Policy labels

Support teams want to know what their tickets are about: which team, which part of the product, which kind of problem. I built policy labels (I did not pick the name) so they can define the categories in plain English, and every ticket gets labelled automatically as it arrives or changes. It came before "decision models" were a thing.

Each label shows the model's reasoning, so people can see why a ticket got it. A ticket can get one label or several, labels can build on other labels, and a new policy can be run over months of old tickets at once. It has been a success: easy for customers to use and easy to explain.

Seeing the trend

With Patrik I built dashboards on top of the labels. They turned out to be surprisingly useful: they show which team, or which kind of failure, is becoming more common over time.

Knowing how good they are

An AI label is only useful if you can trust it. People can approve or correct labels, and those become the ground truth. A customer can then try a new wording for a label, score it against that truth, and see whether it got better, all without touching the labels the team works with.