In most banks and lending organisations, payslip verification is still a manual task. An analyst downloads the document, locates the salary figures, and manually enters them into a credit or onboarding system — a process that is slow, error-prone, and nearly impossible to scale. We built an AI-powered payslip recognition system designed to eliminate that bottleneck.
Why Payslip Processing Is Hard
Payslips look simple until you have to process thousands of them across multiple countries. An Austrian payslip formats gross salary as Bruttobezug while a German equivalent might use Bruttogehalt — and a UK payslip calls it Gross Pay. Beyond terminology, the layout differences are significant: some payslips are structured tables, some are dense text blocks, many include handwritten annotations, and some span multiple pages with the key figures buried in a pension or deduction breakdown on page three.
Traditional template-based OCR — where you define fixed zones to read — works only when every document looks identical. For multi-country, multi-employer payslip processing at scale, that assumption breaks immediately.
A Multi-Engine Ensemble Approach
Instead of building a single extraction model, we designed an ensemble that combines three systems and reconciles their outputs through voting logic.
- Azure Document Intelligence — structural OCR with layout awareness, capable of identifying tables, headers, and key-value pairs across diverse document formats
- GPT-based contextual extraction — large language model analysis providing field disambiguation and context understanding across multiple languages
- Custom normalisation layer — EU payroll-specific logic that resolves field name synonyms, handles comma-decimal conventions, and standardises currency representations
The three engines each produce their interpretation of the key fields: gross salary, net salary, employer name, pay period, currency, and deductions. The normalisation layer then compares these extractions field by field. Where engines agree, the result is accepted with high confidence. Where they diverge, the field is flagged for human review rather than silently accepted — which is exactly the right trade-off for financial data that will inform credit decisions.
The Extraction Pipeline in Practice
- Document ingestion — PDF, scanned image, or Excel input
- Format detection and page analysis
- Azure Document Intelligence structural extraction
- GPT contextual extraction with structured JSON output
- Cross-engine field comparison and confidence scoring
- Normalisation: decimal convention, currency, date format
- Structured output with per-field confidence scores
For multi-page documents, the system identifies which pages are relevant to salary data and which contain unrelated payroll disclosures, processing only what matters and discarding the noise.
Results
- Over 95% field-level accuracy across tested EU payslip formats
- Processing measured in seconds per document rather than minutes of manual review
- Multi-page and mixed-currency documents handled end-to-end
- Selective human review triggered only on low-confidence fields, not blanket manual checks across every document
What This Unlocks for Financial Services
- Faster loan approvals — salary data feeds into credit assessment without manual re-entry, reducing approval time from days to hours
- Automated salary verification — consistent, auditable checks against stated income, usable in onboarding and affordability assessment
- Reduced fraud risk — cross-engine validation is significantly harder to manipulate than a single-reviewer document check
- Compliance-ready outputs — confidence scores and field-level attribution mean every extraction decision is fully traceable
This system is one application of TechZiel’s broader capability in intelligent document processing and enterprise AI. If your organisation processes payslips, financial statements, or other high-volume complex document types, speak with us about what a production-ready extraction system could look like for your environment.