Reading Arabic invoices with AI: what works, what still needs a human

A finance manager in Kuwait showed us her month-end routine once. Three hundred supplier invoices, arriving as PDFs, phone photographs, forwarded emails and the occasional fax that had been scanned by someone who did not look at the result. Two people, four days, typing numbers into an accounting system. Every month, the same four days.

That is the problem document intelligence is supposed to solve, and in the English-language demos it looks solved. Arabic invoices are a different matter — not because the technology fails, but because it fails in specific, predictable places that no demo shows you.

Here is what genuinely works now, where it still breaks, and how to build something you can actually run without checking every field by hand.

What has genuinely stopped being a problem

Reading printed Arabic text is no longer the hard part. A clean, printed invoice — supplier name, invoice number, date, line items, total — comes out accurately, including the bilingual documents that are normal in the Gulf, where the header is English, the description is Arabic, and the terms at the bottom are both.

Layout variation has also stopped being the obstacle it was. The old approach needed a template per supplier: this field lives at these coordinates on this form. Add a supplier, build a template. Change a form, rebuild it. Modern extraction works from a description of what you want rather than where it sits, which means a new supplier is not a new project.

That is a real shift. It moves the work from configuring templates to defining what “correct” means for your business — which is a better use of anyone’s time.

Where it still goes wrong

Arabic-Indic and Western digits. Some invoices write amounts as ٢٥٠, others as 250, and a fair number mix both in the same document — Arabic-Indic in the header, Western in the totals table. Conversion looks reliable until you meet a document where a single digit is silently transposed. The failure is quiet, which is what makes it dangerous.

Dates. ٠٣/٠٤/٢٠٢٦ is either the third of April or the fourth of March. No model can resolve that from the document alone, because the information is not in the document. It needs a rule tied to the supplier, or a convention you enforce, or a human. There is no fourth option.

Tax. VAT rates and whether VAT applies at all vary across the Gulf. A system that assumes one rate produces totals that look entirely plausible and are wrong — the worst possible failure mode, because plausible wrong numbers pass review.

Poor scans. A photograph of a folded invoice taken at an angle in bad light remains the hardest case. Arabic script is connected, so mild blur does not degrade a character gracefully — it converts one letter into another. The output is not obviously broken; it is subtly incorrect.

The split that makes it safe

The design decision that turns this from a risk into an operating system is simple, and it is not a technical one.

Every field that becomes money leaving your business goes past a human. Totals, bank details, IBANs, due dates. No exceptions, no confidence threshold high enough to skip it.

Everything else — supplier name, line-item descriptions, invoice numbers, reference codes — is accepted automatically with a confidence score recorded alongside it.

That split sounds conservative. In practice it is what makes the project viable, because review collapses from twenty fields to three. A reviewer scanning three numbers against a document image moves fast. A reviewer checking twenty fields is just doing data entry with extra steps, and will stop bothering by week three.

Confidence thresholds in practice

Scores are only useful if something acts on them. The rules we tend to land on:

• Financial fields always route to a human, whatever the score.
• Descriptive fields pass automatically above a high threshold.
• Any document failing two or more fields goes to review whole, rather than being patched field by field — because a document with two bad reads usually has a systematic problem, not two unlucky ones.
• Every automatic acceptance is logged with its score, so the thresholds can be tuned against evidence instead of instinct.

How we run the first month

We start with one or two suppliers and a single month of invoices. Not the archive. Not every supplier. One month, one or two formats.

Within two weeks the error patterns are visible, and they are almost never what anyone predicted. It is one supplier whose totals row sits below a signature block. It is a date format that changed in March. It is a scanner setting that halves the resolution on every second page. These are fixable once seen, and invisible until you run real documents through.

By the end of the month you have something better than a demo: a measured hours-saved figure per invoice, a count of how many needed human intervention, and a categorised list of the errors that recurred. That last list is what makes the system improve rather than plateau.

What it means for the finance team

The accountant’s role does not disappear. It changes from complete manual entry to focused review of what actually warrants attention.

The time difference is large. The error difference is larger, and less obvious: human attention is finite, and a person checking twenty fields on three hundred invoices is not attending carefully by invoice two hundred. Concentrating that attention on the three fields where an error costs money produces better accuracy than the manual process it replaced — not merely faster.

That is the argument worth making internally. Not “the AI does the work”, which invites resistance, but “the AI does the typing so the review gets sharper”.

When we tell clients not to do it

If invoice volume is low — a few dozen a month — the maintenance and review overhead is hard to justify against simply typing them.

If every invoice comes from one supplier in one fixed format, ordinary template extraction is cheaper, faster and more predictable than anything involving a language model. Using AI there is paying for flexibility you do not need.

If the documents are overwhelmingly bad phone photographs, fix the capture process first. Better input beats better models, and it costs less. A cheap document scanner and a rule about how invoices arrive will outperform any amount of clever extraction on a folded photograph.

We would rather say this at the start than three months in. It is also, bluntly, the fastest way to tell whether the person quoting you has done this before.

What to measure

Not raw accuracy — it is a comforting number that hides the important detail. (The same instinct applies to costing a WhatsApp assistant: the headline figure is rarely the one that matters.) Measure minutes saved per invoice, the share of documents needing intervention, and the categorised error list. The first tells you whether it was worth doing, the second whether it is trustworthy, and the third is the only one that makes next month better than this one.

Frequently asked questions

Does it work with handwritten invoices?

Handwritten Arabic remains unreliable enough that we do not recommend automating it. Handwritten amounts on an otherwise printed invoice can be flagged for human entry, which is the honest way to handle them rather than pretending accuracy that is not there.

Where does the data go? Is it sent outside Kuwait?

That depends entirely on how it is built, and it is a question you should ask before anything else. Processing can be arranged so documents are handled within a defined region, or so that sensitive fields never leave your systems at all. For financial and government documents this is usually a requirement, not a preference.

Can it push straight into our accounting system?

Yes, and that is where most of the time saving actually lives. Extraction alone still leaves someone copying values across. The integration into your existing system is the part that removes the manual step entirely.

How accurate is it, really?

On clean printed invoices, high enough that review is quick. On poor scans, low enough that review is essential. Any single accuracy percentage quoted without reference to document quality is marketing. The useful measure is how many documents needed a human, on your documents.

What happens when a supplier changes their invoice format?

Because extraction works from field descriptions rather than fixed coordinates, a format change usually passes through without intervention. If it does not, confidence scores drop on that supplier and the documents route to review — which is exactly the behaviour you want, rather than silent failure.