Tips for Best Results

Always upload a manifest — Call date/time is what makes the trend, the date filters and the shift comparison meaningful. Without it, everything is placed by upload time. See Call Metadata
Use stable agent IDs — agent_id from your ACD or workforce system keeps agent statistics correct across name spellings and duplicate names
Upload the manifest whenever it suits you — Rows with no matching file yet are kept and applied to recordings uploaded later
Fill in campaign and direction — They cost nothing to include and become filters in the results table

Use high-quality recordings — Clear audio produces better transcriptions, which leads to more accurate analysis
Ensure speaker separation — The system needs to distinguish between agent and customer to evaluate agent behavior
Review transcriptions first — Check a few transcriptions for accuracy before running analysis on a large batch
See the Tips for Best Results in the Transcriptions Guide for more recommendations on audio quality

Be specific in descriptions — Detailed criterion descriptions help the AI evaluate more accurately. Instead of “Good greeting”, write “Agent greets the customer by name, identifies themselves and the company, and offers assistance”
Use appropriate point weights — Assign higher point values to criteria that have more impact on call quality
Limit criteria count — 5–15 criteria per rubric is ideal. Too many criteria can dilute scoring precision
Use sections for organization — Group related criteria (e.g. “Opening”, “Problem Solving”, “Closing”) to make results easier to interpret

Reserve for non-negotiable items — Critical errors should only flag truly mandatory requirements (e.g. identity verification, legal disclosures)
Don’t overuse them — If too many criteria are marked as critical, most calls will score 0%, making the evaluation less useful
Document clearly — Write clear descriptions for critical criteria so the evaluation is consistent

Start with a small batch — Test your rubric on 5–10 calls first to validate the scoring before evaluating hundreds of calls
Use additional context wisely — The context field in the Analysis Wizard can provide useful background (e.g. “These are technical support calls for a software product”)
Combine analysis types — You can run both Calls Analysis and COPC Evaluation together to get both qualitative insights and quantitative scores

Validate the configuration on a small batch first — Run it manually over a handful of calls, check the output, and only then enable it for everything that arrives
Tune the cadence to your volume — A lower maximum wait makes results appear sooner on a quiet queue; the defaults (20 calls / 300 seconds) suit steady bulk uploads
Don’t edit custom questions mid-project — Answers already collected stay attached to the old question list, so the merged result would mix two question sets
Analyze the backlog once — Recordings uploaded before you enabled automatic analysis are never picked up; run New Analysis over them manually
Refresh contact reasons after a big arrival — Open-ended categories are re-grouped daily; press Refresh contact reasons to do it immediately

Check the date range before concluding anything — The report defaults to Last 30 days and filters by call date, so historical uploads can look like an empty dashboard. The banner tells you when calls exist outside the range
Look at criterion averages first — The Dashboard’s “Score by Criterion” chart and the report’s scorecard quickly reveal which areas agents struggle with most
Use the score range — A wide gap between min and max scores may indicate inconsistent agent performance
Compare against the previous period — The scorecard’s trend column shows whether a criterion is improving or slipping
Review critical errors individually — Expand each failed call to understand what triggered the critical error
Export for deeper analysis — Use Excel exports to filter and sort raw data, and the report’s PDF export for a client-ready summary