The six conversation metrics that catch what voice agent dashboards miss
Most teams have shipped voice agents and most are not happy with them. In Hamming AI's State of Voice AI 2026, 87 percent of companies have deployed voice agents but only 12 percent are satisfied with the quality. The gap is not latency or uptime. It is conversation-level quality: contradictions across turns, missed intents, escalations that fire too late, and sentiment that slides before a hang-up. Standard dashboards do not surface any of that.
The dashboards from a couple of years ago stopped at three numbers: latency, completion rate, and sentiment. Those catch the obvious failures, the calls that time out, do not resolve, or end with an openly angry customer. What they miss is the long tail where every number reads green and the experience still degrades. These six conversation-level metrics catch that long tail. You do not need all six on day one. Start with completion and turn coherence, then add the rest as call volume grows.
- Turn coherence. Does the assistant stay consistent across turns. It catches a fact confirmed in turn 3 and contradicted in turn 7, a tool result that comes back but never gets used, and a reference like "the second option you mentioned" the assistant has forgotten. Each turn reads fine on its own; the failure is in the connective tissue between them, so a single-turn check never sees it. Score it with a multi-turn coherence rubric over the full message history.
- Intent confidence. Did the system identify what the caller actually asked at the entry point. It catches a vague opening routed to a default flow, a mistranscribed accent that feeds the classifier a garbled input, and two requests in one turn where only one gets handled. Intent taxonomy is org-specific, so this is usually a custom check, and it pairs with audio-transcription scoring when STT drift is the real cause rather than the classifier.
- Completion rate. Did the call resolve the caller's goal. Track two versions: customer-perspective resolution (did they get what they came for) and agent-perspective task completion (did the assistant do the right thing). The split is the signal. A correct out-of-policy refusal scores high on task completion and low on resolution, which tells you the failure was policy-induced, not a capability gap. The delta between the two is more useful than either alone.
- Sentiment trend. The slope across turns, not the static end-of-call sentiment. It catches a caller who enters neutral, sours by turn 3, and hangs up before the agent notices, and it credits the agent that de-escalates a frustrated caller back to positive. Score sentiment per turn and alert on negative slopes that cross a threshold; on outbound calls a falling slope often predicts a hang-up a couple of turns out.
- Escalation triggers. When and why a call hits a policy or scope boundary that needs a human. It catches escalate-out-of-caution, where the agent could have handled it but handed off anyway, and the reverse, where the agent should have escalated and tried to muscle through instead. Pair a refusal-justification check with a handoff-quality check that scores whether the human got enough context to avoid re-collecting the basics.
- Repeat-question signal. Did the caller ask the same thing twice because the first answer was not useful. This is the one that catches silent CSAT decline: intent is right, coherence is high, completion fires positive, but the caller had to ask three times. Score it with a custom signal that compares semantic similarity between the caller's turns and flags pairs above a threshold.
The metrics are most useful in combination, not one at a time. Group calls by the set of verdicts rather than by a single score. Low completion plus high repeat-question plus neutral sentiment is the silent-degradation pattern worth fixing. High completion plus declining sentiment plus one escalation is usually recovered-after-friction and fine to leave alone. Clustering the low-scoring calls into named issues, each with a root cause and a fix, is what turns six metrics into an actionable backlog instead of six dashboards to scan.
A few things that go wrong in practice. Running all six on 50 calls is noise, so start with two and add the rest once each score is stable. Treating sentiment as a single end-of-call label throws away the slope, which is the part that carries the signal. Conflating an escalation trigger (the agent decided to hand off) with an escalation outcome (the human resolved it) hides a real failure cluster. And skipping repeat-question because it needs a custom evaluator drops the highest-signal metric for quiet CSAT decline.
The point is that the 12 percent satisfaction gap lives at the conversation level, and infrastructure dashboards never reach it. Completion and turn coherence get you started; the other four cover the failures that only show up once you are looking at the whole conversation rather than a single turn.
