Purchasing, rebuilt
as a system.
The purchasing core of a mining operation ran by hand: every order read, classified and typed into the ERP by a person. We rebuilt that core as an operation of coordinated agents. It went into production, inside the system the company already used.
Every purchase order went through a person reading the invoice, checking the sample, finding the right code in the catalogue and typing it all into the ERP. It took about 30 minutes, depended on whoever was in the chair, and one wrong digit became a wrong entry deep inside. The bottleneck wasn’t the system. It was the manual work on top of it.
Invoice and load sample arrive as PDF, photo or XML.
One agent extracts items, quantities and supplier. Another reads the sample.
The material is classified against the catalogue, with confidence measured and logged.
The buyer reviews and approves. Nothing is written without a human in the loop.
The order is posted inside the ERP, transactionally, and re-read for verification.
Not on top.
Inside.
Most AI tools sit outside the ERP, pushing data through brittle integrations. This was different. The agents write transactionally inside the client’s ERP. Every write is confirmed by a human and re-read right after, to verify that what was written is what was approved. Reversible and auditable.
Integrating inside a system like that is the part nobody solves with an off-the-shelf connector. That’s where the system proved its worth.