Five metrics that tell you whether your voice agent is working
Containment, escalation quality, time to resolution, repeat contact rate, and post-call sentiment — and how each one misleads you on its own.
Most voice AI dashboards report the wrong things beautifully. These five numbers, read together, tell you what is actually happening on your lines.
1. Containment rate
The share of calls resolved without a human. It is the headline number and the easiest one to game — an agent that refuses to transfer will show excellent containment and terrible customer outcomes. Never read it without the next two metrics beside it.
2. Escalation quality
Of the calls that transferred, how many arrived with usable context, and how many made the caller start over? A high transfer rate with clean handoffs is a healthy system. A low transfer rate with confused handoffs is a system hiding failures.
3. Time to resolution
Measure from the caller's first word to the point their problem is actually solved — including any transfer. Average handle time on the AI leg alone is a vanity metric; it improves every time the agent gives up faster.
4. Repeat contact rate
How many callers come back within seventy-two hours about the same issue? This is the single hardest metric to fake and the most honest signal that a call was genuinely resolved rather than merely ended.
5. Post-call sentiment
A short survey or a sentiment read on the final seconds of the call. Treat it as a directional trend rather than a precise score, and always segment it by intent — one badly handled intent can drag an otherwise healthy deployment's average down.
The weekly review that keeps them honest
- Read the five metrics together, never one in isolation.
- Listen to five full calls: two contained, two escalated, one flagged by sentiment.
- Segment everything by intent — averages hide the failing case.
- Write down one change, ship it, and check the same numbers next week.
Keep reading