Accounts payable / AI
Accounts payable / AI case study — how Meritt Dev built and runs Waste management company: 40,000 hauler invoices a month, matched before they are paid Waste management company: 40,000 hauler invoices a month, matched before they are paid
An asset-light waste management company runs customers' waste programmes nationwide and contracts the hauling out to about 1,400 haulers. Around 40,000 hauler invoices arrive every month, and reconciling them by hand forced a bad choice: pay fast and lose margin, or pay late and risk haulers cutting off service.
Overview
We built the accounts payable review system that reads every invoice, finds the purchase order it belongs to and puts a scored match in front of an auditor. Meritt's founders worked on the engineering team with our US partner studio.
- [01]
The challenge
Invoices come in by email as PDFs in hundreds of hauler layouts. The purchase-order data lives in the company's data warehouse, and a match has to be made on service address and dates, not on a reference number that most haulers never quote.
Automatic payment was never the goal. Every decision had to stay with a person, with the system doing the reading and the matching and showing its work.
- [02]
What we built
A shared mailbox is read through Microsoft Graph; PDFs are virus-scanned and kept in blob storage. Document AI and a language model extract the invoice fields. An address-first scoring model, with rules agreed with the client's AP team, ranks candidate purchase orders from a nightly copy of the warehouse.
A leader screen assigns invoices to AP auditors, who record Pay, Escalate or Trap Door on each one. Confirmed payments export to NetSuite as a CSV, with direct write-back planned as phase two. A ranking model learns from auditors' decisions in shadow mode without changing what they see. An assistant answers questions about the queue.
Next.js and Postgres, running in the client's own Azure tenant.
- [03]
Where it is now
Discovery in late July, kickoff on 12 August, on-site sessions with the AP team in mid-September, auditor training in October.
On the first test of 279 invoices already booked in the client's records the right purchase order was found 96% of the time, and it was the top suggestion 82% of the time. On a 2,000-document held-out set, 222 of the 229 invoices the system marked ready for pay review were correct.
Related