Queue noise versus intent

A Midlands retail desk we taught last autumn labelled every second contact as “follow-up.” In the CRM that looks tidy. In the thread it is a collision of three motions: a customer checking because nothing arrived, a customer repeating because the first agent guessed, and a customer confirming a change they already made. Only the middle one is a teaching problem. The other two are logistics and copy.

Customer support signal analytics starts by refusing the label. We ask what the second contact is trying to complete. If the answer is “see a tracking pixel move,” you are in courier theatre. If the answer is “undo a wrong promise,” you are in policy fog. Averaging them into failure demand makes a pretty slide and a useless rota.

In Signal Taxonomy Lab we use a two-column sheet: what the system called it, what the last three sentences were actually doing. The sheet is boring on purpose. People who skip it go back to arguing about reopen rate, which is a count, not a reading.

Intent is not a psychological claim. We do not need to know how the customer “feels.” We need to know which unfinished job pulled them back. Noise is the remainder once that job is named — the duplicate tabs, the accidental double-sends, the bot that asked them to start again. Noise still costs time. It does not deserve the same intervention as a wrong promise.

If your export cannot show the last three sentences, stop. No amount of modelling will recover a signal from a status code. Write to the studio if you want the sheet we use; it is not software.