What fields come back from an invoice?

Vendor and customer names and addresses, invoice number, invoice and due dates, PO reference, currency, subtotal, discount, tax, total, and every line item with description, quantity, unit price and amount. Nulls mean "the document does not say" — never a guess.

What does it cost at volume?

Plans include monthly AI credits — Starter $19 with 250, Professional $59 with 1,000, Business $149 with 4,000 — and one-time credit packs top up without changing plan. One credit per document up to 25 pages.

Can it read invoices in Arabic, Sinhala or Tamil?

The extraction engine is multilingual and we measure these scripts specifically — see the Arabic, Sinhala and Tamil pages for the honest per-language figures, including the ones that are not perfect.

How do we trial it against our own invoices?

Create a free account, buy a $15 credit pack (which also unlocks API keys, no subscription), and run your real paperwork through it. Your own documents are the only benchmark that matters for you.

Last updated 2026-08-03

Solutions · Accounts payable

Invoice OCR and data extraction

Supplier invoices become structured fields — vendor, dates, totals, every line item — with the arithmetic checked before you see it.

Try it on an invoice Plans and credits

Someone on your team retypes supplier invoices into your accounting system. It is slow, it is dull, and once in a while a 7 becomes a 1 and the wrong amount gets paid. Generic OCR does not fix this: it gives you a wall of text, and the typing becomes copy-pasting.

Invoice extraction returns fields instead — vendor, invoice number, dates, PO reference, currency, subtotal, tax, total and every line item — as JSON your system can import directly.

Measured, not promised

Every figure below is a dated measurement recorded in our benchmark log, run through the same pipeline customers use.

Measured 2026-08-03

Field accuracy: 76 of 76 fields exact across a three-layout, three-currency benchmark corpus — and 76 of 76 again on degraded, phone-photo-grade copies of the same documents. On real paperwork: a 7-page purchase order with a one-time discount read exactly, and a 15-line bilingual supplier quotation came back cell-perfect.
The arithmetic is checked before you see it: Deterministic code — no AI — verifies that line items sum to the printed totals, that quantity × price holds on every line, and that dates run in order. A document that does not add up is flagged in plain words, and re-read once on a stronger model at no extra cost.
It says what it is unsure of: The model names any field it could not read confidently — in practice mostly numeric dates where day and month could be swapped — instead of presenting every guess with equal confidence. Your reviewer checks three fields, not thirty.
Honest about what a document is: Send it a purchase order by mistake and it says "purchase_order" rather than pretending to be an invoice. Five document types are supported: invoices, receipts, purchase orders, delivery notes and goods received notes — the set three-way matching needs.
Cost and speed: One credit per document — $0.040–0.059 depending on plan — answered in 2–5 seconds, up to 25 pages per document. No per-page surprises.
Where your documents live: Processed on our own servers in Frankfurt, Germany. Originals are deleted within an hour of processing, results within 30 days, and nothing is ever used to train models — ours or anyone else’s. The security page and DPA spell out the rest.

Method and raw results: how we measure. Every figure is dated and reproducible.

How you use it

In the browser: Drop an invoice on the extraction workbench, get the fields with their checks, copy or download the JSON. Good for trying it on your own paperwork today.
Over the API: One POST with the file, JSON back in the response — no job polling for single documents. Keys are created in your dashboard and every sample in the docs was captured from a live call.
Into your system: The JSON is stable and schema-bound per document type, so mapping it into your accounting or ERP import is a one-time exercise.

What to know before you commit

The limits, stated here rather than discovered in your evaluation.

KNOWN LIMIT

Handwritten invoices are not the printed-invoice product. Our published figures are for printed documents. Handwritten material behaves differently — we measured a handwritten page at 43.9% on classical OCR and 95.3% on our AI engine, so it is possible, but treat it as its own conversation.

KNOWN LIMIT

A verified result is still worth a glance where money moves. The checks catch inconsistent documents and misread totals, and they cannot catch everything. Our terms say it plainly: extracted output is data, not advice, and a person should confirm before payment.

Questions buyers ask

What fields come back from an invoice?: Vendor and customer names and addresses, invoice number, invoice and due dates, PO reference, currency, subtotal, discount, tax, total, and every line item with description, quantity, unit price and amount. Nulls mean "the document does not say" — never a guess.
What does it cost at volume?: Plans include monthly AI credits — Starter $19 with 250, Professional $59 with 1,000, Business $149 with 4,000 — and one-time credit packs top up without changing plan. One credit per document up to 25 pages.
Can it read invoices in Arabic, Sinhala or Tamil?: The extraction engine is multilingual and we measure these scripts specifically — see the Arabic, Sinhala and Tamil pages for the honest per-language figures, including the ones that are not perfect.
How do we trial it against our own invoices?: Create a free account, buy a $15 credit pack (which also unlocks API keys, no subscription), and run your real paperwork through it. Your own documents are the only benchmark that matters for you.

Try it on an invoice Plans and credits

Prefer a conversation first? Write to the people who built it