Submitting this for the September showcase. Built at a regional retailer, about 600 staff, roughly 40 new vendors a month and a procurement team of three who were spending most of a week each month chasing paperwork.
What it does. Reads whatever the vendor sends, in whatever format they send it. Extracts the tax identifier, the banking details, the insurance certificate and the signed terms. Validates the tax identifier format and checksum, screens against the sanctions list, and checks that the legal entity name matches the name on the bank account. Sends one message naming everything that is missing rather than five messages over five days. Escalates anything ambiguous to a person with the discrepancy written out in a sentence.
Numbers after nine weeks. Cycle time from first document to approved-to-pay went from 14 business days to 4. Share of onboardings needing a chase went from 71 percent to 38, and most of that improvement came from rewriting the intake form rather than from the agent, which was a slightly humbling finding to report to the client.
What it does not do, deliberately. It never writes bank details. Not once, not for a correction, not for a new vendor. That write stays with a person plus a phone call to a number from the contract rather than from the email. Two of their peers have been hit by payment fraud in the last three years and the client was clear that this was the line.
The thing I would do differently: I built the chase cadence at every three days with no stop. It sent one vendor eleven messages before anybody noticed. Two attempts then escalate to the requesting department is the version that shipped.