score and a bounding-box reference back to its exact location on the source page, so a reviewer (or an audit trail) can see precisely where "net salary" or "employer SIRET" was read from. This field-level provenance, described in more detail at field-level provenance, is what makes automated extraction trustworthy for financial documents rather than a black box.

Validation rules run automatically: checksum checks on the social security number, arithmetic reconciliation between gross, contributions, and net, IBAN checksum validation, and SIRET lookups against the national registry. Where a payslip needs to be checked against another document, for example matching declared income against a bank statement deposit, Sygnet applies cross-document validation. Fields below a confidence threshold get routed to a human reviewer instead of being auto-approved; everything else flows straight through as structured output. Output is delivered as clean JSON, ready to plug into payroll, lending, or compliance systems, see structured output.

Validation rules worth applying

• Verify the social security number checksum against the INSEE key algorithm.
• Reconcile gross salary minus total contributions against net salary, allowing for standard rounding.
• Confirm SIRET format and cross-check against a live business registry.
• Check that the pay period does not overlap with a previously submitted payslip for the same employee.
• Validate IBAN checksums if bank details are present on the slip.
• Flag payslips where net taxable income falls outside the expected range between net and gross.
• Compare year-to-date cumulative totals across consecutive months for the same employee to catch inconsistent jumps.
• Check font and formatting consistency across numeric fields to flag likely edits.

FAQ

Can payslip extraction detect fraud on its own?

It can catch a large share of common alterations by checking arithmetic consistency, checksum validity, and formatting anomalies, but it isn't a complete fraud solution. Sophisticated forgeries that recompute every line correctly will pass automated checks. Extraction should be treated as a first filter that flags suspicious documents for human review, not a final verdict.

How does Sygnet handle payslips from unfamiliar payroll software?

Sygnet infers structure from the document itself rather than matching against a fixed template, using a multimodal model that reads layout and text together. New formats from unfamiliar payroll vendors are handled the same way as familiar ones: fields are located by meaning, not by fixed position, so accuracy doesn't collapse when a new software vendor appears in the pipeline.

Is it safe to process payslips containing sensitive personal data?

Payslips contain sensitive personal and financial data, so processing them requires attention to data residency, retention, and access controls. See our security and compliance page for details on how data is handled, and GDPR and document processing for the legal framework governing this category of documents.