Industry
The read-match-approve pattern behind it has run for months in a manufacturer's live operations; here is what it does when the document is a COA and the clock is a lot release.
The grind
The PO says 2,000 lbs, the BOL says 1,962.8, and the invoice prices the shipped weight against a bracket that changed mid-contract. Your AP team re-keys weights off PDFs and does the math by hand on every line, because a standard three-way match was never built for variable weights.
Every inbound lot needs a certificate of analysis matched against your spec before it can move to production, but suppliers send COAs as email attachments in a dozen formats — or not at all. QA spends hours a day matching paper to receipts instead of doing quality work.
Broker and distributor deductions land on remittances as a code and a dollar amount, and validating one against the trade agreement, the promo calendar, and the shipment record takes longer than the deduction is worth. So invalid chargebacks get written off, month after month.
The lot data exists — in receiving logs, batch sheets, and shipping records — but a mock recall or a customer's one-up, one-back request means someone reconstructs the chain by hand. That's a fire drill every time, and FSMA 204 timelines are only getting tighter.
The split
We automate
COA intake and spec matching — reads supplier COA PDFs and checks every value against your spec sheet, flagging anything out of spec for QA
Catch-weight invoice reconciliation — extracts shipped weights from invoices and BOLs, then recomputes extended pricing against the contract — deterministically, not by AI guess
Deduction and chargeback validation — matches each remittance deduction to the trade agreement and drafts either the dispute package or the write-off recommendation
Traceability record assembly — pulls one-up, one-back lot records from receiving, batch, and shipping documents into a single audit-ready package
USDA/CFIA and customer paperwork prep — pre-fills recurring export certificates, continuing guarantees, and customer questionnaires from spec and facility data you already hold
We protect
Lot release and hold decisions — no lot moves to production or ships on an automation's say-so; QA approves every release
Out-of-spec dispositions — reject, rework, or accept-on-deviation stays a human quality call, every time
Supplier and broker conversations — disputes and corrective actions stay human; automation only assembles the evidence
Recall judgment — the trace assembles in minutes, but whether and how to act on it belongs to your people
How it runs
AI drafts. A person reviews and approves. Nothing touches a live system without a human gate — that's the architecture, not a phase.
The full build architecture →A worked example
A tanker of liquid sweetener is due Tuesday; the supplier's COA arrives by email Monday night. By the time your QA lead sits down with her coffee, the automation has read the PDF, matched the lot number to the inbound receipt, checked every value against your spec, and staged a release recommendation — with one result flagged as sitting near the spec limit. She reviews the flag, judges it acceptable, and approves the release; her name and timestamp go on the record, not the machine's. Had she rejected it, the disposition — return, rework, or accept on deviation — would have been entirely her call, because we deliberately don't automate that. And until the system has proven itself over months of live lots, the manual check runs in parallel, so her team never loses the ability to do it by hand.
No. Your QA techs stop re-typing COA values and get that time back for supplier development and root-cause work; your AP team stops hand-checking catch-weight math and starts recovering the chargebacks that used to be written off. Same people on the schedule — different work filling their day.
Not yet — and the distinction matters, so we lead with it. Our production system runs at a North American manufacturer and packager, doing high-volume document matching with a person approving every result. A COA checked against spec limits is the same mechanical problem as a vendor quote checked against a last-paid price; what changes is the paper, and the deadline attached to it.
It's designed to reduce it. The AI only extracts text from documents; every check against a spec or a contract is deterministic logic you can inspect, and nothing enters your quality or financial records without a named person's approval. What an auditor sees is the same record you keep today, plus a cleaner trail of who reviewed what and when. Your food safety plan stays yours — we prepare records for sign-off, we don't sign them.
We sit with your team, map how the work actually happens, and draw the split: what's worth automating, and what automation should never touch. You keep the map either way.
Map your workflows