Yes, AI can read many printed Arabic invoices, but whether it reads yours is something you test, not assume. Run it on your own scans, score printed and handwritten pages separately, check totals, vat and line items field by field, and keep a person approving every exception.
#Key takeaways
- A test on your own scans tells you more than any vendor accuracy figure.
- Printed and handwritten pages need separate scores, because one blended number hides the weaker half.
- The fields that matter most are the total, the vat amount, the tax number and every line item.
- Structure causes more trouble than characters, so check that each amount stays attached to its label.
- A person should approve every exception, and the system should say when it is unsure.
#Can AI read Arabic invoices?
An invoice is mostly numerals and tables, which is where reading errors do the most damage. So the check that matters is whether each amount stays attached to its label, not whether the characters look right. Stamps, folds, phone photos and mixed fonts all change the result. Stennir quotes no benchmark figure for Arabic invoices, because none has been verified on your documents: the only number that counts is the one you measure on your own scans.
#How do you test it on your own invoices?
Pull a sample from your own archive, across your real suppliers, scan quality and layouts. Include the ugly ones. Illustrative example, the numbers are assumptions: 200 invoices, of which 150 are printed and 50 are handwritten or stamped. The exact count matters less than covering every kind of invoice you actually receive.
Score at field level, not page level. For each invoice, mark every field right or wrong against what a clerk would have entered. A page that is mostly right can still carry a wrong total. Score printed and handwritten separately, so one blended number cannot hide the weaker half.
#Which invoice fields should you check first?
Start from the published rules and documentation, then test. In Saudi Arabia, zatca sets the e-invoicing requirements on its e-invoicing (fatoora) pages, so confirm there which fields a compliant invoice must carry. For the reading tool itself, vendors such as Microsoft publish language support lists for document intelligence. These lists change, so read the current version and treat it as a starting point rather than proof for your suppliers.
On Arabic invoices, a common risk is structure rather than characters. A vat amount can come back detached from its label, and a table can come back with its columns reversed. Treat this as a prompt to test, not a proven rate.
- The grand total, and whether it equals the sum of the line items
- The vat amount, and whether it sits against the right label
- The supplier's tax registration number
- The invoice number and date
- Each line item: description, quantity, unit amount and line total
#Can it read invoices that mix Arabic and English?
If your suppliers in the UAE or Saudi Arabia send bilingual invoices, test those pages as their own group. Arabic runs right to left, while numbers, product codes and English lines run left to right. Check that a line read in the wrong direction does not swap a quantity with a unit amount.
Stennir supports Arabic and English fully, others on request. Reversed tables are exactly what a bilingual sample should surface, before any invoice reaches your ledger.
#Who approves what the AI gets wrong?
A person should approve every exception. Set the rules before launch. A total that does not match its line items, a missing vat amount, a low-confidence field, or a supplier not seen before goes to a named reviewer in finance. Everything else can pass with spot checks.
Stennir builds this extraction and approval step as part of AI implementation, and you pay for finished work. Data is hosted where your rules require, in the country, cloud or system that fits. This post is about method and makes no legal claim.
We are
Stennir designs the test on your own invoices, then builds the extraction and the approval step with your finance team.
We aren't
A vendor that hands you an accuracy percentage from its own demo files, with no review step behind it.
#What should a finance lead do next?
Start with the test above, on your own scans. The finance page shows how Stennir works with finance teams, and AI consulting covers designing the sample, the scoring and the approval rules. The glossary gives plain definitions. To check readiness first, take the free readiness assessment. Or book a 30-minute call.
#Common questions
Does it handle handwritten Arabic and stamps as well as printed text?
Assume not until you measure it. Handwriting and stamps vary far more than print, so score them as a separate group on your own scans. Stennir does not claim a result for them. If handwritten pages are a small share of your volume, route all of them to a person.
What happens when the AI is unsure about a total or vat line?
The field should be flagged, not guessed. The invoice then goes to a named reviewer with the scan beside the extracted value, and nothing posts until they approve. Set the rule up front: any mismatch between total and line items, or any missing vat amount, always stops for review.
Can it read invoices that mix Arabic and English?
In principle yes, but test those pages as their own group. Arabic runs right to left while numbers and codes run left to right, and tables can come back reversed. Stennir supports Arabic and English fully, others on request. Check quantity, unit amount and line total on every bilingual sample.
Who checks and signs off the extracted data?
Your finance team keeps sign-off. The AI proposes values, a named reviewer approves exceptions, and a finance lead owns the rules for what passes unchecked. Stennir designs that split with you in AI consulting, so accountability for the posted number never moves to software.
What do we need to hand over to test it on our own invoices?
A sample of real invoices from your main suppliers, including poor scans, handwritten and bilingual ones. Add the correct values your team already entered for them, and the list of fields you need. Stennir agrees with you where the data is hosted, where your rules require, before any scan moves.
