The most common lie in an automated desk is a 200 with empty meaning. The job finished. The invoice is missing a line. The quiz has two right answers. The dashboard shows yesterday. The model is not evil. It is done. Done is not the same as checked.
We have upgraded systems that now hold up to 5,000 businesses. At that size the happy path still looks fine. The damage sits in the cases nobody wanted to write down: a renamed column, a duplicate send, a vendor who put the amount in a note. A model that “reviewed” the file will still miss those if you never showed it the ugly pile.
Build a pile of the work people actually did by hand
Before we trust a flow, we sit with the person who used to do the job and we collect the weird weeks. Not ten clean samples. The refund. The missing VAT. The customer who replies in a screenshot. The course lesson that has no still because the outline was a bullet. Those are the tests.
Then we run the flow against that pile with the person in the room. If the model skips a row, we do not “tune the prompt” and ship. We add a branch or we stop the job. Shipping a half-check is how you teach the floor that the tool is decoration.
Compare the machine to last month’s human file
If you have a month of work a person already finished, you have a test. Run the new flow on that month and list every row that differs. Some differences will be the machine being right. Some will be the machine dropping a field. You will not know which until a person who did the job looks at the list.
Do this again after the first live week. Models drift when the input changes. Vendors change templates. Students ask new questions. A flow you blessed in June can lie in August if nobody walks the pile.