Invoice OCR accuracy: where automation stops — Guide pratique pour les restaurants
Une méthode concrète, fondée sur les factures, les chiffres et des contrôles simples.
A restaurant decision is only as reliable as its source and assumptions. Clean PDFs may reach 97–99% field accuracy, while faded or crumpled scans can fall below 85% on line details. Write the baseline, show the formula, and separate an observed result from an illustrative estimate. A good guide gives the manager a practical review path, a threshold for action, and a date to check whether the correction worked.
What the data can and cannot tell you
- Treat field accuracy and line accuracy separately; a correct vendor name does not validate a price.
- Expect decimal shifts, merged lines, tax-column confusion, and pack-size loss on poor images.
- Use confidence thresholds plus materiality thresholds, for example review every line over €250.
- Never overwrite the original image when a reviewer corrects extracted data.
A worked example
A restaurant compares a dated source with an expected result and finds a gap. These numbers show the method only; they are not a verified customer case.
| Clean PDF | 97–99% fields | Consistent supplier layout |
|---|---|---|
| Phone photo | 90–96% fields | Glare and skew need review |
| Faded/crumpled scan | Below 85% lines | Decimals and quantities are fragile |
| Fallback | Human review | Keep image, confidence, and correction |
Use the calculation only when the source, unit, and exception log are visible. Treat it as illustrative until your own records confirm it.
A repeatable process
- 1Define the question. Write the decision, source, unit, time window, and threshold before opening the spreadsheet.
- 2Reconcile inputs. Match identifiers, quantities, dates, credits, and exceptions. Flag missing evidence instead of guessing.
- 3Calculate the gap. Show formula, baseline, observed result, and illustrative impact. Keep price separate from waste or process variance.
- 4Review and act. Have a manager review material exceptions, agree one action, and set a date to measure it.
FAQ
- What is the most common mistake?
- Comparing numbers that look similar but use different units, dates, specifications, or commercial terms.
- How precise should the result be?
- Precise enough to support a decision, not more precise than the source. Preserve the underlying invoice, image, or count.
- When should a person review it?
- Whenever confidence is low, the amount is material, the item is unusual, or an exception changes the conclusion.
- Are the savings figures guaranteed?
- No. All figures in this guide are illustrative. Validate the formula and opportunity against your own records.
Pour aller plus loin
- Une revue de factures de 15 minutes qui attrape ce que les audits ratent
Cinq vérifications, dans l'ordre, qui repèrent les erreurs de prix avant la comptabilité.
- Détectez la dérive des prix ligne par ligne avant qu'elle ne mange votre marge
Comparez chaque ligne de facture aux prix unitaires attendus et agissez sur les écarts réels, pas sur le bruit.
- Problèmes de rapprochement de factures et comment les résoudre
Les six façons dont le rapprochement à trois volets casse : doublons, fournisseurs fantômes, dérive des prix, quantités manquantes, substitutions et écarts de relevé.
Explorer par catégorie
Vous voulez examiner vos achats ?
Envoyez-nous les détails et nous confirmerons quelles factures sont utiles.
Nous contacter