Can AI read Arabic invoices?

The layout runs right to left, so the label and its value sit the opposite way round from an English invoice. A tool that was built assuming left to right and then translated will pair the wrong number with the wrong field, which is a much worse failure than not reading it at all, because it looks fine.

The numerals may be Eastern Arabic. ٠١٢٣٤٥٦٧٨٩ are the same values as 0123456789 and a lot of software has never met them. Many Lebanese and Gulf invoices mix both on the same page, printed figures in one set and the totals in the other.

One document commonly carries two languages. An Arabic header, an English supplier name because that is how the company registers, line items in whichever the person typing knew best. This is normal here, not an edge case, and any tool that has to be told which language a page is in will struggle with it.

And dates are worse than in English. 04/08/2026 is already ambiguous anywhere. Add a supplier who may be following a different convention entirely and a tool that silently resolves it has filed a guess into your accounts.

The model is not told what language to expect and is not asked to translate. It reports what is printed, in the script it is printed in, and the value separately. That keeps the audit trail intact: you can always get back to what the page actually said.

Numerals are normalised to values rather than kept as text, so ١٩٥٠ and 1950 both become a number you can sum.

Dates that cannot be resolved from the page are flagged rather than guessed, and the printed form is kept alongside whatever we resolved. On a Lebanese invoice that flag fires often, which is why it is written as information rather than as a defect: the characters were read perfectly, only the convention is open.

And the totals are checked with arithmetic. That check is language independent, which makes it the most useful thing we do on a document in a script we are less sure about.

Take five of your own documents, and pick badly on purpose. One clean printed Arabic invoice. One phone photograph taken in a hurry. One that mixes Arabic and English. One with Eastern Arabic numerals. One thermal receipt that has started to fade.

Run them through whatever you are evaluating and check three things. Did the total come out right. Did the supplier and the date come out right. And on the two that were always going to be difficult, did it tell you it was unsure, or did it hand you a confident number?

That last one is the whole test. Our reader runs in the browser on the page linked below, and we would rather you tried it on your worst documents than took our word for it.

We have not published a benchmark on Arabic documents, and you should be suspicious of anyone who has published one on yours. Accuracy figures depend almost entirely on the pile they were measured against, and a vendor picks their pile.

What we will say is what happens on a document we cannot read: it comes back marked unreadable with a reason, not with a number in it.