What Production Quality Metrics Should You Track for a Live AI Agent

Quick answer

Track five categories continuously in production: containment rate, escalation reasons (not just escalation count), response accuracy sampled by human review, latency and cost per conversation, and user-reported satisfaction. Pre-launch eval scores tell you the agent was good on a fixed test set; production metrics tell you whether it's still good as real traffic and edge cases evolve.

A strong eval score before launch answers one question: was the agent good on the test set you built. It says nothing about whether the agent stays good three months into production, once real users start asking things your eval set never anticipated. Teams that stop measuring after launch are flying blind on exactly the risk that matters most: silent quality drift.

Why pre-launch evals aren't enough on their own

An eval dataset, however carefully built per building a golden eval dataset, is a snapshot. Real user behavior shifts: new product features get released, seasonal questions change, competitors change their pricing and prompt different comparison questions. A static eval set doesn't catch any of that drift on its own.

The five categories worth tracking continuously

Containment rate, the percentage of conversations the agent resolves without human handoff, is the metric most teams already track, and it's a good headline number, but insufficient alone. See measuring AI agent containment rate for how to calculate it correctly, since naive versions overstate success by counting abandoned conversations as "contained."

Escalation reasons, not just escalation count. A rising escalation rate is a symptom; the reason code is the diagnosis. Segment escalations by cause: agent didn't understand the question, agent gave a low-confidence answer and correctly deferred, user explicitly asked for a human, tool or API failure. Each cause points to a different fix.

Sampled human review of accuracy. Automated metrics catch a lot but not everything, particularly nuanced correctness or tone issues. Have a human reviewer spot-check a random sample of conversations weekly, scored against a simple rubric, to catch what automated monitoring misses.

Latency and cost per conversation. These directly affect both user experience and your margins. A slow but accurate agent still damages the customer relationship; see AI agent latency and cost for customer-facing use cases for target ranges by use case.

User-reported satisfaction, whether a thumbs up/down, a short post-conversation survey, or downstream CSAT and NPS impact. This is the metric that ultimately ties technical quality to actual business outcomes rather than an internal quality score nobody outside the team ever sees.

Building this into a dashboard, not a report

The teams that catch quality drift early are the ones with a live dashboard reviewed at least weekly, not a quarterly report generated after problems have already compounded. Set alert thresholds on containment rate and escalation-reason distribution specifically, since those two shift first when something is degrading.

Frequently asked questions

How often should the eval dataset itself be refreshed? Quarterly at minimum, and immediately after any major product change that could shift the kinds of questions users ask.

Is a thumbs up/down rating actually reliable? It's noisy at the individual level but useful in aggregate trend over time; watch the trend line, not any single day's number.

What's a reasonable cadence for the human accuracy review? Weekly for a growing product, biweekly once the agent's behavior has stabilized and escalation-reason data isn't flagging new issues.

Should cost per conversation be a hard gate, or just monitored? Monitor it continuously and set a soft alert threshold; treating it as a hard gate risks the team quietly degrading response quality to cut cost, which shows up in other metrics later anyway.