Ramp's parsing powers spend management
Geoff Charles, VP of Product at Ramp, on Ramp's AI flywheel
The key point is that Ramp could justify expensive document extraction because each correctly understood receipt or contract unlocked much more revenue and retention than a standalone OCR vendor usually can. Scale, Veryfi, and Ocrolus sell parsing itself as the product, so cost per document matters directly. Ramp uses parsing as an input into a larger spend management system, where better classification improves card usage, bill pay, procurement, contract intelligence, and software savings.
-
Veryfi and Ocrolus show what Ramp was comparing against. Veryfi prices receipt extraction per document, while Ocrolus explicitly pairs AI with human-in-the-loop review for harder files. That model can reach high accuracy, but every edge case adds labor cost, which is painful when the extracted data itself is the thing being sold.
-
Ramp was building for a different unit economics loop. By May 2023 it was using GPT-4 for expense audits, receipt scanning, contract extraction, and price intelligence on top of billions in customer transaction data. In that setup, better OCR is not a standalone revenue line, it improves a broader product that customers already trust with card spend, invoices, and vendor workflows.
-
This is also why foundation models were so disruptive. Ramp described GPT-4 as getting close to specialist OCR vendors on many tasks at a fraction of the cost, and Scale itself had already been framed as a mechanical-turk style data engine. Once generic models got good enough on messy business documents, the advantage shifted from who had the cheapest labeling workforce to who owned the most useful workflow and proprietary finance data.
Going forward, document understanding becomes table stakes and the real prize moves higher up the stack. The winners are likely to be companies that can turn extracted receipt, invoice, and contract data into actions, approve an expense, flag a minibar charge, benchmark a SaaS contract, or route a transaction into the books, with less human touch each year.