Upload
Standard mode vs AI mode
Understand the difference between OCR-based and AI-based processing and when to use each.
Updated: 2026-02-18
Before you start
- File ready to upload
- Entity type confirmed
Open related features
Standard mode overview
Standard mode uses PaddleOCR (PP-OCRv4) to extract text from your document, then passes the text through a deterministic parsing engine. The parser uses date anchors to group multi-line transaction records, removes banking boilerplate, and classifies each transaction using a multilingual merchant dictionary covering English, Finnish, German, French, Spanish, Italian, Dutch, Portuguese, Swedish, and Polish.
No data is sent to any external AI API in Standard mode. Text is processed entirely within the Numwisely infrastructure. This makes Standard mode the privacy-first default.
AI mode overview
AI mode extracts text using the same OCR pipeline, then sends the extracted text to OpenAI's GPT-4.1 mini model with a carefully engineered prompt. The AI model identifies transactions, normalizes merchant names, assigns categories, and returns structured JSON. Results are validated against the canonical category list before being saved.
AI mode requires consent because your financial document text is transmitted to OpenAI for processing. The consent notice shown before upload explains what data is shared and under what terms. Consent is recorded per upload session.
Comparison: when each mode works best
Standard mode works best on clean, structured documents: digital PDFs exported from an online bank, well-formatted CSV files, and printed statements with a clear date-amount-description column layout. It is fast, deterministic, and produces consistent results across repeated uploads of similar documents.
AI mode works better on: scanned handwritten receipts, invoices with unusual visual layouts, documents mixing multiple languages in the same transaction row, and cases where Standard mode leaves more than 20–30% of rows in the "unsure" fallback category.
- Use Standard: digital bank statement PDFs, CSV exports, printed receipts with clear text.
- Use Standard: when privacy is the primary concern.
- Use AI: when Standard mode output has many "unsure" or miscategorized rows.
- Use AI: for handwritten receipts, mixed-language documents, or unusual invoice layouts.
- Never use AI mode for documents you are not comfortable sharing with an external AI service.
Accuracy expectations by document type
For a clean digital bank statement PDF from a major Finnish, German, or UK bank, Standard mode typically achieves over 85% correct categorization on the first pass. Remaining rows fall to "unsure" and need quick manual fixes in Review.
For low-quality scanned images (blurry, tilted, poor lighting), both modes perform worse. The OCR extraction stage cannot recover text that is not legible. For best results, use 300 DPI or higher scans with good contrast.
Privacy implications of each mode
In Standard mode, your document text stays within Numwisely's processing infrastructure (Supabase Edge Functions and the OCR service). No content is sent to a third-party AI API.
In AI mode, the text extracted from your document — which includes merchant names, transaction amounts, dates, and descriptions — is transmitted to OpenAI's API for inference. OpenAI's data processing terms apply. Review those terms before using AI mode for sensitive financial documents.