cx-conversation-samplinglisted

# Sampling conversations Every manual review of support conversations is a sample, including the ones nobody calls a sample. The question is only whether it was drawn deliberately. This is about **how to draw it and how to weight the results**. How many you need for a given precision is a different question — a sample can be perfectly sized and still answer about the wrong population. ## The frame comes first The **sampling frame** is the list of conversations you could have drawn from. Almost every generalisation failure is a frame problem, not a size problem. Write down explicitly: - **What is in the frame** — window, channels, queues, statuses, languages. - **What is silently missing.** Conversations not synced from a channel; deleted, merged or redacted records; anything the API does not return; a channel connected halfway through the window. Each of these removes a non-random slice. - **What you excluded on purpose**, and the rule. **A frame that omits a channel cannot support any organisation-wide claim**, however large the sample. Say which population your result actually describes — that sentence is the deliverable as much as the number is. ## Choose the design to match the question **Simple random.** The default, and correct when you want one overall estimate. Boring and defensible. Use it unless you have a specific reason not to. **Stratified.** Split the frame into strata — channel, queue, team, language, risk band — and sample within each. Use it whe