Table recognition on invoices: lines, columns, pitfalls

To a human, an invoice is a table. To OCR it is often an obstacle course: thin rules, nested headers, right-aligned amounts, footnotes under the net total. When line items break, total checks fail - and the bill of costs or value calculation inherits the error.

Table recognition in the firm is therefore not a nice-to-have; it is what makes invoices machine-usable without retyping every row.

What makes layouts hard

Scanned thermal receipts, coloured background panels, stamps over amounts, multi-column descriptions, combined net/gross columns and handwritten corrections added later. Each element can shift column boundaries. Poor scan angles make it worse.

Before recognition it often pays to: deskew, raise contrast, crop needless margins. Aggressive compression before OCR is the most common self-inflicted wound.

Professional control instead of blind trust

Even good recognition confuses “1” and “7”, thousand separators and line wraps in descriptions. Define mandatory spot checks: grand total against the paper, a random line item, tax rate, invoice number and date. When a dispute turns on particular lines, those fields are exactly what matter.

LexLogik can support table structures in processing; release into the matter and to the client stays human. Use exports only when the spot check has held.

Preparation that pays

• Scan originals digitally or at high resolution, not from a third-generation phone photo.
• One invoice per file where possible - bulk scans hinder column mapping.
• Watch locale of amount formats (1.234,56 vs 1,234.56).
• Remove or segregate handwritten marks before OCR if they cross rules.

When the table fails

Sometimes manual capture of critical rows is faster than a third engine setting. Record then that only totals and disputed lines were taken over. For bulk similar invoices, a fixed template and team training time pay off instead.

From voucher to matter without a media break

Once line items hold, naming and filing decide whether the figures stay findable. Link recognised amounts to the voucher file and matter reference, not only to a loose spreadsheet. When a dispute turns on particular lines, you need the same scan the OCR saw - not a later “pretty” remake without provenance.

For matter types with many similar invoices (construction or insurance, for example) a short internal template pays: which columns are mandatory, which spot check applies, from which amount a second review is required. Duller than a new engine - and it saves more correction rounds.

Good table work cuts typing errors and shortens the path to a reliable figures base - provided nobody trusts the first recognition blindly. The engine proposes; the firm releases. Without that last step even accurate recognition stays a draft with no matter value.

Partner release on disputed amounts

If the disputed invoice total exceeds an internal threshold, the matter lead checks the OCR table against the image before figures enter pleadings. Below that, a support sample suffices. Put the threshold on the same team sheet as the scan profile.

Table recognition often fails on skewed scans and aggressive smoothing, not on “bad AI”. A dedicated invoice profile and an amount threshold for partner release make OCR figures more defensible in costs work.