The question we are asked least, and should be asked most, is what happens in month three.
A system that is built and left alone gets worse. Suppliers change their invoice layout. A new service is added and nobody tells the phone line. A model is updated by its maker and starts answering a question differently. None of that is dramatic. All of it adds up.
So running is a job, and this is what it looks like.
Watching
Every call, message and document the system handles leaves a log. Not a transcript to read, but a record of what came in, what the system did, and where it stopped for a person. We look at the ones it handed over, because that is where it was unsure, and the ones where a person afterwards changed what it did, because that is where it was wrong.
The report
Once a month, the owner gets a short written report. What the system handled. What it got wrong, with the cases. What was changed as a result. It is a page, not a dashboard, because a page can be read in five minutes and argued with.
Changing it
Something is changed almost every month. A new question added to the page. A rule tightened. A supplier's new layout taught. Each change is run against the same test the system passed before launch, and it does not ship if it scores worse. That rule is dull and we keep it without exception, because the alternative is a system that drifts a little each month until nobody trusts it.
Who does this
The person who built it. Not a support queue. That is a deliberate choice: the person who knows why a rule was written is the one who should change it.
The terms this is written into are on the Approach page. They are short, and the running part is most of them.