The OCR extracted the wrong fields

When to use this article

The AI extracted fields from your document and one or more came out wrong: wrong vendor name, wrong amount, wrong VAT rate, wrong date, wrong currency. This article walks through the triage (what's most likely wrong, what to fix first), the per-field-edit flow, the document-level re-scan flow, and when to escalate.

For the broader story of what the AI extracts and how, see what the AI reads, and how to correct it. For the confidence-score-meaning context, see understanding confidence scores. For the per-vendor learning loop, see how to fix a misread vendor name. For pre-extraction problems (upload failed entirely), see upload failed — troubleshooting.

Quick triage — what's most likely wrong

Symptom	Most likely cause	First fix
One or two fields wrong	Confidence was too high for yellow flag; AI guessed wrong	Click and edit the specific fields
All fields wrong or garbled	Poor document quality (blurry, tilted, shadowed)	Re-upload a clearer version
Wrong vendor on a known sender	New vendor template the AI hasn't seen	Edit vendor; system learns for next time
Wrong VAT rate	Country-defaults misapplied, or rate on document is unusual	Edit rate; net / VAT amount / gross recompute
Wrong currency	Ambiguous symbol (e.g. $ could be USD/CAD/AUD/SGD)	Edit currency; conversion re-applies
Date wrong (year off by one)	Common with December invoices read in January (ambiguous date format)	Edit date; conversion-rate re-fetched
Multi-page PDF only extracted page 1	Some scanners produce malformed PDFs	Re-save the PDF in any viewer, re-upload
Line items wrong but totals right	AI saw totals clearly but struggled with table layout	Edit line items individually, or accept totals + leave items blank
Totals wrong but vendor right	OCR misread the totals (often a faded "0" or "8")	Edit totals; AI's confidence on totals is usually high so this is rare
Whole document classified as irrelevant	Classifier missed invoice signals	From Activity log, Mark as invoice + extract

The fix path depends on which of these you're hitting. Most are one-click corrections; whole-document issues are re-scan or re-upload.

Step-by-step — edit fields

The most common fix: just edit the wrong field.

1. Open the document detail page (from the Documents list, or from the in-app notification when extraction finished).
2. Click any field on the right side. A text input appears in place.
3. Type the correction. Autocomplete shows existing values where applicable (e.g. vendor names suggest from your Trading Partners).
4. Tab or click outside the field to save. The change is immediate; the document is marked with a corrected flag on that field.
5. The save propagates — the next invoice from the same vendor (matched by name + sender + layout + page hash) picks up your correction automatically.

See how to fix a misread vendor name for the vendor-learning loop in detail. The same propagation applies to other fields (VAT rate, category) — your correction trains future extractions for that vendor.

Cascading edits

Some edits cascade. Changing the VAT rate recomputes VAT amount and gross / net (or net depending on which you've also edited); changing the currency triggers a fresh ECB-rate lookup and recomputes the converted amount. The cascading is shown in the UI — the affected fields highlight briefly to indicate they've been re-derived.

Manual override of high-confidence fields

A field can be auto-applied with 95% confidence and still be wrong. Click and edit anyway — the system doesn't argue with you. Your manual edit replaces the AI's value; the audit log records both the original and the override.

Step-by-step — re-scan a document

If editing field-by-field is too tedious (e.g. nearly everything is wrong, or you've made vendor corrections elsewhere and want them retroactively applied), force a fresh extraction:

1. Open the document detail page.
2. Click Re-scan in the top action bar.
3. The original file is sent back through the AI pipeline with whatever vendor-corrections you've made since the first scan. The new extraction overwrites the old one — your prior manual edits to individual fields are preserved (re-scan respects manual overrides; it only re-extracts fields that weren't manually touched).
4. Up to 3 re-scans per document, no quota cost. The button greys out after the third.

When re-scan helps:

• You've corrected the vendor name on another document from the same sender; re-scanning now applies the corrected name.
• You've updated the country setting (Settings → Company → Country) which changed VAT-rate defaults; re-scan picks up the new defaults.
• The original was processed during a transient AI hiccup (rare, but happens) and you want a clean retry.

When re-scan doesn't help:

• Same document content + same model = same result. If you haven't made corrections between scans, re-scan returns the same fields.
• Document is genuinely unreadable. Re-scan can't see better than it did the first time.

After 3 re-scans the option greys out. At that point either: re-upload a cleaner version (different file hash = new document), or escalate to support.

Step-by-step — re-upload a cleaner version

If the original photo is blurry, tilted, shadowed, or otherwise low-quality:

1. Take a new photo with better conditions. Tips: flat surface, even lighting (avoid direct flash on glossy receipts), phone parallel to the document, fill the frame with the document, one document per photo.
2. Upload the new version via the standard upload flow (drag-and-drop or file picker).
3. Our duplicate-detector compares file hashes — the new version won't match the blurry one (different bytes), so it's a separate document. Quota counts the new one.
4. Delete the blurry one from the Documents list (checkbox → Delete). Quota isn't refunded but that's expected (deletes don't refund anyway).

You end up with one clean document instead of two. The vendor-learning from the blurry document's corrections (if any) still applies — the system associates corrections with vendors, not with specific document IDs.

When to write to support

If after edit + re-scan + re-upload it's still wrong, and the document is a normal-looking invoice from a normal vendor: write to [email protected] with subject [BUG] OCR extraction wrong. Include:

• The document itself (attach the PDF / image; redact specific sensitive numbers if you want, but the layout and format matter for diagnosis).
• What field is wrong and what the correct value should be.
• What the AI returned (you can copy from the document detail page).
• The document ID (visible in the URL).

We use these reports to improve the model — your specific case can help fix the same pattern for others. Typically 1–3 working days for response.

Common causes per field

Wrong vendor name

• New vendor template the AI hasn't seen → edit once, vendor memory kicks in.
• Vendor's invoice is consolidated (e.g. "Stripe Inc" on the header, "Stripe Payments Europe" elsewhere) → the AI picked the wrong one. Edit to your preferred canonical name.
• OCR mistake on similar letters (e.g. "Hetzner" read as "Helzner") → edit.

Wrong VAT rate

• Country defaults misapplied (German 19% suggested when the invoice is actually French 20%) → edit.
• Mixed-rate invoice with multiple line items at different rates → use the line-items view to set per-line rates, or accept the dominant rate and adjust manually.
• Reverse-charge invoice (intra-EU B2B) where the rate is 0% with the reverse-charge clause → edit rate to 0%, optionally mark as reverse-charge in the document.

Wrong currency

• Ambiguous symbol like $ could be USD, CAD, AUD, NZD, SGD, HKD → the AI uses address and vendor context to disambiguate. If it picked wrong, edit.
• Multiple currencies on one invoice (line items in one currency, totals in another) → edit to the dominant currency.
• Cryptocurrency or non-ECB currency → conversion field is left blank; supply manually. See multi-currency and live ECB rates.

Wrong date

• Year off by one — common with December invoices uploaded in January. EU date format 01.12.2025 is ambiguous if year is partly hidden. Edit.
• Format misread — 05/04/2026 could be May 4 or April 5 depending on locale. Edit.
• Date too far in the past or future — confidence is usually flagged yellow for outliers. Confirm or correct.

Wrong totals

• Faded number in the PDF (printer ran out of toner) → OCR misread. Edit.
• Currency-symbol confusion ("€1,234.50" misread as "1234,50") → edit.
• Comma vs decimal in European notation (1.234,50 vs 1,234.50) → the AI handles both but rare edge cases trip it. Edit.

Edge cases

My re-scan didn't help. Same document + same model = same result. The improvement on re-scan comes from corrections you made between scans (vendor learning). If you didn't correct anything since the first scan, re-scan returns the same fields. Either correct first, then re-scan, or re-upload a cleaner version.

I want to re-extract with a different OCR model. We use one model (Claude via Anthropic). The choice of model isn't user-configurable. If you have a strong-signal use case where a different model would help (specific document type that consistently fails), write [TECHNICAL] to support — we may have alternative paths in development.

Multi-page PDF only shows page 1. Some scanners produce PDFs with JBIG2 compression or other quirks that confuse the multi-page reader. Open the PDF in any viewer (Preview / Acrobat / Skim), choose File → Export → PDF (or Save As → PDF), upload the re-saved version. The re-save normalises the format.

Document classified as irrelevant. From the email Activity log, click Mark as invoice + extract. This forces extraction and trains the classifier for that sender. See what happens when you forward a newsletter.

I correctly typed the right value but the field reverted. A re-scan or a workflow event (vendor merge, currency change) overwrote your manual edit. Manual edits are normally sticky (re-scan respects them), but if the field was last edited before a major migration, it might have been re-extracted. Check the audit log; if it's a real bug, write to support.

Line items extract but totals are blank. Some invoices print totals as images (rendered text rather than selectable PDF text) while line items are in selectable text. The AI handles both but image-rendered totals score lower. Edit totals manually; line items can stay as-extracted.

Hand-written annotations on a printed invoice. The AI ignores handwriting and reads only printed text. If the handwritten amendment changes the meaning (e.g. cross-out + corrected total), treat the printed value as authoritative and use the Notes field on the document to record the amendment for your own reference.

The document is correct but the bank-match scored it against the wrong invoice. Different problem — see how the matching pipeline works. Extraction-vs-matching are separate stages with separate scoring.

Related

• What the AI reads, and how to correct it — broader correction context
• Understanding confidence scores — score-meaning context
• How to fix a misread vendor name — vendor-specific learning
• Upload failed — troubleshooting — when extraction never started