Document Types: From Clean Digital to Camera Scans
Business documents typically fall into four distinct difficulty categories:
Category 1 (Clean Digital PDFs): Modern software invoices (Stripe, AWS, QuickBooks, Google Workspace) exported directly to PDF.
Category 2 (High-Resolution Scans): Clean flatbed scans (300 DPI) of printed paper bills.
Category 3 (Camera Photos): Smartphone photos of paper receipts with slight perspective tilt and lighting shadows.
Category 4 (Degraded Documents): Thermal receipts with faded ink, skewed scans, and multi-page tables without gridlines.
| Document Category | Source Examples | Expected Reliability | What to Expect | Recommendation |
|---|---|---|---|---|
| Digital Vector PDFs | Stripe, AWS, QuickBooks, Google direct exports | Very High | Clean text and distinct table rows | Ready for direct spreadsheet use |
| High-Resolution Scans (300 DPI) | Clean flatbed scanner printouts | High | Accurate totals; verify multi-column tables | Compare grand totals against original |
| Smartphone Camera Photos | Paper receipts photographed on desk | Moderate | Captures main totals; line items may split | Always verify line items and currency |
| Degraded / Skewed / Thermal | Faded thermal register tape or angled scans | Low to Fair | Numbers can misread (e.g. 3 vs 8, missed decimals) | Manual review essential before ledger import |
What Works Exceptionally Well
In practical document workflows, modern multimodal AI excels in three key areas compared to legacy OCR:
- Contextual Header Association: Even when an invoice places the "Due Date" in a top-right decorative card and the "Amount Due" in a bottom banner, AI correctly correlates the labels with their respective values.
- Borderless Table Alignment: Traditional OCR relies on visible grid lines. AI understands whitespace geometry, keeping description, quantity, and unit price aligned in correct Excel columns.
- Currency and Symbol Normalization: Mixed symbols ($1,250.00, €450,00, £89.90) are converted into clean numeric columns without string contamination.
Where AI Struggles (And Why)
Honest engineering means acknowledging failure modes. Here are the common challenges to watch out for:
- Faded Thermal Ink: On old store receipts where digits were physically faded, AI can occasionally confuse "3" with "8" or drop cents decimals (e.g., reading 14.50 as 1450).
- Extreme Rotations (>30 degrees): Files uploaded upside down or tilted at sharp angles experience significant drops in extraction quality.
- Multi-Page Table Column Drift: When a table spans across multiple pages and column widths change slightly, AI may combine two narrow columns into one cell.
Golden rule for accounting data
Never import extracted spreadsheets directly into your general ledger without running a simple sanity check: Subtotal + Tax = Grand Total.
How to Get the Most Accurate Results
Following these three practical guidelines will give you the cleanest, most reliable data:
1. Always prefer digital PDF exports over printed paper whenever available.
2. If photographing paper receipts, lay the document flat in bright, even lighting with no harsh shadows.
3. Specify the exact columns you want in Excel before uploading to avoid misplaced cells.