Accounts Payable
Arabic OCR for Invoices: Why Generic OCR Fails on Bilingual Documents (and What Works)
Most OCR tools treat Arabic as an afterthought, and bilingual invoices with stamps show the gap. This guide covers what fails and what works.

Khushbu Adav
Product at Predflow

Most OCR tools treat Arabic as an afterthought, and bilingual invoices with stamps show the gap. This guide covers what fails and what works.

Arabic OCR is the process of turning Arabic text in a scanned or photographed document into machine-readable text. On a clean printed page, most tools manage. On a UAE tax invoice with Arabic and English side by side, a company stamp across the total and a handwritten note in the margin, most tools fail in ways your AP team only notices after the data reaches the ERP.
This guide explains why Arabic is harder for OCR, where bilingual UAE documents break, how classic OCR differs from AI document extraction, and how to test a tool on your own invoices before you buy.
Why Does Generic OCR Fail on Arabic Invoices?
Why does generic OCR fail on Arabic invoices? Generic OCR engines learned mostly on printed Latin text. Arabic letters join and change shape, text runs right to left while numbers run left to right, and UAE invoices mix Eastern Arabic and Western digits. On bilingual invoices with stamps, these features cause misread totals and TRNs.
English OCR engines learned on billions of pages of printed Latin text. Arabic has less training data and a script with features Latin text does not have.
Connected Script
Arabic letters join to their neighbours, and most letters change shape depending on position: start, middle, end or standalone. An OCR engine has to split a word into letters before reading them. Splitting errors turn one supplier name into three nonsense fragments.
Right-to-Left Text With Left-to-Right Numbers
Arabic reads right to left. Numbers inside Arabic text read left to right. A line such as an invoice total mixes both directions. Engines built for one direction often reverse the order of figures or attach a number to the wrong label.
Eastern Arabic and Western Digits
UAE documents use two digit systems. Western digits (0123456789) and Eastern Arabic digits (٠١٢٣٤٥٦٧٨٩) both appear, sometimes on the same page. An engine without Eastern Arabic digit support reads ٥٠٠٠ as noise or as letters, and the invoice total disappears.
Dots and Diacritics
Many Arabic letters differ only by the number or position of dots. Low-resolution scans and phone photos lose the dots first. One missing dot changes a letter, and the word no longer matches your vendor master.
Where Does Arabic OCR Break on Bilingual UAE Documents?
Most UAE business documents carry both languages. Bilingual pages create problems beyond Arabic script alone.
Supplier names appear in Arabic on the invoice and in English in your ERP. A transliterated name such as "Al Noor" appears as "Al-Noor", "Alnoor" or "Al Nour" depending on who typed the record.
Gulf spellings and abbreviations differ from Modern Standard Arabic training data.
Labels sit in one language and values in the other. "المجموع" (total) in Arabic, then "AED 12,450.00" in Western digits.
Two-column layouts put Arabic on the right and English on the left. Engines reading line by line merge the two columns into one garbled line.
Teams often add a chatbot or a translation step to deal with bilingual text. The better fix sits earlier: extract fields correctly from the source document, then match names with rules built for transliteration.
Where Generic OCR Fails on Real UAE Documents
Document | Typical problem | Result in your ERP |
|---|---|---|
Bilingual tax invoice | Two columns merged, stamp over the total, Eastern Arabic digits in the VAT line | Wrong total or VAT amount, missing TRN |
Delivery note | Handwritten quantities, receiver signature, company stamp | Wrong received quantity, failed GRN match |
Customs declaration | Dense tables, mixed scripts, codes in small print | Missing HS codes and duty values |
Trade licence | Arabic-first layout, licence number near a logo | Wrong licence number or expiry date on supplier onboarding |
Supplier statement | Long tables across pages, Arabic headers | Lines dropped between pages, balance mismatch |
The cost shows up downstream. A misread total fails your 3-way match. A misread TRN fails e-invoicing validation. A misread supplier name creates a duplicate vendor. Your team then spends the time saved by OCR on fixing OCR output. See how one AP team removed this rework in the SAP invoice-to-GRN automation case.

Infographic in text: Arabic OCR vs AI Extraction on Bilingual Invoices. Reading characters is step one. Your AP team needs correct fields.
Extract fields: Arabic and English, Eastern Arabic and Western digits.
Validate: TRN format and status, VAT at 5%, line totals.
Match: Vendor master, PO and goods received note.
Review exceptions: A person sees failed checks, with the reason.
Post to ERP: Clean invoices post with an audit trail.
What generic OCR returns: A block of text with mixed text direction, missing dots, digits read as letters and stamps read as words.
What field extraction returns: Supplier, TRN, invoice number, date, line items, VAT and total as structured data, checked before posting.
What Is the Difference Between OCR and AI Document Extraction?
Classic OCR reads characters. The output is a block of text. Your team or a rules engine then has to find the invoice number, the TRN, the line items and the total inside the text block.
AI document extraction reads fields. The model looks at the whole page, understands which value belongs to which label, and returns structured data: supplier, TRN, invoice date, line items, VAT amount, total. Modern extraction models handle Arabic and English in the same pass and treat a stamp as a stamp rather than as text.
The difference matters most on bilingual documents. Character accuracy of 98% sounds high. On a page with 1,500 characters, 98% means 30 errors, and one error in the total makes the whole invoice wrong. Field accuracy is the measure your AP team cares about.
Arabic OCR Options Compared
Approach | Arabic accuracy on bilingual invoices | Handles Eastern Arabic digits? | Tables and line items | Cost per page | Data residency |
|---|---|---|---|---|---|
Classic OCR engines (open source) | Low to medium on clean print, weak on scans and stamps | Depends on the engine and language pack | Weak, returns text only | Lowest, self-hosted | Your servers |
Cloud document AI (major cloud providers) | Medium to high on print, varies by model and region | Yes on most current models | Good on standard layouts | Low per page, rises with volume | Provider cloud, check region availability |
LLM-based extraction | High on mixed layouts, stamps and handwriting notes | Yes | Strong, reads tables in context | Higher per page | Depends on deployment, private hosting possible |
AI agent with extraction and validation | High, with checks on every field | Yes | Strong, with totals checked | Priced per workflow | Your cloud or a managed cloud |
Check every vendor claim on your own documents. A demo on a clean sample tells you little about your supplier invoices.
How Do You Test Arabic OCR on Your Own Documents?
Build a test set from your own files before you choose a tool.
Collect 200 documents from the last six months: 120 supplier invoices, 40 delivery notes, 20 credit notes and 20 statements or customs papers.
Include the hard ones on purpose: stamps over totals, phone photos, faxed copies, handwritten notes, Eastern Arabic digits.
Define the fields you need: supplier name, TRN, invoice number, date, line items, VAT amount, total.
Key the correct values by hand once. This becomes your answer sheet.
Run each tool on the 200 documents and score field accuracy, not character accuracy.
Record the share of documents with all required fields correct. This number tells you how many invoices go straight through.
A tool scoring 95% field accuracy but 70% full-document accuracy still sends three in ten invoices to a person. Measure both.
A Workflow Built Around Extraction
Extraction alone does not finish the job. A working process for UAE AP teams looks like this:
Extraction: read every field from the invoice, Arabic and English, digits in both systems.
Validation: check the TRN format and status, recalculate VAT at 5%, confirm line totals add up to the invoice total, compare the supplier with the vendor master.
Matching: compare the invoice with the purchase order and goods received note.
Human review on exceptions only: a person sees the documents with a failed check, along with the reason.
ERP posting: clean invoices post to the ERP with an audit trail.
Predflow builds agents to run this whole flow on your ERP and tools, with a person approving exceptions. For a wider view of extraction tools, read our invoice OCR software comparison.
Under e-invoicing, structured invoices from in-scope suppliers arrive as data, not images. Extraction still matters for suppliers outside the first phase, for delivery notes, statements and customs papers, and for every document before 2027. Read the UAE e-invoicing readiness plan for the dates.
Questions to Ask Any Arabic OCR Vendor
Question | Why the answer matters |
|---|---|
What is your field accuracy on bilingual UAE invoices, measured on documents like ours? | Character accuracy hides field errors |
Do you read Eastern Arabic digits and convert them correctly? | Totals and VAT lines often use them |
How do you handle stamps, signatures and handwriting over printed text? | Common on UAE invoices and delivery notes |
Where is our data processed and stored? | Some teams require processing inside their own cloud |
What happens when a field confidence is low? | You need a review step, not a silent guess |
How do you match transliterated supplier names to our vendor master? | Arabic and English spellings rarely agree |
How does the tool learn new supplier layouts? | New suppliers arrive every month |
Ask for a trial on your 200-document test set before you sign. A vendor confident in Arabic accuracy will agree.
Keeping Accuracy High After Go-Live
Extraction accuracy drifts over time, as new suppliers arrive with new layouts and existing suppliers change their templates. Plan for drift from the start.
Review a sample of 50 processed invoices every month and compare extracted fields with the source document. Track field accuracy by supplier, since one supplier with a difficult layout often causes most of the errors. Feed every correction back into the extraction setup, so the same error does not repeat.
Watch the exception queue as well. A sudden rise in exceptions from one supplier usually means a layout change. A slow rise across all suppliers usually means a scanning or image quality problem at your end.
How Does Predflow Help With Arabic Invoice Extraction?
Predflow is an AI agent development company. Predflow builds and runs AI agents for businesses in the UAE, India and the US, on the ERP and tools each business already uses, to save time and cost on repetitive work. Predflow builds agents which extract fields from Arabic and English invoices, check the TRN, VAT and totals, and send only failed checks to a person. In one SAP deployment, 94% of supplier invoices went straight through to posting without manual keying. Read the SAP case study.
Conclusion
Arabic OCR fails on bilingual UAE documents for clear reasons, from connected script and mixed text direction to two digit systems and two-column layouts. Classic OCR returns text and leaves your team to find the fields. AI document extraction returns fields, and validation catches the errors before they reach the ERP.
Test any tool on 200 of your own documents. Score field accuracy, then count how many invoices pass with every field correct. The result tells you the real workload left for your team.
Sources
Bring 20 Arabic and English invoices to a 30-minute extraction test
We will run your actual bilingual invoices through extraction and VAT checks on the call, and show you field-level accuracy on your own documents. No prep beyond the PDFs.
Frequently Asked Questions
Can OCR read handwritten Arabic?
Classic OCR engines read handwritten Arabic poorly. AI extraction models perform better on short handwritten notes, such as a received quantity or a signature date on a delivery note. Long handwritten text in Arabic still needs human review. Test handwriting accuracy on your own documents before relying on any tool.
Does Arabic OCR handle Eastern Arabic digits?
Most current cloud and AI extraction models read Eastern Arabic digits (٠ to ٩) and convert them to Western digits. Older and open-source engines depend on the language pack and often misread them. Include documents with Eastern Arabic digits in any test set, since invoice totals and VAT lines often use them.
Is Arabic OCR accurate enough for accounts payable?
Bilingual UAE invoices need field extraction with validation, since OCR on its own leaves too many errors. When every extracted invoice goes through TRN, VAT and total checks, and a person reviews only the failed checks, AP teams post most invoices without manual keying and catch errors before payment.
What is the difference between OCR and document extraction?
OCR converts an image of text into plain text. Document extraction identifies the fields on the page, such as supplier, TRN, date, line items and total, and returns them as structured data ready for your ERP. Extraction uses OCR inside the process but adds layout and context understanding.
Will e-invoicing remove the need for Arabic OCR?
E-invoicing reduces the need, without removing the need. E-invoicing replaces PDF invoices with structured data for in-scope businesses from 2027. Delivery notes, supplier statements, customs papers, trade licences and invoices from out-of-scope suppliers still arrive as documents. Teams still need extraction for those.
What accuracy should we expect from Arabic invoice extraction?
Expect high field accuracy on clean printed invoices and lower accuracy on scans with stamps or handwriting. The useful measure is the share of invoices where every required field is correct. Test on your own documents, since supplier layouts and scan quality decide the real result.


