cx-emerging-harm-scanlisted

# Scanning for harm you have no category for Every monitoring system in a support operation looks for things someone already thought of. Tags, alerts, complaint categories and QA criteria all encode last year's known failures. Emerging harm is by definition outside that. It arrives as conversations that do not fit existing categories, gets tagged "Other", and is discovered months later — often by a regulator, a journalist, or a customer forum. This is the deliberately open-ended counterpart to categorised monitoring. It is exploratory, it will produce false leads, and it is worth running anyway because the alternative is finding out from outside. ## Where uncategorised harm actually shows up Six places, roughly in order of yield: **1. The "Other" bucket and its growth rate.** The most reliable single indicator. A rising Other rate in a stable taxonomy means something is arriving that the taxonomy cannot describe. Read a sample of Other every period — not the aggregate, the actual conversations. **2. Long conversations with no clean resolution.** Novel problems take longer and end ambiguously, because nobody has a script. Sort by turn count and look at the tail. **3. Agent uncertainty.** Internal notes asking colleagues, escalations with no destination, "has anyone seen this before". **Agents notice new failure modes weeks before any metric moves**, and this signal is usually sitting in internal notes that nothing analyses. **4. Conversations that escalate without a p