An agent, now implemented, that reads invoice emails straight from Gmail, matches each one to its purchase order and production batch, and shows how cost per can would change — with a person approving everything before it touches the books.
Order placement, logistics tracking, and invoice entry into the Excel tolling logs were all manual — and those logs drive the batch-level cost-per-can figure. With the operations analyst role no longer on the team, keeping them current competed with everything else. Every co-packer, ingredient, and freight invoice meant opening the email, reading the PDF, finding the PO, working out which batch it belonged to, and keying it in.
This is where applying AI pays off — and where it deliberately stops. The AI never does the math, and it never changes the books.
AI. An AI model reads each invoice PDF and writes down what it says: vendor, line items, totals, PO reference.
Fixed rules. Deterministic logic matches the invoice to a PO and production batch, and recalculates cost per can.
Fixed rules. Mismatched totals, duplicates, missing POs, and credit memos are flagged with a plain explanation.
A person. Clean invoices are approved in one click per batch. Flagged ones get an individual decision, logged with who and when.
Most of the test invoices contained a deliberate problem, because real invoices do.
Each one built to trip the logic, checked against an answer key. The matching handled all 12 as required.
Demo invoice emails with PDFs attached, read straight from the inbox with no uploading step.
We built this agent on two different AI models to see what each did well. The differences shaped the design we took into implementation.
Intake, matching, and approval, verified end to end.
Built as a standing agent that takes on more of the analyst's workload.
Prototype A's tested extraction, matching, and flagging, with Prototype B's continuous monitoring, drafted POs, weekly status, and parallel-run scorecard layered on.
An illustrative estimate, not a measured result. It assumes about 60 invoices a month, ~12 minutes each by hand versus ~2 minutes to review one. The real figure depends on the client's actual volume.
Open the email, read the PDF, find the PO, match the batch, key it into the cost log.
Review what the agent read and matched, approve the clean ones, decide on the flagged ones.
Beyond hours: reconciliation errors get caught before they reach the books rather than after.
These are the measures we track to confirm it's working. Results get reported once they're measured on real volume.
How long it takes from an invoice, PO, or shipping document arriving to a validated entry in the tolling log. Cost per can is only useful if it's known soon after a run, not weeks later.
The gap between projected and actual cost per can, and how many post-run invoice surprises show up. This is where silent margin leakage hides.
The share of documents logged without a human correction, and accuracy on the critical fields: quantities, line-item pricing, batch numbers, freight surcharges. Anything uncertain goes to a person.
Manual hours saved, net of what it costs to run the agent per invoice. The figure that shows whether it pays for itself.
The agent is now implemented for the client, with a person approving before anything is written to the books. We scoped the build at 4–6 weeks, with messy, inconsistent invoice formats as the main risk. The time-saved figures above stay estimates until we measure them on real volume.
The pattern is the point: find the manual, error-prone work, put AI only where reading and judgment are needed, keep rules and people in charge of the rest.
Book a free scoping call — a short conversation about your current systems and where the team feels the most pain, manual work, or lag.
Book a free consultation