W-Enterprises/Work/AI invoice parsing platform
Case study / KSA · FinOps · 2026
An AI invoice platform that meets suppliers where they already are.
A Saudi finance operations venture needed supplier invoices, arriving as PDFs, photos and WhatsApp messages in Arabic and English, turned into structured, approved records without an army of data-entry staff. The result is a multi-tenant SaaS in production, extracting about 26 header fields plus line items at 93.5 percent accuracy.
Context
Finance teams in the Gulf receive supplier invoices in every format that exists: clean PDFs, scanned photocopies, phone photos taken at an angle, and increasingly, images sent over WhatsApp because that is how business is actually done in the region. Every one of those documents was being read by a person and re-typed into a system, in two languages.
Manual capture scales linearly with volume, produces inconsistent data, and hides errors until reconciliation finds them weeks later. The client wanted a product: multi-tenant, self-serve, and honest about accuracy rather than pretending automation is perfect.
What made it hard
- The documents are hostile. Mixed Arabic and English on one page, Arabic-Indic and Latin numerals in one table, stamps and handwriting over printed text, and totals that must reconcile against extracted line items.
- The suppliers cannot be onboarded. A system that requires a vendor to register, log in and upload correctly does not get used. Ingestion had to meet suppliers on channels they already use.
- Trust had to be earned per field. Pushing a wrong total into an accounting record is worse than no automation. Confidence had to be visible and low-confidence fields routed to a person.
- Multi-tenant from the start. Tenant isolation, roles and billing, because this is a product, not an internal tool.
Approach
Five channels, zero supplier friction
Web upload, QR-initiated mobile capture, inbound email webhooks, WhatsApp via Twilio, and no-login portal links a buyer can send to any supplier. The WhatsApp channel is the one that changed adoption, because it asks suppliers to change nothing.
Hybrid OCR plus LLM, per-field confidence
Gemini as the primary extraction model with Cloud Vision available as hybrid OCR and OpenAI as an alternate provider. Around 26 header fields plus line items per document, each carrying a confidence score, with totals cross-checked against line-item arithmetic.
Draft, approved, pushed
Nothing enters the system of record on model output alone. Low-confidence fields are flagged for review, an approver signs off, and only then does the record export to Excel or PDF or push onward to an ERP through the integration layer built for phase two.
Multi-tenant, stateless, portable
A Node and PostgreSQL backend with app-layer tenant isolation and signed webhooks, and a bilingual React interface with correct right-to-left rendering. Documents stream through extraction and are not persisted to disk. Dockerised, so the whole platform lifts onto a client's own cloud.
Result
The platform is live in production on Google Cloud, serving the Saudi market with a bilingual Arabic and English interface, reaching 93.5 percent field-extraction accuracy on MVP scope. Re-keying is gone; what remains for humans is reviewing flagged fields and approving, which is a fraction of the original effort per document.
What the accuracy number really means
93.5 percent is measured, not marketed, and it is why the review workflow exists. Roughly one field in fifteen needs a human eye, concentrated in the worst documents. The design goal was never zero human involvement. It was moving people from typing everything to confirming the few things the system is unsure about, and being honest about which is which.
The client is credited by sector and region, KSA finance operations, in line with the engagement's confidentiality terms.
Have documents that need to become data?
Send a sample and the monthly volume. You will get a written read on feasibility, expected accuracy on your actual documents, and a fixed scope within one working day.