AI Document Processing: Extract Data from Arabic and English Invoices, Receipts and IDs

Somewhere in your company, someone opens a supplier invoice and types its number, date, tax number and totals into a spreadsheet or the accounting system. Then the next one. At thirty a month that's fine. At five hundred it's a job, and the typos surface at month-end when the numbers don't reconcile.

AI document processing does the reading and typing, checks the result, and asks a person only when it's unsure. Below: how it works, why Arabic is harder, what accuracy to expect, when you don't need AI at all, and the cost.

What it can read, and where the data goes

How it works, in five steps

1. Capture. Documents arrive from a shared email inbox, an upload page, a scanner folder, a phone photo, or WhatsApp.
2. Classify. The system decides what each document is and, for invoices, which supplier sent it.
3. Extract. An AI model reads the text and layout and returns the fields with a confidence level for each.
4. Validate. Plain rules check the data (next section). Most of the reliability comes from here.
5. Review and export. Confident documents that pass every check go straight through. The rest open in a review screen with the original page beside the fields, so a person fixes them in seconds. Clean data goes to Excel or CSV, your accounting software's API, or your ERP.

Validation rules: the part that makes it trustworthy

The AI reads; ordinary code checks. These rules are predictable and catch most extraction mistakes before they reach your books:

• Arithmetic. Line totals add up to the subtotal, and subtotal plus VAT equals the total, within a small rounding tolerance.
• VAT. The rate matches what applies (for example 15% standard VAT in Saudi Arabia, 5% in the UAE, 14% in Egypt), and the supplier's tax number has the right format and matches your supplier records.
• Duplicates. The same supplier, invoice number and amount already entered is flagged, which stops the same invoice being paid twice.
• Matching. Invoice quantities and prices are compared with the purchase order and the delivery note.
• Plausibility. No future dates, no expired IDs, no totals far above what that supplier usually sends.
• Risky changes. A supplier's bank details that differ from the last invoice always go to a person, because that's a common fraud pattern.

Your accountant decides the rules; we turn them into code.

Why Arabic documents are harder

These issues only show up on real samples, so we test on yours:

• Handwriting. Printed Arabic is read well by current models. Handwritten Arabic, especially hurried, connected script, is far less reliable, so handwritten fields usually go to review.
• Mixed languages. Arabic invoices with English item names, or English contracts with Arabic names and addresses, mix text directions on the same line.
• Two digit systems. Arabic-Indic digits (٠١٢٣٤) and Western digits (01234) can appear on the same page, along with different decimal separators. Everything is normalised to one format before the checks run.
• Stamps and poor scans. Stamps often sit on top of totals or dates; thermal receipts fade.
• Name variations. أحمد and احمد, or ة and ه, must still match the same supplier or employee in your records.
• Dates. Hijri, Gregorian, or both.

Accuracy, honestly

No system reads every document correctly; anyone promising 100% hasn't tested on your documents. During the proof of concept we measure:

• Field accuracy, per field, on documents the system hasn't seen before.
• Straight-through rate: the share of documents that need no human touch.
• Errors that get past the checks, which is the number that really matters.

System-generated PDF invoices usually extract very well; crumpled receipts and handwritten forms do worse. The aim is that correct data flows automatically, uncertain data reaches a person, and nothing wrong slips through silently. You choose the line between speed and caution.

Already-structured e-invoices may not need AI at all

In Saudi Arabia, e-invoices under ZATCA's integration phase are XML files in the UBL 2.1 format. In Egypt, e-invoices are submitted to the Egyptian Tax Authority as signed JSON or XML. If your suppliers send you the structured file, or your e-invoicing provider gives you access to the invoices you receive, the data can be read directly: exactly, cheaply, with no AI and no guessing. That's an integration job, and usually a smaller one.

AI earns its place on the documents that aren't structured: PDFs and scans, receipts from small shops, foreign suppliers, delivery notes, IDs, contracts and forms.

Ready-made tools or a custom build?

Many accounting apps include receipt scanning, and the major cloud providers sell document AI services. If your documents are standard, mostly English invoices and your accounting app already offers this, try it first.

A custom pipeline makes sense when documents are mostly Arabic or mixed, when you need your own validation and matching rules, when the documents are specific to your business, when data must land in your own ERP, or when per-page fees grow expensive at your volume. Often the answer is in between: a cloud model does the reading, and we build the checks, review screen and integration around it.

Privacy: IDs, contracts and personal data

• Extract only what you need, mask ID numbers where the full number isn't required, and delete images on a schedule your policies allow.
• AI and cloud accounts in your name, using providers and settings that don't train on your data.
• Access by role, with a log of who viewed or changed what.
• Hosting region chosen to fit your rules. Saudi Arabia, the UAE and Egypt all have personal data protection laws; your legal adviser confirms what applies, and we build to it.

How we start: a proof of concept on 50–100 real documents

1. You send 50–100 real documents of one type, bad scans included, and we agree the fields, rules and destination.
2. In about 5 days you get a working pipeline, a review screen and a field-by-field results sheet.
3. We review the results together: what it got right, what it missed, and why.
4. You decide: production, a changed scope, or stop, having spent little.

The production version adds the real inbox or upload connection, the review workflow for your team, export to Excel, accounting software or your ERP, and a monitoring view.

Prices and timelines

That's roughly SAR 1,900 / AED 1,850 for a proof of concept and SAR 5,600 / AED 5,500 for production (approximate; we price in USD). AI usage is paid by you directly to the AI provider at its own price. Payment is in milestones (typically 30% / 40% / 30%) after a written scope and fixed price, you see a working preview within 72 hours, and you own the code and every account.

What we can show you, honestly

We haven't published a document-processing project yet. The live AI you can try today is the "Q" assistant on this website, which answers in Arabic and English from our own knowledge base. Our own products PharmaRiwaq and Dawawin HR show how we design multi-module systems and assistant experiences (their public demos run on fictional data). And there's On Time Express, a courier operations system we built for a client in Egypt, where every money movement goes into an audited ledger checked on every load. That's also why we start with a low-cost proof of concept on your documents.

When we're not the right fit

• You handle a few dozen documents a month. Typing them is cheaper than any system.
• All your invoices are already structured e-invoices. You need an integration, not AI.
• You need regulated identity verification, with liveness checks and certified providers. Use a specialist service; we can connect it to your app.
• The work isn't Sharia-compliant, such as processing applications for interest-based loans. We don't take it.

If you also want staff to ask questions of your contracts and policies, see our internal knowledge assistant. If documents arrive on WhatsApp, our WhatsApp AI agent can receive them. For AI features inside your existing system, see AI inside apps and ERPs.

FAQ

Can it read handwritten Arabic?

Sometimes, but not reliably enough to trust without checking. Printed Arabic works well; handwriting varies too much. We test on your real handwritten samples during the proof of concept, and handwritten fields are usually routed to human review.

Can the data go straight into our accounting software or ERP?

Usually yes, if the software has an API or a structured import. We check what your plan allows before quoting. Excel or CSV export always works as a fallback.

How much does each document cost to process?

The AI provider charges per use, so the cost per page depends on the model and how much is on the page. We measure it on your real documents during the proof of concept and give you a monthly estimate for your volume.

How do we start?

Collect 50–100 real documents of the type that takes most of your team's time and tell us where the data should end up. Send it on WhatsApp or use the project planner, and you'll get a written scope and fixed price within 24 hours. If you're comparing this with a customer-facing AI chatbot, we can scope both.