Useful conversation data preserves the difference between what someone requested, what was discussed, and what the recording shows actually happened.
Choose the decision moment
State whether the system classifies an opening request, follows a changing intent, or summarizes a completed interaction. Provide input and output examples. A post-call label can depend on later information, so keep label timing and source visible when the model is expected to act earlier.
Make the label schema concrete
Define intent boundaries, multi-intent behavior, uncertain labels, and the summary audience. Separate topics, requested actions, proposed solutions, and completed outcomes. Require evidence intervals for consequential summary claims so reviewers can identify unsupported conclusions instead of judging a plausible narrative as automatically correct.
Preserve conversational structure
Include turns, transfers, follow-up explanations, and unresolved calls when they occur in scope. Decide whether the model uses original audio, transcripts, or both. Review how transcription errors and missing segments affect reference answers, particularly around numbers, negation, and action commitments.
Plan a project-specific review
Bring Fiund the taxonomy, output format, domain priorities, and sensitive-information handling requirements. Potential source fit, permitted use, and available annotations need assessment for the actual project. A varied pilot should test the references and reveal coverage gaps before a larger sourcing discussion.
Include in your inquiry
- Task timing and input artifacts
- Intent or summary schema with examples
- Evidence and uncertainty requirements
- Domain coverage and review workflow