An AI agent that clears the invoice backlog and knows when to ask a person
An agentic AI for bilingualinvoice and purchase-order processing ata distribution company, three-way matched,human-checked, and fully audited.
The Challenge
Every month, thousands of supplier invoices and purchase orders arrived as PDFs, scans, and email bodies, in two languages and no consistent layout. A back-office team re-keyed each one and matched it by hand against the order and the goods-received note before it could be paid: slow work, and exactly where duplicate payments slip through. Any automation here had to be trusted with the company's own ledger, which meant proving itself before it could act.
Mixed, messy inputs at scale
Invoices and purchase orders arrived as native PDFs, phone-camera scans, and email bodies, in inconsistent layouts across hundreds of suppliers, in English and Arabic. No fixed template survived.
Manual matching bred costly errors
A team matched every document line by line against the order and the goods-received note. It was slow, and exactly where duplicate payments and quiet miscodes creep in.
Automation had to earn trust
Finance would not let software post to the ledger unless every entry could be explained and reversed, and unless the system handed off the moment it was unsure.
Our approach
Every build follows the same software development life cycle, from requirements and design through build, testing, and support. Each phase is planned, demoed, and signed off before the next begins, so quality is engineered in rather than checked at the end.
Discovery & requirements
We turned the back-office problem into a written, testable specification: which documents the agent could own, what a person must always sign off, and the risk attached to each decision, mapped across hundreds of supplier formats.
(Outcome):
Architecture & design
We designed an agentic architecture around trust: a document-AI pipeline feeding a tool-constrained agent that can only act through validated, reversible operations, never writing freely to the ledger.
(Outcome):
Build
We built the pipeline and the agent in milestones, demoing real documents at every step, so the finance team saw exactly how each invoice was read, matched, and posted.
(Outcome):
Testing & UAT
Before the agent was given any autonomy, we measured it against the company's own history with an evaluation harness, then ran user acceptance testing with the finance team against agreed criteria.
(Outcome):
Deployment & support
We ramped from assisted to autonomous only on the categories that cleared the accuracy bar, with monitoring and a standing review to catch format drift as new suppliers onboard.
(Outcome):
Outcomes
Around 85% of documents now processed with no human touch
Processing time per document cut from roughly 12 minutes to under one
Hundreds of staff-hours a month returned to higher-value work
Every automated decision logged, explainable, and reversible