started planning a real analysis of call transcripts for the first wave of clients on a thing that launched at work. the instinct is to jump straight into building the tagging pipeline, but step one is boring and it is the bottleneck: there is no transcript data yet. nothing local, and the recordings sit behind a compliance process. one access request for the right cohort and time window, with the purpose stated as post launch gap analysis, and that request is the long pole of the whole project, not the analysis itself.
the order matters more than the tools. the client list from the launch tracker first, then per client cutover dates, then transcripts from cutover forward. calls before cutover are the baseline, they are not the subject. and sample before you automate. code twenty calls by hand, fix the reason codes, then tag the rest. hand coded calls are how you find out the codes are wrong before you have burned a week labeling garbage.
the schema came together clean: reason code, which system was named, sentiment, resolution, and whether the client called back within two weeks. roll it up by repeat contacts and severity, not by volume. two guardrails carried over from earlier arguments: the transcripts only justify the agreed subset of fields, they do not rebuild the whole system, and report by client id, not by name, unless the audience is cleared for names.