Convert Invoice PDFs to CSV

The output is structured rows, not the text of the page. Header fields — vendor, invoice number, date, the totals — are separated from the line items, and each line item is its own row with its own description, quantity, unit price and line total where the invoice prints them.

Invoice layouts vary more than bank statements do, because there is no convention behind them. A software subscription invoice, a freight bill and a contractor invoice put the same information in entirely different places, and some of them omit fields others treat as mandatory. StackExtract reads the document in front of it, which is why the review screen matters as much as the extraction.

Each line item becomes its own row, with the invoice-level fields repeated alongside it. That shape is deliberate: it means a CSV of many invoices can be concatenated, sorted and filtered without losing which invoice a line belonged to, which a nested format would not allow.

Where the invoice gives StackExtract two figures that should agree, it checks them: quantity times unit price against the line total, line totals plus tax against the invoice total.

These are checks against the document itself, and that is their whole nature. They can tell you the invoice does not add up, which usually means a figure was misread. They cannot tell you whether the invoice is correct, whether the vendor is who it says it is, or whether the amount is what was agreed. Nothing on the page can answer those.

Where an invoice gives nothing to cross-check — no line-item breakdown, or a total with no components — no math check runs, and the review screen says so rather than showing a pass. An absent check and a passed check are different facts.

Nothing is recalculated to make a check succeed. Where an invoice does not add up, the figures are shown as printed and the row is flagged, which leaves the question where it belongs: with the invoice, not with the extraction.

CSV is the output when the invoice data is going somewhere else — a spreadsheet you already maintain, a template, a script, an import into a system that accepts transaction rows.

What it is not is a universal import file. Systems that ingest invoice data each want their own columns in their own order, and some want a format that is not CSV at all. StackExtract gives you clean structured data; the mapping into a particular system is a step you own.

The file is UTF-8 with the accents intact, so vendor names and currency symbols arrive as printed rather than as replacement characters. Anything that will be matched on a name downstream depends on that surviving.

Scanned and photographed invoices go through the OCR beta path, and they need more review than digital PDFs — recognition can produce a plausible wrong character where reading would have produced nothing.

Where a document matters enough that you want a person to check it rather than checking it yourself, Assisted Review is available as a paid add-on on the document it applies to. A reviewer looks at the flagged fields against the source and reports what they found.