Meet the documents where they arrive
Web upload, email webhooks, WhatsApp, mobile capture, and no-login portal links for suppliers. The channel matters, because a system that requires vendors to change their behaviour does not get used.
W-Enterprises/AI document extraction
Document AI / Arabic and English
The re-keying is the cost. That is what automates.
Invoices, forms and statements arrive as PDFs, photos and email attachments, and someone types them into a system by hand. AI extraction removes that step: OCR plus large language model normalisation, in Arabic and English, with a human review workflow instead of blind trust. Shipped and running in production for the Saudi market.
93.5%
Field extraction accuracy, MVP scope
~26
Header fields plus line items per document
AR / EN
Mixed-script documents handled natively
0
Documents persisted to disk by default
How it actually works
The naive version of this is "send the PDF to a model and hope". The production version has four stages, and each exists because something fails without it.
01 / Ingest
02 / Extract
Hybrid OCR plus LLM normalisation
OCR gives positional ground truth, the language model gives interpretation: dates in mixed formats, totals that must reconcile against line items, vendor names that mix Arabic and Latin script, VAT numbers validated against their checksum shape.
03 / Review
A workflow, not a leap of faith
Low-confidence fields are flagged for a human, not silently guessed. Draft, approved, pushed. The economics still work because a person confirming three flagged fields is twenty times faster than a person typing thirty.
04 / Deliver
Into the system of record
Export to Excel and PDF, or straight into an ERP or accounting platform through an integration layer. Extraction that ends in a folder is not automation. See also ZATCA Phase 2 integration where compliance is the driver.
Where this pays for itself
- Finance operations re-keying supplier invoices across formats and two languages. This is the shipped, proven case.
- Accounting and bookkeeping practices processing client source documents at volume, where extraction time is margin under fixed fees.
- HR and admin teams lifting data from contracts, IDs and forms into structured records, with the document itself never retained.
- Logistics and field operations digitising delivery notes and proofs of delivery photographed on a phone.
The data protection posture
Documents stream through and are not written to disk by default, which keeps the PDPL, POPIA and GDPR surface small by construction rather than by mitigation. Where retention is genuinely required, it is deliberate and configurable, and audit trails record metadata only, never field values.
Common questions
How accurate is it, honestly?
About 93.5 percent field accuracy on the shipped platform's MVP scope. No extraction system is 100 percent, which is exactly why the review workflow exists: low-confidence fields go to a human, and unverified data never lands in accounting records on its own.
Does it handle Arabic documents?
Yes, natively: right-to-left text, Arabic-Indic and Latin numerals, and vendor names mixing scripts. This is where generic tools degrade, and it is the reason the work exists as a specialisation.
What does an engagement look like?
A short discovery against your real documents, a fixed written scope, then typically a few weeks to a working system on your own infrastructure. Accuracy is measured on your document set, not quoted from a brochure.
Which models and stack are used?
Gemini as the primary extraction model with Cloud Vision as optional hybrid OCR and OpenAI as an alternate provider, behind a Node and PostgreSQL backend and a bilingual React interface. The stack flexes to your cloud and your constraints, and you own the system at the end.
Drowning in manual capture?
Send a sample of the documents and the volume per month. You will get a written read on feasibility, expected accuracy and scope within one working day.