A Vendor Bill Only Becomes Payables Data After a Human Confirms It

Filings aren’t the only paper a finance team lives with. Vendor bills, purchase orders, receipts (the everyday flow of accounts payable) arrive just as often and get trusted far less carefully, usually because nobody has time to check them line by line. Rexfin applies the same instinct it applies to a filed statement, every value needs to point back to where it came from, to this pile of documents too. It just doesn’t pretend the result carries the same weight.

The same discipline, a different document

A bill, PO, or receipt goes through the same upload pipeline as a filing, but it’s extracted generically rather than matched against a per-issuer profile: there’s no filing structure to key off, just a vendor’s own invoice layout. The record it produces is shaped in four levels: party, document header, line items, tax lines, mirroring the fields a standard invoice actually carries: vendor name and tax registration number, invoice number and dates, the line-by-line quantities and prices, and the VAT breakdown by rate. Bank statements are handled separately and aren’t covered by this extraction at all: transaction rows aren’t line items, and a statement usually arrives as a structured file that wants a deterministic parser, not field-by-field reading.

Every value carries its source, never just a number

Nothing on an extracted bill is a bare value. Every leaf field, a vendor name, a line total, a tax rate, carries the value itself plus the page and bounding box it was read from, a confidence score, and which extractor produced it. That confidence score is advisory, not a trust grade: a field read at 0.99 confidence is still only as trustworthy as the document’s overall review status says it is, the same discipline behind the multi-backend extraction agreement that filings go through.

The extraction itself runs in order of preference: first a match against a known template for that vendor, which is deterministic and doesn’t touch a model at all; failing that, a schema-driven pass with a vision-capable model; then a separate anchoring pass that locates each extracted field on the rendered page so it has a real bounding box, not just a guessed value. A field nobody could anchor to the page comes through flagged with low confidence rather than a confident-looking placeholder.

Confirmed by a person, not by a score

Extraction produces a draft, not a fact. Every extracted bill sits in a quarantined extracted state (visible, drillable back to the source page, but not usable as data) until a person reviews it and confirms it or rejects it as a whole document. Confirmation is deliberately per-document, not per-field: someone looks at the extracted fields next to the source page, corrects anything wrong, and signs off once. A rejection requires a reason, so a discarded read still leaves a trail instead of just vanishing. Before either happens, the totals get a sanity check (net plus tax should equal the gross, and the lines should foot to the total), surfaced as a flag for the reviewer to see, not a silent block.

A corrected reissue (a vendor’s amended invoice or a credit note) never overwrites the original record. It arrives as its own upload and its own extraction, linked back through a supersedes chain, so the newest confirmed version wins without erasing what came before it. That’s the same non-destructive posture behind reconciliation against printed totals: correct forward, never mutate in place.

Why it never becomes a citation

Even fully confirmed, an extracted bill has a hard ceiling: it’s unaudited internal data, badged as such everywhere it appears, and it structurally cannot carry a citation the way a figure from a filing can: there’s no field for it. That’s not a gap to be closed later; it’s the honest classification of what a single confirmed read of one invoice actually is, distinct from the multi-backend agreement bar that governs what counts as citable in a filing. A confirmed bill is real, useful, reviewable payables data (spend by vendor, spend by category), it’s just never asked to pretend it’s something a filing’s own numbers had to earn independently.

See how extraction, confirmation, and citation fit together across the rest of the platform at the product tour hub.