Choose the decision level
Decide whether intent belongs to the whole call, the opening request, a turn, or an evolving segment. A customer may begin with a billing question and later ask for a service change. A single final label can hide that progression and may not match a model intended to act at intake.
Write clear category boundaries
Use operational definitions with examples and counterexamples. Distinguish the customer’s requested action from the topic discussed and the eventual resolution. Include an unknown, ambiguous, or multi-intent path where needed rather than forcing every conversation into a tidy category that the audio does not support.
Keep label timing visible
Record whether labels came from a source system, an agent, or a later reviewer, and when they were assigned. Post-call labels may depend on information unavailable during the model’s decision window. Review disagreements and changing category names before treating the taxonomy as a consistent reference across sources.
Pilot realistic coverage
Inspect frequent and infrequent intents, transferred calls, unsuccessful interactions, and ambiguous openings. Measure source concentration and repeated scripts in each class. Share the intended decision moment, taxonomy, and evidence requirements with Fiund so a sourcing discussion can focus on calls that support the task and its review constraints. Test the taxonomy on a caller who asks about a charge and then requests cancellation. Ask reviewers which intent applies at the opening, after clarification, and at the end. This makes the timing rule tangible and shows whether the model task requires one label, several labels, or an explicit sequence of changing intents.
Review checklist
- Choose call, turn, or segment labels
- Define intent versus topic and outcome
- Record label origin and timing
- Include ambiguous and multi-intent examples