Skip to content

Illustrative: a regional distributor and logistics operator

Reading Arabic and English supplier invoices into the ERP

An illustrative scenario: a distributor receives supplier invoices and delivery notes in Arabic and English — PDFs, scans and phone photos — and keys them into its ERP by hand. This is how we would build document intelligence that reads, checks and posts them, with people reviewing whatever the system is unsure of.

Illustrative scenario

Illustrative scenario: a composite of common requirements, not a description of a specific client project. Figures are targets we would agree and measure against — not results achieved.

The challenge

In this scenario, the finance team receives several thousand supplier invoices and delivery notes a month, by e-mail, through supplier portals and on paper. Some are clean PDFs; many are scans or phone photos, in Arabic, in English or in both. Clerks read each one, find the purchase order, check quantities against the delivery note and key the lines into the ERP.

Errors surface late — at month-end, or when a supplier chases payment. Duplicates slip through, VAT details are keyed wrongly, and the team spends its time typing instead of resolving the exceptions that need judgement.

ZATCA’s e-invoicing (Fatoora) requirements mean many suppliers now issue invoices that carry a QR code and structured data. Not every document does, and delivery notes fall outside those requirements altogether. The process has to handle both.

Our approach

We would start with a two-week discovery: sample a few hundred real documents, measure how long each type takes today and where errors occur, and agree targets with the finance owner before building anything.

Structured data first, AI for the rest

Where an invoice carries e-invoice data — an embedded XML file or a QR code — the pipeline reads it directly. Everything else goes through OCR tuned for Arabic and English, then a language model that extracts supplier, VAT number, dates, totals and line items into a fixed schema.

Checks before anything is posted

  • Match each invoice to its purchase order and delivery note, line by line
  • Validate VAT numbers, totals and tax arithmetic
  • Flag likely duplicates by supplier, invoice number, amount and date
  • Send any low-confidence field or failed check to a review queue, with the source document beside it

What we would build

What we would build

  • 01

    Intake

    A connector watches the accounts-payable mailbox and the supplier upload folder, splits multi-document PDFs and classifies each page as an invoice, credit note or delivery note.

  • 02

    Arabic and English extraction

    OCR and a language model, hosted in the client’s environment, extract fields into a fixed schema. Each field carries a confidence score and a pointer to where it was found on the page.

  • 03

    Validation and matching

    Rules check VAT numbers, arithmetic and duplicates, then match against purchase orders and goods receipts read from the ERP.

  • 04

    Review queue

    Clerks see the document and the extracted fields side by side, correct what is wrong and approve. Corrections are logged and feed a weekly accuracy review.

  • 05

    ERP posting

    Approved invoices are posted through the ERP’s standard APIs — for example SAP S/4HANA or Oracle Fusion Cloud — never by screen scraping.

  • 06

    Monitoring

    A dashboard shows volumes, the share processed without re-keying, exceptions by reason and accuracy on sampled documents, so progress against the targets can be reported every week.

Target measures

Documents posted without re-keying
70% or more
Target agreed at kick-off, against the measured baseline
Accuracy on key fields
98% or more
Target, checked weekly on a reviewed sample
Receipt to posting
Within 1 working day
Target agreed at kick-off
Duplicate invoices reaching payment
Zero
Target agreed at kick-off

Expected outcome

The expected outcome is a finance team that reviews exceptions instead of typing documents. Most invoices would be read, matched and posted without re-keying; the rest would reach a person with the problem already identified.

These are targets, not results. They are agreed with the finance owner at kick-off against a measured baseline, and reported against weekly from the first month in production.

Have a process that should run itself?

Tell us about it. We will help you work out whether AI is the right answer, what it would take, and what it should return.

Newsletter

Practical AI, in your inbox.

Real use cases, Saudi regulations and what works in Arabic — once or twice a month. No spam; unsubscribe any time.

We’ll e-mail you to confirm. Unsubscribe any time — see our privacy policy.