Skip to main content
All articles
Published on6 min read

The 6 KPIs to monitor on an AI voice agent in production

Once your AI voice agent is in production, gut feel stops working. Here are the 6 metrics that actually matter — and the alert threshold on each.

Plenty of call centers running an AI voice agent in production track 'calls handled' and stop there. Not enough. Here are the 6 metrics we pilot on at VocazAI — and the threshold at which we trigger a review.

1. Completion rate#

Share of calls that reach the intended action (booking taken, info given, human transfer). Healthy threshold: > 85 %. Below 75 %, the system prompt or scope is poorly defined.

2. Human transfer rate#

Share of calls that end transferred. Too low (< 5 %) = agent tries to do everything alone, risky. Too high (> 25 %) = it's no longer doing its job. Sweet spot: 10-20 %.

3. Average call duration#

Target: 60-90 seconds for a simple booking, 90-150 seconds for qualification. If you drift to 3+ minutes, either the prompt is too chatty or the user is fighting the agent. Listen to 5 long calls.

4. p95 latency per turn#

  • p50: under 500 ms.
  • p95: under 800 ms.
  • p99: under 1,200 ms.
  • Above that, agent feels broken and triggers hangups.

5. Post-call CSAT#

A post-call SMS with 1-5 rating or just thumbs up/down. Target: > 4/5 average. Important: don't measure after a missed emergency — biased. Exclude those from the math.

6. Business conversion#

The only one that truly matters: how many calls become honored bookings, sales, qualified leads. Measure it CRM-side, not platform-side. Compare to the prior month without the agent.

Review cadence#

Weekly the first 4 weeks, monthly after. If a KPI drifts 2 weeks in a row, you correct. Otherwise, let it run. VocazAI ships the dashboard for these 6 metrics by default. First month free to set thresholds on your activity.