A mid-market distributor was paying three people to reconcile invoices across two systems. The accounting platform said one thing. The vendor portal said another. Every week, three people spent a chunk of their job finding the gap and arguing about who was right.
Nobody called it an AI problem. They called it "the Tuesday thing." That is usually where the money is.
What it actually was
Strip the frustration away and the task is mechanical. Pull the invoices from system A. Pull the matching records from system B. Compare them on a handful of fields: vendor, PO number, amount, date. Most of the time they agree. When they do not, a human needs to look.
The work that felt like judgment was mostly lookup. The actual judgment was a thin slice at the end.
The build
A scheduled job reads both systems overnight. It matches records on the key fields and applies the same rules the team had been applying in their heads, written down for the first time. Anything that matches cleanly is closed. Anything that does not is queued for a person, with both versions side by side and the likely cause flagged.
No model sits in the critical path. It is rules until the last step, and even then a person makes the call. The boring choice was the correct one.
The number
The team got back most of the time they were spending on the lookup. Call it an estimate in the range of a few hours per person per week, reclaimed and pointed at work that actually needed a human. The build cost less than one of those people costs in a month.
It is not interesting. It did not use a single feature anyone would put in a keynote. And it was the highest-ROI thing they shipped all year, because the line to the P&L was one step long.
Why this one and not the flashy one
The same company had a list of exciting ideas. A chatbot. A forecasting model. A "copilot" for the sales team. Every one of them had a longer line to the money and a fuzzier definition of done.
The reconciliation build had a clear before, a clear after, and a number you could check. That is the test. If you cannot draw the line from the build to the P&L on one slide, it goes to the bottom of the list.