OCR and invoice capture: what it solves and what it can't
OCR and invoice capture convert a supplier's invoice – a PDF, a scan, an emailed image – into structured data your AP system can use: supplier name, invoice number, dates, amounts, line items. Capture solves data entry. The work that actually makes accounts payable slow – matching against purchase orders, resolving price and quantity discrepancies, deciding who approves a charge nobody remembers ordering – begins after capture has done its job.
Capture is pattern recognition, and it stops at the page
Classic OCR turns pixels into characters. Modern capture adds layout models and machine learning that locate fields on documents the system has never seen before, which is genuinely useful when every supplier formats invoices differently. What comes out the other end is a faithful transcript of whatever the supplier wrote. If the supplier typed the wrong PO number, capture will extract the wrong PO number flawlessly. The tool has no opinion on whether the data is true; it can only report what the page says. Most AP problems are disagreements between documents, and a transcript of one document settles nothing.
Header fields are easy; line items decide the outcome
Invoice number, date, supplier name, total due: these sit in predictable places in standard formats, and machines read them reliably. Line items are where capture earns or loses its keep. One supplier merges part numbers into free-text descriptions, another splits a shipment across three pages, a third bills in cases when the purchase order was written in individual units. Line-level data feeds matching, price verification, and GL coding, so one garbled line can push an otherwise clean invoice into a manual queue. When you evaluate capture vendors, ask how they handle your ugliest supplier's line items, because that is where your invoices will actually fail.
Matching begins where capture ends
Once the data is extracted, something has to compare it against the purchase order and the receiving record – the work of three-way matching. Capture can tell you the invoice says 96 units at $14.10. It cannot tell you why the PO says 100 units at $13.75, or whether the difference is a short shipment, a negotiated price change that never reached the ERP, or a supplier error. Those discrepancies have causes upstream, in the reasons invoice exceptions happen, and resolving them means reading the change order history, checking the contract, and sometimes emailing the buyer. Perfect extraction of a mismatched invoice yields a perfectly documented mismatch.
Non-PO invoices give capture nothing to check against
A meaningful share of invoices arrive with no purchase order behind them – utilities, legal fees, subscriptions, one-off services. For these, capture performs its transcription and then the hard questions start: who ordered this, which cost center absorbs it, does a contract govern the rate, is this the third time the same firm has billed for the same matter. Non-PO invoices turn on judgment about your own organization, and reading the document more accurately answers none of it.
The layers above capture determine the return
In the four-layer AP automation stack, capture is the foundation, and foundations are worth getting right. Even so, teams that buy capture expecting shorter cycle times often find the queue has simply moved: invoices enter the system in seconds and then sit in exception status for the same two weeks, because the people who once keyed data are the same people who chase discrepancies. Capture pays off when the layers above it – matching, exception resolution, approval routing – can act on what was extracted. Digitized data that nobody reasons over is a well-organized backlog.
Capture vendors solved the part of AP that could be solved by reading documents better, and the industry has spent a decade waiting for someone to work the queue that better reading creates. Fragment picks up exactly there: its agents reason across the ERP, supplier portals, email, and spreadsheets to resolve the exceptions that extracted data surfaces – autonomously where you trust them, with human breakpoints where you want a person to look first. Browse the workflows Fragment covers or see one run in a demo.