Invoice data extraction

Where invoice extraction fits in accounts payable

AP clerks and controllers use it at the front of the payables process, in place of typing invoices into the ERP. Approval and payment stay where they are. The output also feeds three-way matching, the control that compares the order, the delivery record and the supplier's bill and blocks payment on a variance. That is why line-level quantity, unit price and PO reference matter as much as the total.

Keying and chasing exceptions is where the cost sits. Ardent Partners puts the average cost of processing an invoice at $9.90, against $2.67 for best-in-class teams, with an average exception rate of 19.9% against 11.8% (a Medius-sponsored survey, not our measurement). A fixed supplier list sending the same forms every month is the best case for docRefinery: each layout is learned once and remembered. See document extraction for accounts payable.

Invoice fields docRefinery extracts

All 62 definitions are below, counting the two repeating tables and their columns. Your organisation decides which are switched on, sets a threshold for each, and can add custom fields.

Example invoice JSON

This is a trimmed excerpt of a finished run, with invented values. bbox gives the value's position as fractions of the page, and each alias is the name set under Return as, shown here at its default. The full response is in Extract a document.

{
  "document_type": "invoice",
  "result": {
    "fields": [
      { "alias": "Invoice ID", "text": "INV-0042", "field_confidence": 0.98,
        "location": { "page": 1, "bbox": [0.66, 0.09, 0.81, 0.11] } },
      { "alias": "Supplier Tax ID", "text": "GB000000000", "field_confidence": 0.71,
        "location": { "page": 1, "bbox": [0.08, 0.17, 0.26, 0.19] } },
      { "alias": "Purchase Order", "text": null, "location": null },
      { "alias": "Line Item", "table": [[
        { "alias": "Description", "text": "Sample widget, blue", "field_confidence": 0.96,
          "location": { "page": 1, "bbox": [0.08, 0.42, 0.44, 0.44] } },
        { "alias": "Unit Price", "text": "12.50", "field_confidence": 0.97,
          "location": { "page": 1, "bbox": [0.62, 0.42, 0.70, 0.44] } }
      ]] }
    ],
    "extraction_approval": "human-approval",
    "low_confidence_fields": [
      { "alias": "Supplier Tax ID", "text": "GB000000000", "confidence": 0.71,
        "threshold": 0.85, "reason": "low_confidence" }
    ]
  }
}

What docRefinery brings to invoice extraction

docRefinery is a next-generation extraction system built for invoices that keep coming back. Eight things change on an AP desk.

• Your names on every value. Each field has a Return as alias, so Supplier Tax ID can arrive as vendor_vat if that is the column your ERP import expects.
• One format for the whole inbox. Invoices return the same field and table shape as purchase orders, receipts and statements, so one parser reads them all.
• Line items as rows. Quantity, unit price, tax and PO reference stay together per line, ready for matching.
• Suppliers remembered. Each layout is learned once, and a redesign is re-trained when it first arrives.
• A source behind the figures. Extracted values and table cells carry page and box citations, so an approver checks the invoice itself.
• Corrections that pay back. A clerk's fix becomes a suggested fix for that supplier's layout, tested on that layout's own invoices before anyone applies it.
• Supplier-level insight. Insights shows vendors, flagged fields and auto-approval trends for your account, so a problem supplier stands out.
• Overseas suppliers. 70+ languages and scripts are supported, so a foreign invoice doesn't need a second tool.

Data reaches your ERP through the API or an export today. Inbox pickup, webhooks and ERP connectors are in design on the roadmap.

What makes invoices hard to read

• Four parties, four addresses. Supplier, receiver, remit-to and ship-to all look alike, and the wrong one is still a plausible address.
• Line tables that break. Tables run across pages, lose their header row, or carry subtotals mid-table.
• Mixed tax. One invoice can carry several VAT rates, stated per line and again in a summary block.
• Rules differ by country. GOV.UK's list requires a unique number, the supply date and the amounts, not where they sit or what they are called.
• Ambiguous dates. 03/04 means different months in different countries.
• Marks on scans. Stamps and handwritten approvals often land on top of the figures.

How low-confidence invoice values are handled

Every value is checked against the threshold set for that field, so Supplier IBAN can demand far more than Supplier Website. Below the bar, the run is marked human-approval and the value waits in the review queue, where a clerk sees the box on the page and corrects it. A required field that can't be found is flagged as missing. You can also write a workflow rule such as "line amounts add up to the net amount, within a cent", and a failed rule is flagged the same way. Exports to JSON, CSV or XLSX carry the corrected values.