Skip to main content

Understand AI extraction

Clerked reads an invoice source document and prepares header and line data for a person to review.

Supported source files

The invoice upload flow accepts:

  • PDF files.
  • Common image formats, including PNG, JPEG, TIFF, WebP, GIF, BMP, HEIC, and HEIF.
  • Spreadsheet files in CSV, XLSX, XLS, or XLSM format.
  • DOCX files.

What extraction fills

Extraction follows your company schema. It can fill configured header details, totals, dates, references, and invoice lines, but it does not force every company to use the same field names.

For example, one company may configure fields for an invoice number, vendor, date, total, and purchase order reference. Its line sections may include description, quantity, price, and coding fields. Another company can use a different set and order.

What happens next

  1. Clerked receives the source file from email or upload.
  2. Processing reads the document and prepares the configured invoice data.
  3. The invoice becomes available for review when processing finishes.
  4. You compare the extracted fields and lines with the source document.
  5. You correct any value that does not match the source.
  6. Autosave stores the correction and validation checks the current invoice again.

Extraction does not replace review. The source document remains the reference when an extracted value and the invoice disagree.

Review line data

Invoice lines appear in stacked, schema-defined sections. Use Review view to check every authored field, or use Table view for a smaller set of configured columns. See Review invoice line items.

When a field has a sparkle rationale control, open it to read Clark's reasoning for that value. The rationale can help explain the choice, but the source document is still the reference.

If something goes wrong

  • If processing reaches a terminal error, open the processing error notice instead of creating a duplicate invoice.
  • If a field is missing, check whether your company schema includes it before treating that as an extraction error.
  • If a line is in an unexpected section, review its classification fields and use an allowed Move to {segment} action when the line needs to move.

We use cookies to understand how visitors use our site.