Accounts payable OCR reads the supplier documents landing in your inbox and turns them into named fields your accounting system can accept. Nobody types an invoice number again.
Key takeaways:
- Accounts payable OCR returns named fields, not a page of raw text. That one difference is the whole point.
- One invoice costs $12.88 to process at the average organization against $2.78 at best-in-class AP teams, and manual handling is most of the gap (Ardent Partners, State of ePayables 2024).
- AI OCR needs no templates, so a supplier redesigning their invoice stops being your problem.
- A review step before export beats any headline accuracy percentage.
- Parseur handles capture and extraction. Approvals and payments stay in the accounting software or ERP you already run, which is the point rather than a limitation.
- Pricing is per page, and a free tier of 20 pages a month lets you run your own ugliest invoices through before anyone signs anything.
- EU-hosted and GDPR compliant. SOC 2 Type II is in progress and not certified, which you should know before your procurement team asks.
What is accounts payable OCR?
Accounts payable OCR is software that reads supplier invoices, receipts and purchase orders and returns named data fields such as vendor, invoice number, date, total, tax and line items, rather than returning the raw text on the page.
That distinction does the work. Plain OCR hands you a transcript, and somebody still has to find the total in it. Accounts payable OCR hands you total: 1880.83 and a place to send it.
Why plain OCR kept losing to invoices
Optical character recognition has read printed pages for decades. Reading was never the problem. The problem is that an invoice is not a form.
Every supplier lays theirs out differently. One puts the invoice number top right. The next buries it in the footer. Template-based OCR handled that by having somebody draw boxes on a sample document to say where each field lives. It worked beautifully, right up until a supplier redesigned their invoice. Then somebody redrew the boxes.
Two hundred suppliers, two hundred templates. Somebody on the team owns that maintenance now, and it is a worse job than the one it was supposed to replace.
How AI OCR killed the template
AI OCR identifies fields by understanding what they mean rather than where they sit, so a document layout it has never seen before still yields the right values.
Ask for the invoice number and it hunts for the thing that behaves like an invoice number, the way you would if you glanced at the page. There is nothing to annotate first and nothing to maintain afterwards.
Two engines do the reading, depending on what arrives:
- Vision AI handles PDFs, scans, photographs and images, which covers most supplier documents in practice.
- Text AI handles emails and text documents, including the invoice details vendors put in the email body rather than attaching.
Neither asks you to build anything first.
No engine reads everything, though, and you should distrust any page that says otherwise. When one supplier's layout keeps coming back wrong, you pin a template to that single layout and leave the AI to handle everyone else. That fallback is not an admission of defeat. It is the difference between a tool that survives your awkward vendors and a tool that gets quietly abandoned six months in.
What accounts payable OCR extracts from an invoice
| Field group | Typical fields | Why AP needs it |
|---|---|---|
| Vendor | supplier name, address, tax ID, remit-to | Matching the invoice to the right vendor record |
| Document | invoice number, issue date, due date, PO reference | Duplicate detection and payment scheduling |
| Amounts | subtotal, tax, total, currency, discount terms | Ledger entry and early-payment discount capture |
| Line items | description, quantity, unit price, line total, GL hint | Three-way matching and cost allocation |
| Payment | bank details, payment terms, method | Payment run preparation |
Line items are where cheap tools fall over. Pulling a header total is easy. Pulling twelve rows of a table, every row intact, off a page somebody scanned at an angle, is the harder half, and it is the half a polished demo never shows you.
So test it that way. Take the five ugliest invoices in your inbox, the multi-page tables and the one with a handwritten note across the total, and run those first. Any tool worth buying gets judged on your worst documents, not on the clean sample PDF in its own screenshots.
What it costs to keep typing
Doing nothing has a price. It just never arrives as an invoice of its own.
Ardent Partners' State of ePayables research puts the fully loaded cost of processing a single invoice at $12.88 for the typical organization against $2.78 for best-in-class AP teams. The same research puts the exception rate at 9% for top performers and 22% for everyone else. An exception is never just a percentage on a slide. It is an email to a supplier, a week of aging, and somebody chasing an approval.
Cycle time follows the same shape: 3.1 days for best-in-class AP teams against a general population measured in weeks.
Parseur's own survey of 500 professionals found manual data entry costs US companies $28,500 per employee each year. Across a two-person AP desk that is close to the cost of a headcount nobody hired, spent on typing. That last figure is our own research rather than an independent study, so weigh it accordingly. The Ardent numbers above are not ours.
And what Parseur costs
You cannot defend a purchase with only one side of the arithmetic, so here is the other side.
Parseur is priced per page processed, not per seat and not per user. One page is one credit, so a one-page invoice costs one credit and a three-page invoice costs three. There is a free tier of 20 pages a month, which exists precisely so you can run your own worst suppliers through it before you send anyone a purchase order. Plans scale to a million pages a month, and past that you ask for a quote. The pricing page has a calculator that produces the before-and-after numbers a CFO will want to see.
Can you trust what the OCR pulled out?
Not blindly. And any vendor who answers that question with a single accuracy percentage is selling, not explaining.
Accuracy on a clean native PDF from a familiar supplier is not the same number as accuracy on a phone photo of a crumpled delivery docket. One figure cannot honestly describe both. What matters is what the software does when it is unsure.
Extraction without a review step does not remove errors from your ledger, it relocates them somewhere nobody is looking, which is worse than manual entry because at least a human keying an invoice occasionally notices something odd.
The optional manual validation step is what makes that workable. Switch it on and every extracted record pauses for a person before it exports, so a wrong total gets corrected at intake rather than found in a bank reconciliation six weeks later. Parseur does not assign confidence scores to individual fields, so you choose which mailboxes carry that check rather than a threshold choosing for you. Normalization and validation run on every field regardless, holding dates, numbers and choices to the schema you defined. Our AI invoice processing benchmarks go deeper on how the numbers behave across document types.
Do not take any of that on trust either. The free tier is enough to run your own worst invoices through and see which fields come back right before you have a single procurement conversation.
Security, and what is not certified yet
Supplier invoices carry bank details, so this stops being a finance decision and becomes an IT one. Here is the version you can forward without editing it.
Parseur is EU-hosted and GDPR compliant, and the standard terms are aligned to UK GDPR, California CCPA, Swiss FADP and Singapore PDPA. Document retention is a setting rather than a policy you inherit: choose anything from one day upward, per mailbox. Access is role based with audit trails. Higher plans can request the report from our external penetration test, and enterprise agreements cover a custom security questionnaire.
SOC 2 Type II is in progress and Parseur is not SOC 2 certified today. If your procurement process requires the certificate before signature, we will not clear that bar right now. If it accepts an audit in flight plus a penetration test report and a signed DPA, we will.
Longevity is the other thing worth checking on a vendor. Parseur has run in production since 2016. Emonkey, an accounting integrator in Scandinavia, has kept it inside a client-facing product for years and reports no downtime across that span, which you can read in their own words rather than take from us.
Setting it up without code or a kickoff meeting
Finance teams end up in six-month implementation projects because they bought a platform when the thing that hurt was the typing. Capture does not need a project.
Step 1: Create a mailbox
Sign up and create a Parseur mailbox. It gets its own email address.
Step 2: Forward your invoices to it
Point a forwarding rule at that address, drop files in, or send them through the API. Your suppliers never need to know anything changed.
Step 3: Name the fields you want
Tell Parseur you want vendor, invoice number, due date, total and line items. No boxes to draw, no sample documents to annotate, no training set to assemble.
Step 4: Pick a destination
Connect QuickBooks, Xero or Google Sheets directly, route through Zapier, Make or Power Automate, or post to your own endpoint with a webhook or the API when your ERP prefers to be fed on its own terms. See invoice data to ERP if that is where yours ends up.
An afternoon, not a quarter. Which deserves a definition, because every vendor who ever burned you also said something like it.
An afternoon buys you this: mailbox created, forwarding rule pointed at it, fields named, one destination connected, real invoices arriving as structured data. What it does not buy you is a finished project. Reconciling your field names with what your ERP expects takes a conversation with whoever owns that system. The supplier whose invoice is a photograph of a photograph will want a second pass. Both of those are days, not months, and both happen after data is already flowing rather than before anything works at all. That ordering is the whole difference between this and an implementation project.
Where accounts payable OCR stops
Parseur reads supplier documents and delivers structured fields. It does not route approvals, execute payments, or perform three-way matching against your purchase orders. Those belong in your accounting software or ERP, where your controls and audit trail already live.
Worth saying plainly, because AP suites from SAP, Oracle NetSuite, Tipalti and Medius bundle capture in with everything else. Already running one of those is not a reason to close this tab. It is the common case. Most finance teams reading this own a system that handles approvals, matching and payments perfectly well, and are stuck one step earlier, where a person opens a PDF and types what it says into that system. That step is the capture layer.
So the split is simple. No AP workflow at all and you want approvals, matching and payments in one place: buy a suite. Workflow already fine and the typing is the wound: buy capture and point it at what you own. The second is a far smaller purchase with a far shorter meeting attached.
The accounts payable automation overview covers how the capture layer fits the rest of your finance stack.
Related reading
- Traditional invoicing vs AI invoicing
- How to automate invoice extraction
- How to extract invoice data
- What is invoice data capture?
- Extract data from invoices with Python
Last updated on
