Ask anyone to sketch an AI system that sends invoices and you get the same five boxes: a folder of documents, an agent to read them, a template to fill, an invoice it produces, and an arrow marked send. Draw it in five minutes and a whole room will nod. We call this the Mickey Mouse architecture — not for the character, but in the old business sense: trivial, flimsy, working only on the surface and certain to buckle under real-world weight.
The sketch is still worth drawing, because it settles quickly whether everyone agrees on what the thing is for. It just isn't a design, and the honest move is to name the failure hiding under each box before anyone builds it.
Under folder sits the problem of running again and again: an hourly job sees the same document twenty-four times a day, and without a memory of what it already handled, one invoice cheerfully goes out three times. Under agent sits the wrong-client trap, where a near-enough address match quietly sends Client A's private figures to Client B. Under template sits the risk of leaking your own bank details. Under invoice sits the confident wrong number a model will produce when a figure isn't actually there. Under send sits the deepest one: does a human ever actually look before it goes?
Engineering it properly means answering each of those in turn — a memory so nothing is processed twice, tiered matching so uncertain cases wait for a person, a sealed template, figures that carry a source, and a genuine stop-and-check before anything leaves the building.
The sketch isn't wrong. It is unfinished. The value of drawing it is that it shows you, in five boxes, exactly the five places you now have to do the real work.
“The sketch isn't wrong — it's unfinished. Its value is showing you, in five boxes, the five places the real work lives.