A ticket comes in. The routing model reads it and finds nothing it has seen before. It returns no category. The ticket lands in the intake queue with a blank field. The SLA clock starts anyway. The clock was never told that blank is a state.
By mid-afternoon, someone has filed a second ticket. “AI routing is down.”
Nothing is down. The model had nothing to say. The queue had nowhere to put it.
Empty Was Always a Failure
Software has never had a slot for none. A null is a bug. An empty response is a timeout. A blank field fails validation. Every system downstream was built on the assumption that something comes back. The dashboard needs a number. The report needs a row. The next step in the workflow needs an input. All of it was built for software that gives the same answer every time.
A model that declines to answer breaks all of that at once. Not because it was wrong, but because it was honest in a format nothing was built to receive.
Last month I wrote that a model hallucinates when there is a gap, and that the work is removing the guessing. This is what comes out the other side. Remove the guessing and some questions get no answer. That is the design working.
Fewer Hits
Every detection product makes the same trade at some point. Flag every correlation, and people drown. Flag only when there is risk, and the system goes quiet.
The first design is trusted because it is loud. Something appears every hour. People click it, dismiss it, complain about it. The complaints are proof it is running.
The second design has to earn trust while producing less evidence that it exists. Same workflow. Same screens. Fewer things on them. The only visible change is the drop.
People built their sense of “working” on the count. Adoption dashboards built theirs on clicks. Less to interact with reads as less adoption. A more accurate product can show up in a usage report as a product nobody opens.
The honest system looks like the broken one.
What the Silence Says
A guess hides the gap. The model does not know, writes something plausible, and the gap disappears into a confident paragraph. Nobody learns where the data stopped connecting.
Silence points at the gap. The model does not know and says so. That output tells you exactly where the work is undone.
A false positive says the same thing from the other side. The system proposed, a person looked, and the person said no. That no is information. It says the system connected things that do not belong together, which is a map of what the data has not taught it yet.
Most teams throw both away. They log the empty result as an error. The false positive gets logged as a mistake. Both go into a table nobody reads, and the next quarter the system guesses the same way.
The team I work with proposes on everything, lets a person mark the misses, and uses the pile of misses to make the next proposal better. That loop is the product. Volume never was.
The Number That Went Missing
A system that got quieter did not get worse. The organization lost the number it was using to tell.
The old number was easy. Hits per day. Alerts per week. Something to point at in the review. The new number is harder to read. How often did it decline? How often was it right when it spoke? How many misses did it take to stop making one?
Nobody has that number on a dashboard yet. The ticket that says “AI routing is down” is what gets filed instead.