A receipt is a small object with an inflated opinion of its own importance. It fades, curls, hides in pockets and waits until tax time to become urgent. Ernest Semerda and Dmitry Birulia knew the species well. Before Veryfi, they had spent years managing expenses with desktop software, losing hours to data entry and chasing colleagues for the story behind each scrap of paper. So, around the end of 2016, the two engineers began with a modest proposition: let a machine do the typing.
- Veryfi converts documents into structured data and checks them for signs of fraud.
- Its wedge was an expense app; its larger business became APIs and capture SDKs.
- More than 1,000 organizations use it, from fintech products to consumer loyalty programs.
- The free tier suits experiments. Production starts at a $500 monthly minimum.
- Copy the sequencing: solve one narrow workflow, learn from it, then sell the reusable layer.
Veryfi entered Y Combinator's Winter 2017 batch as a personal and small-business expense product. Its founders talked to users constantly, connected the app to QuickBooks when customers asked, and aimed the software at the tasks people postponed: scanning, coding and explaining expenses. That application did real work. It also did something more valuable in private. Every uploaded receipt and invoice gave the company another difficult example on which to train.
For roughly three years, millions of documents accumulated. The cheerful phrase for this is a data flywheel. The less cheerful phrase is that someone had to make sense of every smeared total, multilingual item and merchant name printed by a machine whose ribbon had plainly lost the will to live. The expense app became a laboratory. The trained extraction system became the export.
“We started with a simple premise to automate data extraction.”Ernest Semerda and Dmitry Birulia
The app was the apprenticeship
This was not a public failure followed by a cinematic rescue. The first product worked; by one early account it had processed four million receipts and reached break-even. What changed was the founders' understanding of what they owned. An expense tracker served one workflow. A fast, pre-trained system for turning disorderly documents into clean JSON could serve hundreds.
Today, a developer can send Veryfi a receipt, invoice, hotel folio, bank statement, check, purchase order, tax form or identity document. The system captures the image, classifies it, extracts fields and line items, returns structured data and, when the relevant tools are enabled, looks for evidence of tampering. Veryfi Lens handles the camera side inside mobile apps, including awkward long receipts. A browser extension, dashboard and mobile software serve teams that do not want to start with an API call. Workflows adds no-code routing around the extraction engine.
Four jobs hiding inside one scan
That last step explains why the company sits in several markets without quite belonging to any one of them. For Navan, the document becomes an expense inside a travel product. For PepsiCo, it becomes proof of purchase in a loyalty campaign. For a builder, it becomes job-cost data. For a healthcare application, it may become structured information from an insurance card or superbill. Veryfi is not selling the final experience so much as the ability to make the paper disappear inside it.
The product is speed you do not notice
The strongest customer results are not really about OCR. Navan says its expense product reached up to 90 percent adoption; Veryfi is one component in that outcome, not the sole cause. The useful detail is why Navan bought rather than built. A home-grown extraction system would require engineers, data specialists, maintenance and time. Navan wanted its people improving travel and expense, while a specialist worried about receipt geometry. The invisible infrastructure helped the visible product feel immediate.
PepsiCo supplies a still sharper contrast: purchase validation that once took 11 days could happen in seconds. In a loyalty program, that delay is the difference between delight and a consumer wondering whether the promotion was a clerical prank. In another case, Travelit cleared a backlog of more than 70,000 invoices in a week. The pattern is consistent: the value appears when extraction removes a queue, not when it merely produces an impressive demo.
Veryfi competes with developer-first services such as Mindee, AI document platforms such as Nanonets and Rossum, older OCR specialists such as ABBYY, fraud-oriented processors such as Ocrolus, and the raw document services offered by Amazon, Google and Microsoft. Its argument is specialization. The models arrive trained for financial documents; line items are a first-class concern; mobile capture is part of the system; and fraud signals sit beside extraction rather than several integrations away.
That positioning is especially useful for product teams that want an embedded component, not a new kingdom of software. It is less decisive when a company needs a complete accounts-payable suite with elaborate exception queues and ERP approvals. General cloud OCR can also be cheaper at enormous scale if the buyer has engineers ready to build and maintain the missing layers. Veryfi removes plumbing; it does not repeal architecture.
The price of avoiding the plumbing
The pricing makes the intended customer obvious. Developers can test up to 100 documents a month for free. The Starter platform then carries a $500 monthly minimum. Published unit rates include eight cents per receipt and sixteen cents per invoice, with volume pricing above the standard tier. A separate expense product is priced per active user.
The unit economics are pleasant only when the minimum is busy. At 6,250 receipts, $500 works out neatly to eight cents each. At 400 production receipts, the same floor becomes $1.25 each. That is not trickery; the minimum is plainly published. It simply means a hobby project, a tiny office or an uncertain prototype should remain on the free tier, choose a lower-floor service or postpone production. Veryfi works best when document volume is already a problem worth paying to make boring.
What a founder can steal
There is a reusable lesson here, and it is more demanding than “add AI.” Veryfi earned its models through a product that solved a genuine problem. The narrow application created examples, edge cases, integrations and credibility. Only then did the company expose the reusable machinery beneath it.
Choose a wedge that manufactures learning. The expense app did not merely acquire users; it produced the documents and feedback needed to improve extraction.
Listen for requests that reveal a platform. QuickBooks integration requests showed that customers wanted data to move, not another isolated scanner.
Sell the relieved outcome. Seconds instead of 11 days is legible. “Advanced multimodal architecture” is material for a procurement appendix.
Know when the arithmetic rebels. A transaction model with a monthly floor needs steady volume. Below it, even excellent automation can be an expensive ornament.
The strategy would be hard to copy where documents are scarce, privacy rules prevent model improvement, the first application cannot attract meaningful use, or a general model already solves the task well enough. It also depends on patience. Three years of expense software is a long apprenticeship if one begins by insisting the infrastructure layer must be visible on day one.
Veryfi now describes its software with a brisk promise: extract data, catch fraud, one API. The second clause matters. Reading a receipt is no longer the entire job when a convincing fake receipt can be generated as easily as a birthday invitation. The company has added checks for manipulation, duplicates, suspicious metadata and AI-generated images. The machine that learned to read the paperwork is being asked to develop judgment about it.
The valuable trick was not digitizing a receipt. It was turning the receipt into a decision while nobody was waiting.
There is something pleasingly unfashionable about the result. Veryfi did not begin with a universal intelligence and search for a nuisance. It began with a nuisance, studied it until the nuisance became expertise, and packaged that expertise for everyone else. The paper was merely where the lesson was printed.