Phase 7 · 7.2
Monitoring and Alerting
Effective monitoring means watching the right thing rather than watching everything. In a voice system running at enterprise scale, the indicators worth watching group into a few layers:
| Layer | Example indicators |
|---|---|
| Conversation quality | End to end latency, completion rate, escalation rate |
| Infrastructure health | Resource usage, concurrent active calls, line and channel occupancy |
| Voice line quality | Packet loss, jitter, dropped call rate |
| Business outcome | Task completion rate, write back errors |
Alert thresholds should be layered so that they separate "slightly outside normal" from "needs immediate attention". For example, if the escalation rate rises above 30 percent of calls in the last ten minutes, that alone should trigger an alert. It is usually an early sign that something is wrong upstream, in an integration or a data source, rather than in the Voice AI Agent itself. An alarm for every small deviation, on the other hand, buries the real problems in noise.
At CBOT a webhook based event layer can stream these metrics in real time into the organisation's own monitoring systems, Prometheus and Grafana among them.