Call Sample Review Gate (QA)

Call Sample Review Gate (QA)

1. 01

   Pull a minimum of 20 real call recordings per client before signing off on any deployment

   Split the sample across call types the client actually receives: new inbound, after-hours, existing-customer support, and outbound follow-up. Twenty calls is the floor for spotting a pattern; five calls only shows you anecdotes.

2. 02

   Score escalation accuracy against a written rule set, not a gut feel

   Define in advance which intents must reach a human (billing disputes, clinical questions, legal threats) and count how often the agent routed them correctly. Trillet and PolyAI both expose routing logs you can reconcile against the client's own ticket system.

3. 03

   Measure end-to-end latency on the worst 10 percent of calls, not the average

   Average response time hides the calls where the caller talks over the agent or hangs up. Flag any turn exceeding roughly 1.5 seconds of dead air and check whether the cause is model inference, telephony, or a slow calendar lookup.

4. 04

   Verify consent and recording disclosure on every outbound campaign

   Two-party consent states and regulated verticals require an explicit disclosure before the agent speaks its script. Confirm the disclosure fires on the first turn and that the transcript stores the caller's acknowledgment.

5. 05

   Reconcile usage cost per resolved call against the retainer line item

   Pull the platform's per-minute or per-resolution billing for the same 20 calls and divide by outcomes the client counts as resolved. If cost per resolution exceeds the human alternative, the offer needs repricing before renewal.

6. 06

   Document the labor that remains after deployment

   List every task a human still performs: transcript review, CRM cleanup, script edits, escalation callbacks. Agencies that skip this step quote a retainer that quietly loses money by month three.

7. 07

   Run a blind listen with the client's front-desk lead

   Have the person who currently answers the phone score the same recordings without knowing which agent handled them. Their objections surface handoff gaps that internal QA misses.

8. 08

   Freeze the reviewed script version and log the change date

   Once the sample passes, tag the prompt and flow version in the platform so the next QA cycle compares like with like. Unversioned edits are the most common reason a passing deployment degrades within 30 days.