Somewhere in your business, someone compares two lists. They do it every week or every month, they do it line by line, and the reason nobody has automated it is that the hard ten percent is genuinely hard.
The job today
Two records of the same thing, produced by two different parties, that ought to match.
The supplier's invoice against what you actually ordered and received. The payment that arrived against the invoices it was meant to cover. The hours a crew logged against the jobs they were dispatched to. The commission statement from a carrier or manufacturer against what you calculated you were owed. The bank against the books.
Someone opens both, and goes down them. Most lines agree, and agreeing them takes a few seconds each. Some do not, and each of those becomes its own small investigation: pull the order, find the delivery note, email the supplier, ask the crew leader what happened on the fourteenth.
The pile of exceptions is the real job. The clean ones are just the tax you pay to find them.
And underneath all of it is the thing nobody says: on a busy month, the clean ones get skimmed rather than checked, because there are two hundred of them and the deadline is real. So the exceptions that look clean at a glance slip through, and those are precisely the ones that cost money — a two percent price increase nobody agreed to, applied quietly across every line.
What the agent does
It does the matching, all of it, every line, at the same level of attention on line two hundred as on line one.
It reads both sides — including where one side is a PDF that arrived by email, which it is more often than anyone would like. It lines up the records, which is harder than it sounds and is where the actual engineering goes: the same item described by your part number, their catalog number, and a free-text description that changed last spring is still the same item, and knowing that is most of the work.
Anything that matches within the tolerance you set is agreed and moves on.
Anything that does not is presented as an exception, and the presentation is the whole point. Not a flag saying something is wrong — the two records side by side, the specific difference highlighted, the supporting document one click away, and what it looks like this is: a short shipment, a price change, a duplicate, a freight charge nobody mentioned. The person's job becomes deciding, not investigating.
Tolerances are set per counterparty, because they genuinely differ. A small price drift from a supplier whose material cost moves weekly is expected. The same drift from a fixed-price contract is a problem. One global tolerance number is how these projects fail — it either buries the person in false exceptions or waves through the real ones.
What changes
The easy ninety percent stops consuming the week, and the person who did it spends their time on the exceptions, which is where their judgment was always needed.
The attention stops degrading. This is the part owners underestimate. A person on line two hundred of a bad month is not checking as carefully as on line five, and everybody knows it, and nobody says it. A system checks line two hundred exactly as hard.
And you find money you were losing quietly. Not dramatically, usually. A few tenths of a percent of purchasing, a handful of unclaimed credits, some commission that was never paid. It is rarely the headline; it is often more than the project cost.
What it connects to
Both sides of the comparison. One is usually a system with a proper connection — accounting, the ERP, the operational platform. The other is frequently a document: an emailed PDF invoice, a statement, a portal you have to log into. Reading that side reliably is the part to test first, with your real documents from your real counterparties, before anything else is built.
Your tolerance rules, written down per counterparty. For your biggest suppliers or carriers, what variance is normal and expected. If nobody can currently answer that, working it out is the first and most valuable part of the project.
A way to write the result back — matched items marked as matched, exceptions logged with what was decided and why. The log matters more than it looks: it is what lets you see, six months on, that the same supplier generates a third of your exceptions.
And the source document, always one click away. A person asked to approve a match they cannot inspect will re-check it by hand, every time. Showing the working is what makes the queue trusted rather than clicked through.