Transcript Sampling and Scoring Cadence (QA)

Transcript Sampling and Scoring Cadence (QA)

1. 01

   Define the scoring rubric before the first call is ever recorded

   Lock 5 to 7 dimensions (intent accuracy, escalation correctness, compliance disclosure, tone, resolution rate) and weight them. A rubric written after listening to calls bends toward whatever went wrong that week.

2. 02

   Set the sample size as a percentage of monthly volume, not a fixed count

   A 200-call sample means 2% coverage on a 10,000-call retainer and 40% on a 500-call one. Express the target as a percentage floor (commonly 3 to 5%) so it scales with client growth.

3. 03

   Stratify the pull across outcome, channel, and time of day

   Random sampling under-samples the failures that matter. Force quotas for abandoned calls, transfers to humans, after-hours contacts, and any channel the client added in the last 30 days.

4. 04

   Score AI-handled and human-handled calls on the same rubric

   Separate rubrics hide the real comparison. When a platform like NICE CXone or Talkdesk reports autonomous resolution, the same quality bar should apply to the calls it hands off.

5. 05

   Route every failed dimension to a named owner with a 5-business-day fix window

   A score without an owner is a report. Assign prompt edits, routing changes, or knowledge-base updates to a specific person and log the due date in the client's delivery tracker.

6. 06

   Publish a one-page scorecard to the client before the monthly retainer review

   Lead with trend lines, not raw scores: resolution rate, escalation rate, and the two dimensions that moved most. Clients renew on direction of travel, not absolute numbers.

7. 07

   Re-baseline the rubric every quarter or after any model or prompt change

   A vendor model swap can shift behavior overnight. Treat any platform update, voice change, or new intent as a trigger to re-score a fresh sample before declaring the change a win.