Invoice OCR accuracy: where automation stops
Realistic OCR error rates on crumpled and faded restaurant invoices, line-item pitfalls, and safe human-review fallbacks.
OCR errors are uneven. A crisp computer-generated invoice may identify the supplier, date, total, and most descriptions reliably, while a phone photograph of a folded thermal print introduces glare, skew, shadows, and missing corners. Faded ink often turns 8 into 3; crumpled paper can shift a decimal; a narrow tax column can be joined to the net price. Even a line that passes a confidence threshold can fail arithmetic validation when quantity × unit price does not equal the extension. Treat extraction as triage: it should make review faster while preserving a human decision at the boundary of financial risk.
Where real invoices break the model
- Compare line quantity, unit price, extension, and pack description; one correct field does not validate the other three.
- Flag €12.50 versus €1250, 6 × €4.20 versus 6 cases, and a credit shown as 48 rather than −€48.
- Require review for low-confidence lines and for every line above a chosen materiality level such as €250.
- Use arithmetic, duplicate-line, date, and supplier-history checks as evidence, not as permission to hide the original.
Error rates by document condition
These ranges are illustrative operating expectations, not a benchmark or verified customer result. Accuracy depends on model, layout, language, and validation rules.
| Digital PDF | 97–99% fields | Text layer and stable layout |
|---|---|---|
| Phone photograph | 90–96% fields | Glare, perspective, and cropping |
| Faded or crumpled scan | Below 85% line details | Decimals, quantities, and signs are fragile |
| Safe fallback | Human review | Retain image, confidence, and correction |
Do not turn an illustrative accuracy range into a promise. Calibrate on your own supplier layouts and route uncertain or expensive lines to a reviewer.
A fail-safe review path
- 1Preserve the source. Store the original PDF or image with a stable identifier. Do not compress away the evidence needed to read a disputed line.
- 2Extract and validate. Capture fields, confidence, arithmetic checks, pack units, tax treatment, and credit signs. Mark missing values instead of guessing.
- 3Route by risk. Send low-confidence, unusual, duplicate, negative, or high-value lines to a person; a €250 threshold is an example, not a universal rule.
- 4Record the decision. Keep the corrected value, reviewer, timestamp, and reason beside the original extraction, then sample the next invoices for drift.
FAQ
- Is 99% OCR accuracy realistic?
- It can be realistic for selected fields on clean digital PDFs, but it is not a guarantee for every line. Faded or crumpled scans can perform below 85% on line details.
- Which invoice fields are most fragile?
- Decimals, quantities, pack sizes, tax columns, line extensions, and negative credits are common failure points. Validate them separately.
- When should a human review a line?
- When confidence is low, the line is financially material or unusual, arithmetic fails, or the source quality makes a plausible error likely.
- Can I delete the original after correction?
- No. Retain the original image or PDF and the reviewer decision so the corrected ledger value remains explainable.
Keep reading
- A 15-minute invoice review that catches what audits miss
Five checks, in order, that surface pricing errors before they reach your books.
- Catch invoice line-price drift before it eats your margin
Compare each invoice line against expected unit prices and act on real variances, not noise.
- Invoice matching issues and how to fix each one
The six ways a three-way match breaks — duplicate invoices, phantom vendors, price drift, quantity gaps, substitution fraud, and statement mismatches.
Browse by category
Want to review your purchasing?
Submit the details and we will confirm which invoices are useful.
Get in touch