The most useful agent we have ever built runs once a week, at three in the morning on a Tuesday. It costs roughly four dollars a month to operate. It does one thing.
It reads the previous week's vendor invoices, compares them to purchase orders, flags any discrepancies above a threshold, drafts a clarification email to the vendor, and queues the email for human approval before close of business.
That is the entire scope. It is also a $40,000-a-year saving for the client we built it for, and a useful pattern for thinking about where agents actually belong inside a business.
The problem before the agent
The client is a 14-person operations team inside a mid-sized importer. Every Tuesday morning, the AP lead would arrive at her desk to find between twenty and forty new vendor invoices in her inbox. Most matched their purchase orders. A handful did not. Finding the handful was the work.
The work took her, on average, three and a half hours a week. It was tedious, it was unforgiving, and the cost of getting it wrong was either an overpaid vendor or a pissed-off vendor. She did it carefully, and she resented every minute of it.
What the agent actually does
The system has four parts, each small. It pulls invoices from the shared mailbox. It matches them to purchase orders in the ERP. It runs a simple comparison: line items, unit costs, totals, payment terms. When something is off by more than 2% or by more than $250, it drafts an email to the vendor — formal, polite, specific — and puts it in a review queue.
By the time the AP lead arrives Tuesday morning, three or four pre-drafted emails are waiting for her. She reads each one, edits if needed, hits send. Total time: about twenty minutes.
Why this works
Three reasons. The work was bounded — the same task, every week, with the same inputs and the same kind of output. The judgment was concentrated — the AP lead's expertise was needed for the four ambiguous cases, not the thirty-six clean ones. And the failure mode was visible — a wrong email gets caught at the review step, not in front of the vendor.
Most of the agents that fail in production fail because they are trying to do too much. This one succeeds because it is trying to do almost nothing.
The economics
Three and a half hours a week, fifty weeks a year, at the AP lead's loaded cost, comes to about $13,000 of reclaimed time. That is the conservative number. The real number is higher, because the time she got back was not low-quality time — it was Tuesday morning, the most productive part of her week, which she now spends on closing-cycle work that used to slip.
The build cost was eight hours of our time and a small monthly bill for compute. The payback period was about three weeks.
Why this is the pattern
When clients ask where to start with agents, the answer is almost never the impressive thing. It is almost always the small recurring task that someone competent has been doing on autopilot, that nobody has thought to question, and that is quietly costing more than its visibility suggests.
Find the Tuesday morning. Build the small thing. Move on to the next one.
