How we automated invoice-to-PO matching in six weeks
Back office automation is having a loud year. The big platforms are racing to put agents into finance and operations, and the marketing makes it sound like you flip a switch. The real work is quieter than that. Here is exactly how one build went, start to finish.
When the large vendors announce back office agents that complete work across your systems, they are pointing at a real and large problem: invoice processing, reconciliation, and data entry that still runs on people copying numbers between screens. The promise is right. What the announcements skip is the part that actually decides whether it works, which is the unglamorous mapping of how one specific company does the job.
The starting point
A 60 person logistics company was spending three days a week matching freight invoices to purchase orders by hand, across four different systems. The process lived in spreadsheets, in tribal knowledge, and in one person who knew where everything was. It was slow, but it was not random. That mattered.
What we mapped before we built anything
We went inside and watched the work for what it really was. The finding surprised them more than us: the matching logic was consistent. The same rules applied every time. A human was doing it only because nobody had ever written the process down end to end. Most of our first week is this, sitting with the people who do the job and turning their habits into explicit rules, including the rules they did not know they had.
What we built
A custom agent that does the matching automatically across all four systems, clears the clean matches, and flags anything that does not fit for a human to review. It sends a daily exceptions report so the finance team starts the day with a short, ranked list instead of an inbox of raw invoices. The human did not get removed from the loop. They moved up it, from doing every match to judging the handful that need judgment.
Why six weeks and not six months
Because we did not try to boil the ocean. We took one painful, well defined workflow, built the system around how this company actually ran it, and shipped a working version fast with a human in the loop for the exceptions. Then it earns the right to expand to the next workflow. That is the whole model: embed, map, build, support. If you have a back office process that is slow but not actually random, it is very likely a candidate for exactly this.