← sources

Our voice agent passes every metric we wrote and customers still complain. What are we not measuring?

source: reddit (reddit:AI_Agents)published 2026-11-17 06:44ingested 2026-11-18 02:36

retained text

We have about 15 metrics on our voice agent: task completion, tool call success, latency, a couple of LLM judges. All green. Support keeps forwarding complaints from calls that look fine by those numbers. Listening to calls one by one is the only way I've found the real issues, like the agent repeating itself after a long pause and one confusing way it offers appointment times. I can't do that for hundreds of calls a day. How do you find failure modes you haven't written a metric for yet? Anyone clustering failed calls or auditing batches?

document fetched 2026-11-18 00:20 · canonical https://reddit.com/r/AI_Agents/comments/1xydhou/our_voice_agent_passes_every_metric_we_wrote_and